Martin Odersky's Scala Levels

I just saw Martin's post on levels of expertise in Scala, and Tony Morris's response. The concept of levels is something that many in the community have called for repeatedly, in order to help ease people - especially team members who may be a little put off by some of Scala's more advanced features - into the language more gently. I understand the motivation, but it has always struck me as somewhat odd. The reason is that I think one needs to have very compelling reason for such a significant change to be justified, especially a change from a mature, mainstream technology such as Java into one as new an immature as Scala. To being, I'll recount my experience with adding Python as a stable in my programming toolbox.

Learning Python

I dabbled it Python a lot before really using it extensively. Much of my initial usage basically amounted to using Jython as a repl for interactively poking at Java libraries, especially ones that weren't particularly well documented. I thought Python was neat, but at the time most of my programming experience was in C++ and Java. I valued static typing. I valued having a compiler. When I was in college, I had a professor who said something along the line so of "The compiler is your friend. Write your code so that the compiler is more likely to catch your mistakes. Don't try to trick it or bypass it." This was in the context of C++ programming, and given all the ways that one can shoot himself in the foot using C++, it always seemed like sage advice. So giving up have a compiler to catch my mistakes, and an explicit structure clearly present in my code, seemed like an awful lot to ask. Sure, Python is more concise, but in my opinion conciseness is often more a matter of how closely the libraries you're using match the problem that you are trying to solve. About a year ago I had some experimental Python code that I had thrown together that did some regular expression heavy text processing then some XML processing (I was processing output of an application that produced invalid XML, so the first phase was the fix it so the XML parser wouldn't crash and the second was to actually process it). I can't remember why, but I decided to convert it to Java. I think it was because I wanted to integrate it with something that integrated well with Java, but not as well with Python. Anyway, the point is that the Python code and the Java code weren't that different in length. Sure, Java has a lot more boiler plate, but once your code starts being dominated by actual logic instead of getters and setters, it's the libraries that make the difference.

Anyway, I didn't really get into Python until I discovered and thoroughly learned metaclasses and descriptors. In my mind, they took Python from being a toy with a clean syntax to a language that opened up an entire new dimension of abstraction. A dimension of abstraction that had been missing all me life. A dimension that I had glimpsed in C++ template metaprogramming but had never quite managed to enter. All of a sudden I had this tool that enabled me to drastically and repeatedly reduce code sizes all while making the business logic cleaner and more explicit. I had a means to encode my ideas such that they enabled my coworkers to use them without fulling understanding them (an admittedly dangerous proposition). It was beautiful! They remain an important part of my toolbox to this day.

Incidentally, my first uses metaclasses and descriptors was to concisely implement immutable, lazy data structures. This was long before I knew anything about functional programming.

Learning Scala

I also dabbled in Scala a lot before really becoming enamored with it. I was caught by the conciseness over Java, especially the absence of getters and setters, the presence of operator overloading, and multiple inheritance using traits. But I had operator overloading and multiple inheritance in both C++ and Python. I could usually abstract away explicit getters and setters in Python, and by the time I started picking up Scala I had already made a significant shift towards immutability and laziness, so setters weren't as necessary. Scala was neat, and the compatibility with Java was a real benefit, but at first it didn't seem compelling. In fact, when I first started experimenting with Scala it seemed too buggy to really be worth it. Something worth watching, but not yet worth using. Also, by the time I started learning Scala I had already learned Common Lisp and functional programming with full tail-call optimization seemed awkward (and still does).

What eventually sold me on Scala was higher-kinded types, companion objects, and objects-as-modules. You see, but the time I started picking up Scala I had already developed a habit of using a lot of generic programming. This is easy in Python where there's no compiler check a type system to tell you that you are wrong. But a lot of the patterns I commonly used couldn't be translated into Java generics (in fact, in my first foray into Java generics I crashed javac writing code that was compliant with the Java specification). Scala gave me concise programs and type safety! And with the introduction of collections that actually use higher-kinded types in 2.8.0, along with the leaps and bounds made in robustness, I think it finally reached threshold for serious use.

That's not to say these are the only key features. Understanding folds was a real eureka moment for me. I'm not sure how one can do effective functional programming without them, but perhaps there is a way. Implicit definitions are also key, and they are used so extensively in Scala libraries that I think writing Scala without a decent understanding of them is nothing short of voodoo programming. In short, there's a lot, and the features tend combine together in powerful ways.

Summary

I understand the motivation for establishing levels, but I personally think if one avoids many of the features that Martin has flagged as advanced, then the benefits of Scala don't justify the risk of using such a new technology and the need to retrain people. I think without them one is better off finding a better library or framework for their current stack. Also, between the fact that advanced features are omnipresent in the libraries, and teams will always have someone who wants to use the more powerful features, avoiding them works a lot better in theory than it does in practice. You can't not use them, unless you don't use Scala's standard library and avoid almost any library written explicitly for Scala. You can only pretend that you are not using them.

This is probably going to get me flamed, but reward must justify risk. In many circumstances Scala has the features to justify the risk associated with it, but the justification quickly vanishes as the more powerful features are stripped away. So while I think a programmer may be productive at Martin's lower levels, they don't justify Scala over Java.

Sphere: Related Content

Easier object inspection in the Scala REPL

If you're like me, you spend a fair amount of time poking at objects of poorly documented classes in a REPL (Scala or otherwise...). This is great compared to writing whole test programs solely for the purpose of seeing what something does or what data it really contains, but it can be quite time consuming. So I've written a utility for use on the Scala REPL that will print out all of the "attributes" of an object. Here's some sample usage:

scala> import replutils._
import replutils._

scala> case class Test(a: CharSequence, b: Int)
defined class Test

scala> val t = Test("hello", 1)
t: Test = Test(hello,1)

scala> printAttrValues(t)
hashCode: int = -229308731
b: int = 1
a: CharSequence (String) = hello
productArity: int = 2
getClass: Class = class line0$object$$iw$$iw$Test

That looks fairly anti-climatic, but after spending hours typing objName. to see what's there, and poking at methods, it seems like a miracle. Also, one neat feature of it is that if the class of the object returned is different from the class declared on the method, it prints both the declared class and the actual returned class.

Code and further usage instructions is available on BitBucket here.

So what exactly is an attribute?

I haven't quite figured that out yet. Right now it is any method that meets the following criteria:

1. the method has zero arguments
2. the method is public
3. the method is not static
4. the method's name is not on the exclude list (e.g. wait)
5. the method's return type is not on the exclude list (e.g. Unit/void)
6. the method is not a "default argument" method

Please post any feedback in the comments here, or better yet in the issue tracker on Bitbucket.

Sphere: Related Content

Playing with Scala Products

Have you ever wanted something like a case class, that magically provides decent (for some definition of decent) implementations of methods like equals and hashCode, but without all of the restrictions? I know that I have. Frequently. And recently I ran into an answer on StackOverflow that gave me an idea for how to do it. The trick is to play with Products. I'm not sure how good of an idea it is, but I figure the Internet will tell me. Here's the basic idea:

import scala.runtime.ScalaRunTime

trait ProductMixin extends Product {
  def canEqual(other: Any) = toProduct(other) ne null
  protected def equalityClass = getClass
  protected def equalityClassCheck(cls: Class[_]) = equalityClass.isAssignableFrom(cls)
  protected def toProduct(a: Any): Product = a match {
    case a: Product if productArity == a.productArity && equalityClassCheck(a.getClass) => a
    case _ => null
  }
  override def equals(other: Any) = toProduct(other) match {
    case null => false
    case p if p eq this => true
    case p => {
      var i = 0
      var r = true
      while(i < productArity && r) {
 if (productElement(i) != p.productElement(i)) r = false
 i += 1
      }
      r
    }
  }
  override def productPrefix = try {
    getClass.getSimpleName()
  } catch {
    //workaround for #4118, so classes can be defined in the REPL that extend ProductMixin
    case e: InternalError if e.getMessage == "Malformed class name" => getClass.getName()
  }
  override def hashCode = ScalaRunTime._hashCode(this)
  override def toString = ScalaRunTime._toString(this)
}

Basic Usage

There's a couple ways to use ProductMixin. The first, and probably the simplest, is to extends one of the ProductN classes and provide the appropriate defs:

class Cat(val name: String, val age: Int) extends Product2[String, Int] with ProductMixin {
  def _1 = name
  def _2 = age
}

The second way is to provide them yourself:

 class Dog(val name: String, age: Int) extends ProductMixin {
  def productArity = 2
  def productElement(i: Int) = i match {
    case 0 => name
    case 1 => age
  }
}

Either way, the result is something that has many of the benefits of case classes but allows for more flexibility. Here's some sample usage:

scala> val c1 = new Cat("Felix", 2)
val c1 = new Cat("Felix", 2)
c1: Cat = Cat(Felix,2)

scala> val c2 = new Cat("Alley Cat", 1)
val c2 = new Cat("Alley Cat", 1)
c2: Cat = Cat(Alley Cat,1)

scala> val c3 = new Cat("Alley Cat", 1)
val c3 = new Cat("Alley Cat", 1)
c3: Cat = Cat(Alley Cat,1)

scala> c2 == c3
c2 == c3
res0: Boolean = true

scala> c1 == c2
c1 == c2
res1: Boolean = false

Dealing with Inheritance

Equality in the presence of inheritance is very tricky. But, possibly foolishly, ProductMixin makes it easy! Let's say you have a hierarchy, and you want all the classes under some root class to be able to equal each other. Here's how it would be done by overriding the equalityClass so that it is the root of the hierarchy (using a not-so-good example):

scala> abstract class Animal(val name: String, val age: Int) extends Product2[String, Int] with ProductMixin {
     | override protected def equalityClass = classOf[Animal]
     | def _1 = name
     | def _2 = age
     | }
defined class Animal

scala> class Dog(name: String, age: Int) extends Animal(name, age)
class Dog(name: String, age: Int) extends Animal(name, age)
defined class Dog

scala> class Cat(name: String, age: Int) extends Animal(name, age)
class Cat(name: String, age: Int) extends Animal(name, age)
defined class Cat

scala> val c = new Cat("Felix", 1)
val c = new Cat("Felix", 1)
c: Cat = Cat(Felix,1)

scala> val d = new Dog("Felix", 1)
val d = new Dog("Felix", 1)
d: Dog = Dog(Felix,1)

scala> c == d
c == d
res0: Boolean = true

The reverse can also be accomplished by overriding equalityClassCheck such that it checks that the classes are equal instead of using isAssignableFrom. That would mean two objects could be equal if and only if they are the same class.

Wrapup

I don't know to what extent this is a good idea. I've only tested it a bit in the REPL. It's neat, but there is one problem that comes to mind: performance. Any primitive members that are elements of the product will end up being boxed prior to usage in the hashCode and equality methods, the extra layers of indirection aren't free, and neither is the loop. That being said, case classes use the exact same method for hashing, so in what way they pay the same penalty, and unless code is really hashing and equality heavy it probably wouldn't be noticeable.

Sphere: Related Content

Scala is for VB programmers

Normally I don't go for flame bait titles. But I haven't finished my morning coffee yet so I can't help myself. There's once again a debate raging across the internet about whether Scala is more or less complex than Java, along with the more nuanced argument that yes, it is more complex but only framework and library developers have to contend with this complexity. This argument usually happens when someone posts a little bit of code like this:

def map[B, That](f: A => B)(implicit bf: CanBuildFrom[Repr, B, That]): That

And then someone responds like this:

Why do not people realize that Java is too difficult for the average programmer? That is the true purpose of Scala, to escape the complexity of Java code! Framework code in Scala, with heavy use of implicit keywords and all kinds of type abstractions, is very difficult. This is correct, but this code is not meant for innocent eyes. You do not use that sort of code when you write an application.

I've seen this type of thinking before. A few years ago I had a bout of insanity and lead an ASP.NET project using ASP.NET 2.0. I had no .NET experience prior to this project. The project failed, although the reasons for that failure are legion and unimportant here. But I noticed something about ASP.NET developers: they have no clue how the technology works. It's a black box. Do you why? Because it is a black box. I searched and searched and couldn't even find a good illustration of the lifecycle for an ASP.NET page that's using events. This type of information is put front and center in the Java world. It's absolutely buried in the Microsoft world. Or at least the parts of it that target the hoards of VB programmers that are undyingly loyal to MS. The framework is magic dust created by the great wizards in Redmond so that you can build solutions for your customers. Do not question the dust. Think about VB. Or, no don't, it might damage your brain. My coffee hasn't quite kicked in, so I should have some immunity, so I'll do it for you. VB is a black box (well, at old school VB6). It was designed to allow people who do not know how to really program, and who will probably never know how to program, to create applications. It's completely flat, opaque abstraction. The dichotomy between the application developer and the framework developer is as high and as thick as the gates of Mordor.

There are many people in the Scala community that claim Scala's complexity can be hidden from the application program. I don't believe them, but there's a chance that they are right. It's technically feasible, and I can see how it could happen if Scala started attracting lots of VB programmers. I can't see how it's going to attract lots of VB programmers, but apparently many people in the Scala community think Scala is for VB programmers. So we'll just have to wait and see...

Sphere: Related Content

Scala Actors: loop, react, and schedulers

One of the unfortunate aspects of many of the "published" (meaning blogged) Scala Actor benchmarks out there is that they rarely pay much attention, if any, to the affects of seemingly idiomatic patterns on performance. Some of the main culprits are:

1. react versus receive (event-based versus threaded)
2. loop/react versus recursive react
3. loop/receive versus receive/while
4. tweaking (or failing to tweak) the scheduler

I've been working on setting up a "benchmarking framework" in conjunction with experimenting with modifications to the underlying thread pool so that all the possible permutations are automatically tested. What I have right now is a classic "ring" benchmark setup to permute the schedulers and loop/react versus recursive react. The loop/react pattern is more idiomatic (or at least more common), but higher overhead, and it looks something like this:

loop {
  react {
    case Msg(m) => // do stuff
    case Stop => exit()
  }
}

The reason it is high-overhead is that both loop and react raise control flow exceptions that result in the creation of new tasks for the thread pool when they are hit, so for each loop, two exceptions are raised and two tasks are executed. There's overhead in both of the operations, especially raising the exceptions. The recursive react pattern looks like this, so it can avoid the extra exception/task:

def rloop(): Unit = react {  //this would be called by the act() method
  case Msg(m) => {
    // do stuff
    rloop()
  }
  case Stop => // just drop out or call exit()
}

Using loop instead of recursive react effectively doubles the number of tasks that the thread pool has to execute in order to accomplish the same amount of work, which in turn makes it so any overhead in the scheduler is far more pronounced when using loop. Now, I should point out that the overhead really isn't that large, so if the actor is performing significant computations it will be lost in the noise. But it's fairly common to have actors do fairly little with each message. Here's some results from the ring benchmark using 10 rings of 10,000 actors passing a token around them 100 times before exiting. I'm using multiple rings because otherwise there is no parallelism in the benchmark. These are being run on my dual core Macbook.

Scheduler	ReactMethod	Time (sec)
ManagedForkJoinScheduler	LoopReact	45.416058
ManagedForkJoinScheduler	RecursiveReact	25.509482
ForkJoinScheduler	LoopReact	65.268584
ForkJoinScheduler	RecursiveReact	45.85605
ResizableThreadPoolScheduler	LoopReact	98.084794
ResizableThreadPoolScheduler	RecursiveReact	53.379757

The fork/join schedulers are faster than the ResizableThreadPoolScheduler because rather than have all of the worker threads pull tasks off of a single, shared queue; each thread maintains its own local dequeue where it can place tasks directly onto if they are generated while it is running a task. This creates a kind of "fast path" for the tasks that involves much less overhead.

I believe the primary reason ManagedForkJoinScheduler is faster because ForkJoinScheduler does not always leverage the "fast path," even when in theory it could be used. I'm unsure about some of the rationale behind it, but I know some of the time the fast path is bypassed probabilistically in order to reduce the chances of starvation causing deadlock in the presence of long running or blocking tasks. ManagedForkJoinScheduler escapes this particular issue by more actively monitoring the underlying thread pool, and growing it when tasks are being starved. The second reason, and I'm somewhat unsure of the actual degree of the affects, if that ForkJoinScheduler configures the underlying thread pool so that the threads work through the local dequeues in FIFO order, while ManagedForkJoinScheduler configures the pool such that the local dequeues are processed in LIFO order. Processing in LIFO order allows the pool to take advantage of locality with regard to the tasks, basically assuming that the last task generated is the most likely to use data that's currently in cache, and thus reduce cache misses.

The benchmark outputs a lot more information than I captured in the above table. If you'd like to run it, you can obtain the code here. The project uses sbt, so you'll need to have it working on your computer. After you run update in sbt to download all of the dependencies, you can run the ring benchmark as follows:

$ sbt
[info] Building project ManagedForkJoinPool 1.0 against Scala 2.8.0
[info]    using ManagedForkJoinPoolProject with sbt 0.7.4 and Scala 2.7.7
> ringbenchmark
[info] 
[info] == compile ==
[info]   Source analysis: 1 new/modified, 0 indirectly invalidated, 0 removed.
[info] Compiling main sources...
[info] Compilation successful.
[info]   Post-analysis: 79 classes.
[info] == compile ==
[info] 
[info] == copy-resources ==
[info] == copy-resources ==
[info] 
[info] == ringbenchmark ==
[info] RingBenchmark ManagedForkJoinScheduler LoopReact 2 ....output truncated...

You can tweak the benchmarks by modifying the sbt project file. If you do run them, I'm very interested in the results.

Sphere: Related Content

Concurrency Benchmarking, Actors, and sbt tricks

Have you ever noticed that other people's microbenchmarks tend to be hard to run and often impossible to duplicate? And are frequently caveated to the hilt? When it gets down to it, a benchmark is really an experiment, and ideally a scientific experiment. That means all factors that are relevant to the results should be clearly recorded, and the tests should be easy for others to duplicate.

Custom sbt actions for benchmarks

In order to test and run benchmarks on the work I'm doing around creating a managed variant of the JSR-166y ForkJoinPool along with supporting infrastructure for use with Scala Actors, I'm creating a test harness that captures a host of environmental factors about how it was run, and writing sbt actions to make it easy to run the benchmarks and automatically permute the variables.

It still needs a lot of work, but I had some trouble figuring out a really basic task so I thought I'd share it. Basically I wanted to build a Task object that consists of several tasks based on information in the project definition and permuted parameters. It actually pretty easy, as you can see in the snippet below from my project definition:

  /** this task executes the PingPong benchmark using each available scheduler */
  lazy val pingpongbench = pingpongTaskList
  /** produces a sequence of run tasks using all the available schedulers  */
  def pingpongTaskList = {
    val pairs = 100
    val messagesPerPair = 10000
    val tasks = for(sched <- schedulers) yield pingpongTask(sched, pairs, messagesPerPair)
    tasks.reduceLeft((a, b) => a && b)
  }

You can see the whole file here. Basically Task has an && operator that essentially allows you to concatenate one task with another task. This allows you to build a whole chain of tasks. In the example above, I'm having it run the benchmark once for each scheduler configuration. Soon, I'm going to make it permute other parameters. But right now my test harness isn't playing nicely with the schedulers included in the Scala distribution, so first things first.

There's also one other little customization, which is documented, but I think it's important for benchmarking. By default, sbt runs your code in its own process. This can cause problems with multithreaded code, especially if it doesn't terminate properly. It also means the next benchmark to run has to content with any junk that the previous benchmark left around. So I configured sbt to fork new processes. It just required one line:

override def fork = forkRun

Important variables

Here's what I'm capturing for each run right now so that the results can all be dumped into a big spreadsheet for analysis. I'd like to capture more information about the host machine, such as more information about the CPUs and the loading when the benchmark is being run, but haven't got that far yet. Currently these are all captured from within the benchmark process, mostly using system properties and the Runtime object.

1. Test Name - obviously needed so that results from multiple benchmarks can be stored in the same file
2. Scheduler - this is my primary variable right now, I want to run each benchmark with each scheduler while holding everything else constant
3. # of Cores/Processors - essential so that anyone looking at the results has an idea about the hardware used
4. Java VM Name - different VMs can perform quite differently
5. Java VM Version - performance characteristics change from version to version (usually getting better)
6. Java Version - same reason as above, but this is probably the more publicly known version number
7. Scala Version - this could be important in the future, as it becomes more common for different projects to be on different version of Scala
8. OS Name and version - again, it can affect performance
9. Processor Architecture
10. Approximate Concurrency (number of simultaneously alive actors) - this allows us to examine concurrency levels versus resource consumption, more concurrency does not necessarily mean that more cores or threads would be helpful
11. Approximate Parallelism (number of simultaneously runnable actors) - this measures how many cores/threads the benchmark can really keep busy
12. Approximate Total Messages - this estimates the amount of activity that takes place during the benchmark, generally the benchmarks I'm looking at contain very little logic because they are intended to measure overhead introduced by the framework
13. Total Wall Clock Time (seconds) - as measured using nanoTime within the benchmark process
14. Initial Thread and Maximum Observed Thread Count - used to examine automatic expansion of the thread pool
15. Initial Free Memory and Minimum Observed Free Memory - threads use a fair amount of memory, so performance impacts may show up as pressure on the GC as well has contention for the CPU
16. Initial and Maximum Observed Total Memory - threads use a lot of memory, so it's important to track usage
17. Verbose - debugging output pretty much invalidates any of these tests

Sphere: Related Content

Improving Schedulers for High-level, Fine-grained Concurrency Frameworks

Quite a while ago, Greg Meredith made a comment that really made me stop and think, and that has been lingering in the back of my mind ever since:

Erik, Alex, John, et al, i'm loathe to hijack this thread -- which is a good one -- but the experience with the lock up described below is really just the tip of the iceberg. Unless specific scheduling contracts and semantics can be worked out, a pluggable scheduling architecture is just asking for disaster. Worse, it means that a whole host of compiler support that could be provided to apps that use actor-based concurrency is null and void because those solutions (be they model-checking or types-based) will critically depend on specific ranges of scheduling semantics. Actors and other higher level concurrency constructs are a major appeal of languages like scala. If they prove themselves to be "not ready for prime time" the impact on perception might not be contained to just that usage of the language. Best wishes, --greg

That wasn't the first time that Greg made a comment along those lines, and it certainly wasn't the last. At one point he offered to allow a few individuals look at some older code that he believed could help. Greg's a really smart guy. If he's worried, we should all be worried.

Empirical Evidence of the Problem

Contrary to what many may think this is hardly an academic problem. A significant portion of the issues people new to Scala Actors have trace to the fact that they expect the scheduler to be making guarantees that it simply does not make. Here's a sampling:

1. Actors not creating enough threads
2. Scala Actors different behavior on JRE 1.5 and 1.6
3. Actor starvation problem (related to previous)
4. Actor as new thread

There are others, including cases where people have both rightly and wrongly blamed the default scheduler for serious problems with their programs. Most of the above can be traced back to the fact that the default scheduler for Scala Actors uses a slightly modified version of the JSR166y ForkJoinPool, which has issues described below (source):

Like all lightweight-task frameworks, ForkJoin (FJ) does not explicitly cope with blocked IO: If a worker thread blocks on IO, then (1) it is not available to help process other tasks (2) Other worker threads waiting for the task to complete (i.e., to join it) may run out of work and waste CPU cycles. Neither of these issues completely eliminates potential use, but they do require a lot of explicit care. For example, you could place tasks that may experience sustained blockages in their own small ForkJoinPools. (The Fortress folks do something along these lines mapping fortress "fair" threads onto forkjoin.) You can also dynamically increase worker pool sizes (we have methods to do this) when blockages may occur. All in all though, the reason for the restrictions and advice are that we do not have good automated support for these kinds of cases, and don't yet know of the best practices, or whether it is a good idea at all.

The Scope of the Issue

The issue here is neither limited to Scala nor to actor based concurrency. The general consensus in the concurrency community is that locks and threads are not the right abstractions for concurrency in application code (or hardly any code, for that matter), but there isn't any consensus on what the right abstractions are, or if there is consensus it is that the right abstractions are problem specific. There's no one-size-fits-all concurrency abstraction.

The one trait that all high-level or higher-order concurrency abstractions have in common is that under the hood they rely on some sort of managed thread pool and/or scheduler. For purposes here, when I say "scheduler," I mean the layer with which a concurrency framework interacts directly and likely contains framework-specific logic. When I say "thread pool," I mean the layer that directly manages the threads and is concurrency framework agnostic and may even be shared by several schedulers (sharing will be problematic, but I think ultimately necessary). This isn't a hard line, and often times they may be lumped into one. However I think it's a useful dichotomy. I'm also assuming the scheduler and thread pool rely on cooperative thread sharing, and that preemptive multitasking is left to the virtual machine and/or operating system.

The point is, any concurrency abstraction where the user can submit arbitrary code to be executing can ultimately run into serious issues if that user code does something it does not expect. For example, the fork-join scheduler from JSR-166y that the Scala Actors library uses (I think Clojure uses it as well, but it may not) doesn't automatically enlarge its pool in the presence of unmanaged blocks (including normal blocking IO) or simple long-running computations. This results in the majority of the problems I previously cited, because blocking operations are extremely common tasks, and thus a very leaky abstraction.

Key Characteristics of a Good Scheduler

Unfortunately my command of type theory is rudimentary at best, so while Greg could probably describe a well-behaved scheduler using types, I'm going to have to resort to plain English:

1. The user should not have to think about choosing a scheduler.
2. The user should not have to think about the size of the thread pool managed by the scheduler.
3. The user should be able to write code naturally without worrying about any assumptions the scheduler may make
4. The user should not have to worry about starting or stopping the scheduler
5. Multiple flavors of concurrency abstraction should be able to share the same thread pool.
6. The scheduler should impose minimal overhead.

There's probably others, but I think this is a good start.

The Performance Problem

So why hasn't this problem already been solved? Perhaps it has been already and I just don't know about it. If someone has, please let me know. But as far as I can tell it's because there's a perceived nasty performance trade. Here's a quote from Philipp Haller:

The starvation detection mechanism in Scala 2.7 was too expensive to support in the new default scheduler. However, there is a starvation detection mechanism for thread-blocking methods of the `Actor` object (`receive` etc.).

Similar answers have been provided when people have questioned the wisdom of using the FJ framework over the classes included with the JDK. I've tested it a bit and it's true, the FJ framework does yield a significant improvement for certain classes of tasks over JDK classes (although I believe using exceptions for control flow imposes a far larger penalty). Basically it appears that the overhead associated with known solutions to the problem overcomes the introduced benefits of fine-grained concurrency,

ForkJoinPool in a nutshell

What makes ForkJoinPool so fast (other than highly optimized code) is that uses a two-level architecture for its task queues. There's one queue for the pool, and the one dequeue per thread. Most of the data associated with the dequeues is only updated from the owning thread so that tasks can be added and removed without synchronization and minimal volatile and CAS operations. The worker threads also steal work from one another in order to spread out the load. When a worker is deciding the next task to execute, it performs the following checks, using the first one to return a task:

1. The thread's own local dequeue
2. The dequeues of other threads in the pool, scanned in a psuedo-random pattern with guiding hints
3. The common task queue

Threads are only added to the pool if a ManagedBlocker that the pool can see if used to block causing the target concurrency to not the be met. The default target concurrency for the ForkJoinPool is set to the number of available processors. In Scala Actors this is overridden to be twice the number of available processors. The idea behind this is simple: If you're keeping all your processors busy, then adding more threads will just slow you down. As we've seen, problems arise when a higher level of concurrency is needed due to unmanaged blocks, long running computations, or simply the required concurrency being inherently higher than the default concurrency. This happens because as long as a worker is busy or has something in its local dequeue, it won't look elsewhere for tasks. If all the workers are in this state, tasks can simply accumulate, resulting in starvation.

A Crack at Solving the Problem

As you might have guessed by now, I'm trying to solve this problem. Right now I wouldn't use the code for anything beyond testing the code, but despite the fact I'm still wrestling with the internals of ForkJoinPool I've experienced good results so far. My approach is simple: I added in a manager thread that monitors the changes in size of the various queues in the pool, and if they've been flushed since the last check, and grows the pool it tasks appear to be being starved. The overhead imposed on each worker is minimal, as it only has to update a single additional volatile field when it clears out its queue. The overhead of the manager thread is something, but in the benchmarking I've done so far it doesn't seem to add noticeable overhead. Ironically, Scala Actors already maintain a dedicated thread for the scheduler, its just that the heuristic check on the size of the pool were removed in 2.8 (although honestly they never worked that well).

I have two simple tests/micro-benchmarks using Scala Actors with a custom scheduler built on my ManagedForkJoinPool. The SleepyActor benchmark performs extremely poorly if the scheduler doesn't significantly grow the pool, because each actor sleeps on the thread. The PingPong benchmark spawns a bunch of pairs of actors that simply message each other back-and-forth. It is the type of use case where the ForkJoinScheduler shines, at least in terms of throughput. The results on my MacBook are as follows (these are NOT scientific):

Benchmark	Default Scheduler	ManagedForkJoinScheduler	ResizeableThreadPoolScheduler
SleepyActor	92 sec	30 sec	not tested
PingPong	59.49 sec	47.17 sec	66.02 sec

As you can see, performance actually improved with my scheduler. This is because the default scheduler for Scala Actors does not always use the "fast path" and put new tasks on the dequeue of the thread creating them. It only does it about half the time. So for the PingPong test you can see how the thread local queues impact performance.

Conclusions

It's too early to draw solid conclusions, but based on what I've done so far I think I can develop a solid heuristic for managing the thread pool size, and that the overhead will be negligible. The key is to not impose any more overhead on the worker threads, and to keep the computational overhead of the manager thread a low as possible. This means that the heuristic needs to be simple, and that the wait times between checks should be a long as possible, probably dynamically sized based on how smoothly the pool seems to be running.

What I need right now are more tests and benchmarks to cover "real world" scenarios. I also need to test on machines with more cores. My MacBook doesn't exactly present a powerful system. If anyone has suggestions, please post them!

Sphere: Related Content

Higher-Level versus Higher-Order Abstraction

Engineering in general, and software engineering in particular, is all about abstraction. The creation and utilization of abstractions is a core part of the daily activities of every good engineer, and many engineers are on a never ending quest to increase their level of abstraction. Abstraction both increases productivity and increases the complexity of problems that can be tackled. Few would argue that increased abstraction is a bad thing.

But people do argue about abstraction, and often condemn abstractions that they view as unfit or too complex. It is common to hear a software developer praise the merits of one means of abstraction and then demean another in the same breath. Some abstractions are too general. Some aren't general enough. Some are simply too complex or, ironically, too abstract. So while engineers almost universally agree that abstraction is good, and more abstraction is better; they often disagree fiercely about what genuinely constitutes "better."

What is an abstraction?

Simply put, an abstraction is something that hides the details involved in realizing a concept, and thus frees the engineer to focus on the problem at hand rather than the details of some other concern. This can take many, many forms. The simplest and most common, so common that the term is often used to describe almost all "good" abstractions, is higher-level abstractions.

Higher-level Abstractions

Higher-level abstractions are fairly simple: they encapsulate details so that they can be used without knowledge or concern about the details. Let's consider a simple example in Scala 2.8 (you can copy and paste these examples directly into the Scala 2.8 REPL):

scala> def sum(data: Array[Double]): Double = {
     |   var i = 0
     |   var total = 0.0 
     |   while(i < data.length) {
     |     total = total + data(i)
     |     i = i + 1
     |   }
     |   total
     | }
scala> def mean(data: Array[Double]): Double = sum(data) / data.length

So we have two functions: one that computes the sum of an array of doubles, and one that computes the mean. Pretty much any programmer (professional or otherwise) would feel confident both writing and using such functions. They are written in a modern multi-paradigm programming language yet I bet if you went back in time and showed them to programmers in some of the very first high-level languages they would recognize exactly what they do. They clearly encapsulate the details of performing certain computations. But what's more interesting about them is what's missing from them:

1. Polymorphism
2. Closures or first-class functions
3. Usage or creation of higher-order functions
4. Other Scala magic such as type classes and implicit parameters

In other words, they provide a simple, layered, hierarchical abstraction in a very bounded way. If we step away from software for a minute, you can imagine a digital designer placing a symbol for an adder or a register on a schematic without worrying about the arrangement of transistors that will be required to realize them in an actual circuit, or a mechanical engineer selecting a standard screw or clamp. These are parts that can be simply built, tested, and composed into larger devices.

Higher-Order Abstraction Take 1: The Mathematics of Fruit

Imagine I have some apples. But unfortunately I'm here, off in the internet, show I can't show them to you. I need an abstraction to tell you about them. For example, I could say I have two apples. Two is a number, and numbers are abstractions. I can use the same number two to describe the number of McIntoshes in my refrigerator or the number of Apple Macintoshes I own. Now let's say I also want to talk about my strawberries and blueberries. I have 16 strawberries and 100 blueberries. How many pieces of fruit do I have? 118! How did I figure that out? I used arithmetic, which is an abstraction of higher order than numbers. How let's say I want to know how many days it will be before I will have eaten all my strawberries. I can write an equation such as: current_fruit - pieces_per_day * number_of_days = fruit_after_number_of_days. I can do even more with this by solving for different variables. This is algebra, and it is a higher-order abstraction than arithmetic. Now let's say I want to build upon that so that I can study the dynamics of the amount of fruit in my refrigerator. I purchase fruit and put it in my refrigerator. I take fruit out and eat it. I forget about fruit and it gets moldy, the I remove it and throw it away. I can capture all of these as a system of differential equations, and using calculus describe all sorts of things about my fruit at any given point in time. Calculus is a higher-order abstraction that algebra. In fact, abstractions similar to the ones built with calculus are what I mean when I say "higher-order abstraction."

At each step up the chain of mathematics both the generality and number of concepts that can be conveyed by a single abstraction increased. In the case of calculus it becomes essentially infinite, and that's the essence of higher-order abstractions: they deal with the infinite or near-infinite. Also, observe that almost everyone from relatively small children on up understand numbers and arithmetic, most adolescents and adults can stumble through applying algebra, and only a very small portion of the population knows anything about calculus, much less can effectively apply it. Also observe that the functions I defined earlier are just below algebra in terms of their order of abstraction.

Higher-Order Abstraction Take 2: Programming in Scala

Let's see if we can sprinkle some higher-order abstraction into our previously defined functions:

scala> def sum(data: Traversable[Double]) = data.foldLeft(0.0)(_ + _)

This definition of sum exposes one higher-order abstraction (polymorphism - it now can use any Traversable[Double] instead of just an Array[Double]), and it uses higher-order functions in its implementation to perform the summation. This new definition is both much shorter and much more general, but it's still relatively approachable, especially for use. Calling it with an Array[Double] works as before, and now it can be used with any number of collections, so long as the collections contain doubles. But forcing the collections to contain doubles is very limiting, so let's see if we can do better:

scala> def sum[T](data: Traversable[T])(implicit n: Numeric[T]): T = {                                  
     |   import n._
     |   data.foldLeft(zero)(_ + _)
     | }

Ahhh, that's better! Much more general. Not only will this work for any numeric type defined in the Scala standard library, but for any numeric type for which the Numeric type class has been defined! It doesn't even need to be in the same class hierarchy! In order to introduce this new level of generality, we've also introduced the following higher-order abstractions:

1. Classic polymorphism (Traversable instead of Array)
2. Parametric polymorphism (the type parameter T for various classes)
3. Higher-order functions and closures (foldLeft and it's argument that does addition)
4. Type classes (well, Scala's flavor of type classes, e.g. Numeric)

Now, this is the point where people start screaming that abstraction has gone too far. Many professional programmers would look at it and think "WTF?" They could still guess what it does, but the mechanics are quite elusive for anyone that doesn't know Scala reasonably well. That being said, the code is still far more compact than the original imperative version and is extremely general (to which someone replies "Yes, but it's slow a heck compared to the original!"). At this point I would say the order of abstraction has went from being just below algebra to being on par with integral calculus. Just like with mathematics, we see a significant drop off in the number of people who readily understand it.

Higher-Order Abstraction Take 3: Conclusions

Let's consider a short, incomplete list of higher-order abstractions, means of abstraction, and fields that rely upon higher-order abstractions:

1. Higher-order functions
2. Polymorphism and parametric polymorphism
3. Metaclasses such as in Python and Smalltalk
4. Rich macros such as in Lisp or C++ Templates
5. Calculus and all the mathematics built upon it (and various other forms of advanced math)
6. First-order Logic and its kin, and all the higher-order logics
7. Any physics or other sciences that are built upon calculus
8. Any engineering or other applications that are built upon physics

Higher-order abstractions tend to exhibit one or more of the following traits:

1. They deal in infinities (e.g. integration and differentiation in calculus, universal and existential quantification in first-order logic, polymorphism and type classes in programming)
2. They deal in transformations (e.g. integration and differentiation in calculus, metaclasses and macros in programming)
3. They rely upon composition rules (e.g. chain rule and integration by parts in calculus, higher-order functions in programming)
4. Learning to use any given form of higher-order abstraction requires significant effort and discipline, therefore they tend to be the tools of specialists
5. They form the foundation of our modern, technological society.

The last two can be converted into conclusions:

1. Higher-order abstractions are critical for most interesting forms of science and technology
2. You can't expect a colleague or customer from another field to understand the higher-order abstractions in your field, even if his own field is totally immersed in higher-order abstraction

Sphere: Related Content

Pondering Actor Design Trades

There's been a lot of discussion of the Scala actors library lately, much of it critical, and a recent flurry of alternate implementations.  The alternate implementations (except my languishing state-based one ;-) all have one thing in common:  They are several orders of magnitude simpler.  Writing a basic actor implementation is actually pretty trivial, especially given java.util.concurrent classes that provide a decent chunk of the functionality in Scala actors, all for free on JDK5+.  So this begs the question few questions:

1. Why is the standard Scala actor implementation so complex when others have done it in a such simpler fashion?
2. Is it better to have one, big actor library that supports a wide variety of use cases, or a bunch of smaller ones targeted at specific niches and programming styles?
3. If there are to be a bunch, should they just be conceptually similar (e.g. all based on the actor model), or should there be interoperability among them?

I'm not going to answer these questions now.  Instead, I'm going to try to start laying out some of what I believe to be the key characteristics of an actor implementation, and how they detract or enforce one another.  So here it goes:

1. Guarantees
2. Expressivity
3. Extensibility
4. Performance
5. Scalability

Guarantees

The purpose of a concurrency framework is to make concurrency easier.  Concurrency is hard largely because it is extremely difficult to reason about, and thus concurrent code tends to be hard to write, laden with bugs, and subject to various odd pitfalls.  By providing various guarantees, a concurrency framework makes it easier to reason about concurrent code.  Actors are intended to free the programmer from worrying about things like locks, semaphores, thread management, etc. by encapsulating all that complexity behind a simple interface, assuming the programmer follows some basic rules like "no shared mutable state among actors."

The problem with guarantees is that in they tend to break down in the presence of limited CPU and memory resources.

Expressivity

Expressivity is difficult to define.  For purposes here, I'm going to define it as the degree to which a concise, natural expression of the programmer's intent is supported, and illustrate it by comparing Scala Actor to Lift Actor.  Scala Actors allow you to execute logic independent of message processing (note: this a violation of the theoretical model for actors) by simply placing it in the act method.  Lift Actors, on the other hand, are only triggered when they receive of message (this is consistent with the theoretical model).  For example, this makes it so that Scala Actors can do things such as perform possibly costly setup operations in their own thread before they start listening for messages.  In order to accomplish this in the Lift model, the programmer must create the actor and then send it some sort of "init" message.  The same effect can be achieved with both implementations, but it is more naturally supported by Scala Actors.  Of course there is a tradeoff here, as deviating from the theoretical model potentially weakens any guarantees that the model may provide.  The Scala Actor way also implies that an Actor has an explicit lifecycle, which as we'll see later has other significant implications.

Another example is what I'll call the "nested react pattern."  It is relatively common to want an actor to take on a different behavior after processing a message, thus altering which messages are ignored and how the received messages are processed.

loop {
 react {
    case 'foo => { 
      // do some stuff...
      react {
        case 'bar => // do some other stuff... 
      } 
    } 
  } 
}

The code above alternates between processing 'foo messages and 'bar messages.  This can be done with Lift Actor as well, but the expression is a little less natural:

class MyActor extends LiftActor {
  private val fooMode: PartialFunction[Any, Unit] = {
    case 'foo => {
      // do some stuff
      mode = barMode
    }
  }
  private val barMode: PartialFunction[Any, Unit] = {
    case 'bar => {
      // do some other stuff...
      mode = fooMode
    }
  }
  private var mode = fooMode
  protected def messageHandler = mode
}

Finally, Lift Actors exclusively use an event-based model and have no support for blocking on a thread while waiting for a message, and thus looses the ability to express patterns such as the following:

loop {
  react {
    case 'converToNumber => {
      val i: Int = receive {
        case 'one => 1
        case 'two => 2
        case 'three => 3
      }
      reply(i)
    }
  }
}

Extensibility

For purposes here, I'm going to use "extensible" to mean that a piece of software is extensible if capabilities can be added without modifying the core or breaking its semantics in a amount of effort proportional to the size of the extension.  This is narrower than the traditional definition of extensibility, which also covers the ability of a system to evolve internally.  A good example of extensibility is the ability of both Scala Actors and Lift Actors to allow the user to specify a custom scheduler.  Other examples could include adding control structures, using a different data structure for a mailbox.

The challenge with extensibility is that in order to enable it, what could otherwise be treated as the internal bits of the library must instead have well defined interfaces for components along with appropriate hooks for inserting them.  For example, a while ago I did some work to make the MessageQueue used for the mailbox overrideable (it has temporarily been overcome-by-events due to other changes).  This is a small example, but it shows how extensibility requires a greater degree of forethought.

Extensibility also benefits substantially from simplicity.  Scala Actors are almost impossible to extend from outside the scala.actors package because of their heavy reliance on package-private methods and state (mostly fixed here, but I broke remote actors in the process so no patch yet).  Lift Actors, on the other hand, are very extensible, at least within the bounds of their design (purely event-based actors with no explicit lifecyle).  Many of the flow control mechanisms could be implemented on top of the baseline approach.

At this point we see that extensibility has an interesting relationship with expressivity.  I previously claimed that Scala Actors were more expressive because the wide variety of control structures they provide (and I didn't even touch on some of the DSL-like functionality that enables all sorts of interesting things).  However, given Lift Actors far simpler and more extensible foundation, there is much more opportunity to create custom control structures as extensions to Lift Actors without modifying the core.  Thus, if you are willing to do some lower-level programming, it could be argued that Lift Actors are in reality more expressive due to their extensibility.

Performance and Scalability

For purposes here, I'm going to treat performance as the rate a which an actor can receive and process messages at a relatively small, fixed number of simultaneous actors.   This means that improving performance in largely a matter of reducing the time it takes from when a message is initially sent to when user-code within the actor begins processing the message, including minimizing any pause between when an actor finishes processing one message and is available to start processing the next.  For moderate numbers of actors, performance is often maximized by having one thread per actor, and having the actor block while waiting for a message.  Given enough actors, the memory requirements of using a thread for each actor will eventually cause more slowdown than cost of scheduling a new reaction for each message.  This is illustrated in Philipp Haller's paper, "Actors that Unify Threads and Events" in the following graph:

Note that the above graph covers a microbenchmark running a simple, non-memory intensive task, and that the thread line is not a measurement of thread-bound actors, but rather of a simple threaded implementation.  However, my own benchmarking has shown that receive-based (ones that block on a thread) compare to event-based actors in almost the same way as threads to event-based actors in the above graph.  Also, remember that given a real application where heap space is needed for things besides the stacks of thousands of threads the point where the JVM throws an OutOfMemoryError will be much farther to the left.  There are also more subtle issues.  One of my first experiences with the Scala Actors library was creating a deadlock.  I created more thread-bound actors than the scheduler wanted to create threads, and thus actors were stuck blocking on threads waiting for messages from an actor that hadn't started yet because there were no available threads.  In other words, blocking can lead to situations such as deadlock, starvation, and simply extreme forms of unfairness with respect to how much CPU time is allocated each actor.  These all go against highly desirable guarantees that a actor library should provide outside of extreme circumstances.

Ultimately event-based actors make the better model.  For one, part of the reason why event-based Scala Actors are so expensive is that they suspend by throwing an exception to return control from user code to the library.  While exceptions have been heavily optimized in the JVM, especially in recent versions, they are still substantially slower than normal return paths.  Scala Actors need to use exceptions to suspend is a consequence of their expressivity.  Basically, because the library as little or no knowledge of what an actor is doing within a reaction, it cannot rely on traditional returns without introducing special control structures (see reactWhile numbers in one of my previous blogs).  Lift Actors, on the other hand, have do not need to use exceptions for control flow because the message processing cycle is essentially fixed - user code cannot intersperse weird (or even not-so-weird) patterns within it, or mix in blocking receives with event-based ones.  Another potential optimization of event-based actors is to have them block if there are plenty of threads available, and then release it if the thread they are on is needed by the scheduler.  To my knowledge this optimization is not implemented anywhere, but I think it would be relatively straight forward.  The only problem is that the actor becomes more tightly bound to its scheduler.

Parting Thoughts

Ultimately, time and community willing, I'd like to evolve what is here, plus solid treatment of a lot of lower-level details, into a Scala Improvement Document (SID).  There are a lot of subtle trades involved, and I think producing a general-purpose actors library is at least an order-of-magnitude more difficult than producing a special-purpose one.  I also believe that if an actor implementation is part of the standard library, then it should provide the necessary extension points for when users need something special-purpose they can create it and still leverage components of the standard library and interoperate with other actors.  In order words, I think it should define both the interface portion of an API along with providing a solid implementation.  I don't think we'll even get their without a clear and common understanding of the various considerations involved.

Sphere: Related Content

I'm on Twitter!

For those of you with ADD or it's internet induced equivalents, I've started posting on Twitter. I long avoided it because I feel like the last thing people need is yet another half-baked information stream, but then people seem to like it so I'm giving it a shot. I'll post links to bugs, patches, and other comments regarding my efforts (and those of others) with Scala actors...along with other less important matters. http://twitter.com/ErikEngbrecht

Sphere: Related Content

Refactoring Scala Actors: Progress Update

It's been a while since I've posted, so I thought I'd give everyone a status update.  This post covers several different semi-disjoint topics at a fairly high level.  I plan on diving into some of the issues later this week and beyond, but for now...

State-machine Based Actors

A while back Philipp Haller, the original author and current maintainer of the Scala actors library, contacted and basically said he found the changes I was making really interesting, but he really needed a smaller, more gradual set of patches.  It's a perfectly reasonable request, as I had pretty much completely ripped apart his library.  I had rethought my approach, anyway, so I went about moving my state-machine based actor implementation into its own package and rewiring some of the pieces so that they could share common base traits, common infrastructure, and interoperate with one another as if they were the same library.  So I shoved my code into scalax.actors, and started hacking insertion points for my code into the main library.

The first thing I thought I needed was a base trait that defines the basic structure and operations of an actor, so created a BaseActor in between AbstractActor and Actor (as well as my own StateActor):

trait BaseActor extends AbstractActor {
  def react(f: PartialFunction[Any, Unit]): Nothing
  def reactWithin(msec: Long)(f: PartialFunction[Any, Unit]): Nothing
  def receive[A](f: PartialFunction[Any, A]): A
  def receiveWithin[R](msec: Long)(f: PartialFunction[Any, R]): R
  /*def loop(body: => Unit): Nothing */
  /*def loopWhile(cond: => Boolean)(body: => Unit): Nothing */
  protected[actors] def mailbox: MessageQueue[Message[Any]]
  private[actors] final def mailboxForChannel: MessageQueue[Message[Any]] = mailbox
  def mailboxSize: Int
  def send(msg: Any, replyTo: OutputChannel[Any]): Unit
  def forward(msg: Any): Unit
  def reply(msg: Any): Unit
  /*protected[actors]*/ def sender: OutputChannel[Any]
  def ? : Any
  def start(): AbstractActor
  def freshReplyChannel: Channel[Any] = new Channel[Any](this)
  def scheduler: IScheduler //TODO: restrict access to scheduler??
}

I don't think BaseActor is going to be a permanent fixture because its contents probably belong in AbstractActor instead, but for now it serves its purpose.  One of the first things you should notice is that way to much stuff in there is public.  Most of it should be protected, or perhaps somewhere else entirely (like an InputChannel encapsulated by the actor).

Reworking MessageQueue

There's also the issue of the mailbox, which is a rather important and a den of mutable data that is passed all around with private[actors] qualifiers.  Basically it separates the Message from the elements within the MessageQueue, so that the MessageQueue can keep its internal structure private, and thus facilitating making it a trait so that an actor can provide its own specialized implementation.  I was about to submit a patch for the change, but a fix for a memory leak in FJTaskRunner came about that relied on clearing mutable fields in the message when a task is done processing.  I have an alternative fix by changing pieces of FJTaskScheduler2, but schedulers in general and FJTaskScheduler2 in specific are in flux right now due to bugs (here and here and probably elsewhere), and I want to tweak the design a bit, so I'm holding off.

Fixing Schedulers

Which brings me to schedulers...  Problems with plugging in custom schedulers (mostly fixed) are what originally caused me to dive into the guts of the actor library.  Closely related to schedulers is ActorGC, which is absolutely essential to actors (almost) transparently abstracting threads, but can also be problematic due to it's fundamentally non-deterministic nature (it relies on the garbage collector for some of its more advanced capabilities).  That being said, now in trunk ActorGC is optional, so environments that don't require an implicit shutdown of the actor worker threads can avoid the added complexity.  I intend to cover the details of ActorGC very soon.  There should also be a default scheduler with daemon semantics coming, which has a number of use cases.

Closing Matter

There's a lot more going on.  Some recent flare-ups on Scala Internals mailing, despite being a tad melodramatic, brought a welcome focus on actors for the next release of Scala.  The issue has also given rise to two minimalistic actor implementations, one in Lift and the other in Scalaz.  They both make interesting data points for design and potential interoperability (remember: one of my primary goals is an actor implementation that lets you plug in what you need).  There's issues around ActorProxy that I think will be a little hairy to sort out, but I'm confident they will be.  And finally, there's the omnipresent issue of ensuring actor's really make the guarantees that they claim (right now I think they do, but I wouldn't place money on it until I have tests to prove it).

That's it for now.  I'm going to try to take the time to blog about many of the above issues and more in depth in the coming weeks, and hopefully gain some insights from out in the cloud.

Sphere: Related Content

Refactoring Scala Actors: Rethinking the Approach

When I started refactoring Scala's actor library, I really had several goals:

1. Reduce the coupling within the library so that the implementation can be more easily extended and customized
2. Create a more transparent and "programmer friendly" actor implementation
3. Improve performance of various use cases
4. Make actors that interact better with their environment, particularly non-actor code
5. Maintain API compatibility with the existing library to the maximum extent practical

Thus far I've done my work by completely overhauling the Actor trait directly in the library.  While I have plans for how this will reduce coupling, the haven't come to fruition yet.  I think my state machine model is considerably more transparent and programmer friendly than the current implementation.  The state an actor in is always clear, transition are reasonably well defined, and performing correct locking and unlocking is now pretty straight forward.  I've substantially improved the performance of event-based actors for the common case where an actor loops while it receives messages until some termination criterion is reached.  I haven't done anything with making them interact better with their environment yet, as I believe Philipp Haller is in the process of incorporating some changes for which I submitted patches several months back that will help considerably (he doesn't appear to be using the patches directly, but the changes I've seen are close enough).

A few days ago David MacIver asked me a couple interesting questions on IRC:

1. Have you considered using Kilim?
2. Have you looked at the continuations support that appears to be emerging as a compiler plugin?

Using Kilim would almost certainly disqualify the changes I'm making from incorporation into the standard library because of the external dependency on a prerelease framework that uses bytecode manipulation, and I don't think the fledgling continuation support is mature enough to experiment with yet (or maybe just not well documented enough yet, who knows...).  That being said, both of these would be interesting avenues of development.  Event based actors rely heavily on continuations, and the performance of those continuations has a substantial effect on the performance of the actors.  Ultimately a properly decoupled implementation would allow someone to build an API compatible actor implementation on either of these technologies, or something else entirely.

I also received a recent nudge on the mailing list, when someone pointed out that it would be easier to experiment with my library if it was in it's own package.  I somewhat disagree with that.  It would be easier to experiment with it if you didn't have to do an entire build of the Scala distribution that has my code in it, and unfortunately for the foreseeable future I'm going to be depending on both code that is being committed into the trunk and on the ability to modify said code.  Also, the way I have it setup today, if someone wanted to test their Scala application/library/framework against my state-based actors, they would just have to pull my Mercurial repository, build it, and rebuild against the resulting distribution.  On the downside, it's a pain to do side-by-side comparisons between the two implementations, because you have to switch distributions and rebuild every time.

That being said, decoupling is one of my primary goals, and Martin & Company have already set the precedence of doing major work in a separate package with their efforts on redesigning the Scala collections library.  So as soon as I finish up some of my immediate tasks and have a good build again (I'm in the middle of redesigning the control flow pattern to minimize the stack size when exceptions are thrown), I'm going to push most of the public and protected interface on Actor into a new trait that will be between Actor and AbstractActor as abstract methods, move my code out of scala.actors and into scalax.actors in my own directory.  I definitely keep it within the same repository as the Scala distribution code, and will probably just make another directory for it.  This means I'll have to mess with Sabbus to add my portion to the build, which won't be fun, but shouldn't be too hard.  I'm sure there's going to be a lot more to do to extract it out, so I'll be adding items to my issues list as I think of them.

The end result should be my state based actors and the existing actors being able to live side-by-side and interact with one another in the same program.  Assuming I can at least get some patches into the main library, it will also mean that the future of my work will not be dependent on being incorporated into the standard distribution.  If it is, great, if not, I can distribute it separately.  I would have done this from the start, but initially I was highly dependent on the existing code and infrastructure into order to get something working reasonably quickly, and to be able to smoke it out to see if it indeed worked.  Back in December I was basically breaking partest with every change.  But now things are reasonably stable, I rarely break partest, and I don't think it will take much for my code to be able to stand on its own.

So what does everyone think?  Is this a good direction to head in?

Sphere: Related Content

Profiling the ExceptionBlob

Last week on the Scala IRC channel (#scala) I chatted with several people about the need to substantially reduce the use of exceptions for flow control within the Scala Actors library.  My remarks were met with a certain amount of healthy skepticism, as significant progress has been made on improving the performance of Java exceptions, as stated by John Rose in his blog Longjumps Considered Inexpensive.

I highly suggest you go read the post.  I don't doubt anything stated in it, and in fact noticed a substantial performance improvement when I stopped benchmarking against Java 5 and started doing it against Java 6 (I usually work on a Mac, and Java 5 is the default VM), and expect even more improvements with Java 7.  Once you're done, I'd like to call your attention to a couple fragments of it (additional emphasis mine):

The Hotspot JIT can optimize a throw of a preallocated exception into a simple goto, if the thrower and catcher are together in the same compilation unit (which happens because of inlining).

Here is a concrete example of non-local return, implemented efficiently via a cloned exception. I have observed the Hotspot server compiler emitting a machine-level goto for the throws. With current server VM optimizations of this code, the cost of a non-local return is less than three times the cost of a plain (local) return; with the client VM the ratio is ten. See the code example and benchmark for details: NonLocalExit.java (javadoc).

These statements have a couple implications:

1. Reducing a throw to a long jump requires a certain amount of HotSpot magic take place to ensure that the throw and catch are in the same compilation unit.  The more complex the code is, the less likely this is to happen.
2. Long jumps are still substantially slower than normal returns

I think John's post received a lot of attention, and people (including myself) saw it, didn't read it carefully, and assumed that HotSpot now performed miracles with cached exceptions.   What we missed was a post to the MLVM development mailing list about a year later by Charles Nutter, who had noticed that his non-local returns in JRuby where not being compiled down into long jumps, and subsequent response from John Rose.  There's a lot of very good technical discussion in there, but I think the gist is that HotSpot's ability to optimize throws into long jumps is still limited, it often needs help to do so, and very smart people are working at improving it.

Of course that all still leaves the question:  Just how much slower are cached exceptions than a normal return?  I can't really answer that in the general case, but I did put together a couple mircobenchmarks in Java that somewhat simulate the type of stack creation that happens within my state-based actors.  The benchmark is pretty simple.  It recursively descends down, incrementing a member variable, and then on the way back up it decrements another member variable.  In the case of the exception-based return, it does the decrements within a finally block.  This admittedly is rather pathological code, but it is also simple and easy to understand.

Here's the normal return version:

public class NormalReturn {
    public static void main(String[] args) {
        NormalReturn m = new NormalReturn();
        m.doTest(10000000);
    }
    public void doTest(int times) {
        for(int i = 0; i < times; i++) {
            down = 0;
            up = 0;
            deepRecurse(1000);
        }
        System.out.println(down);
        System.out.println(up);
    }
    private int down = 0;
    private int up = 0;
    public void deepRecurse(int i) {
        down++;
        if (down <= i) {
            deepRecurse(i);
            up--;
        }
    }
}

...and here's the version that uses a cached exception:

public class TryFinally {
    private static class Signal extends RuntimeException {
        @Override
        public Signal fillInStackTrace() { return this; }
    }
    private Signal signal = new Signal();
    public static void main(String[] args) {
        TryFinally m = new TryFinally();
        m.doTest(10000000);
    }
    public void doTest(int times) {
        for(int i = 0; i < times; i++) {
            try {
                down = 0;
                up = 0;
                deepRecurse(1000);
            } catch (Signal s) {
                // do nothing
            }
        }
        System.out.println(down);
        System.out.println(up);
    }
    private int down = 0;
    private int up = 0;
    public void deepRecurse(int i) {
        try {
            down++;
            if (down == i) throw signal;
            else deepRecurse(i);
        } finally {
            up--;
        }
    }
}

With profiling turned off on and -server on Java 6 on my Mac, the normal return version takes about 60 seconds to complete, while the TryFinally version takes over 10 minutes.  That's certainly much better than the 30x figure John Rose cited for the client VM, but it's still an order-of-magnitude worse than a normal return.  One interesting thing about the benchmark is that if you decrease the recursion depth and increase the number of times deepRecurse() is executed the total time stays relatively constant.

So with this in mind, I'm going to work hard to refactor the actors library to keep exception usage to a minimum, an where it must be used to minimize the depth of the stack.

Sphere: Related Content

Scala Actors versus the ExceptionBlob

I've begun some very preliminary benchmarking of my state-based actors library against the current standard actors library and seen some rather interesting results.  The test I'm running spawns four pairs of actors and sends about 40 million messages between each pair.  The only logic in the actors is to count how many messages they have received, ensure they are received in order, and occasionally send back a "More" message.  This is so the mailbox doesn't overfill and cause out-of-memory errors.

The benchmark has three different incarnations, each testing a different way of processing messages until some termination condition occurs.  The versions are:

• loop
• recursive react
• reactWhile (only available with the State-based actors, but relatively easy to port to "classic" actors)

The code for the loop version is below:

class LoopPingSender(val totalPings: Int, val pingReceiver: LoopPingReceiver) extends Actor {
  private var pingsSent = 0
  def act() {
    Actor.loop {
      pingReceiver ! Ping(pingsSent)
      pingsSent += 1
      if (pingsSent == totalPings) {
        println("Sender Done")
        exit("sender completed successfully")
      } else if ((pingsSent % 1000) == 0) react {
        case More => 
      }
    }
  }
}
class LoopPingReceiver(val totalPings: Int) extends Actor {
  private var pingsReceived = 0
  def act() {
    Actor.loop {
      react {
        case Ping(n) if n == pingsReceived => {
          pingsReceived += 1
          if (pingsReceived == totalPings) {
            println("Receiver Done")
            exit("receiver completed successfully")
          } else if ((pingsReceived % 1000) == 0) {
            reply(More)
          }
        }
        case Ping(n) => {
          println("error")
          exit("invalid n value: " + n)
        }
      }
    }
  }
}

The non-scientifically gathered results are as follows, all times in seconds:

Version	Standard Actors	State-based Actors
loop	298.22	96.84
Recursive react	90.25	113.57
reactWhile	N/A	40.63

Standard Actor loop

Let's consider the loop version first.  loop in the standard actor implementation is rather convoluted.  Rather than directly supporting looping, the standard actor implementation provides a package private hook for a function that is executed when the actor is terminating.  loop uses this hook to subvert the termination process by throwing an exception and subsequently restarting execution of the actor.  This makes it so that two exceptions are thrown per message processed rather than the normal one.  Here's a small snippet from the profiling results:

     Compiled + native   Method                        
 40.0%  5148  +     0    ExceptionBlob
 13.9%   658  +  1133    scala.actors.Scheduler$$anon$2.run
 13.1%   391  +  1299    scala.actors.Actor$.loop
  3.4%   441  +     0    scala.actors.FJTaskRunner.run
--- snip ---
 73.7%  6705  +  2771    Total compiled

         Stub + native   Method                        
 20.6%     0  +  2654    java.lang.Thread.isInterrupted

What you can see is that the ExceptionBlob is dominating the runtime, followed by Thread.isInterrupted.  ExceptionBlob is a class in the JVM that observes the unwinding of the stack.  As far as I could discern through so separate microbenchmarking (which I'll discuss another day, after I've spent some time in HotSpot code), the amount of time consumed by ExceptionBlob is proportionate depth of the stack being unwound by exceptions.  In other words, throwing fewer exceptions doesn't do any good if you use them to unwind deeper stacks, and unwinding shallower stacks doesn't do any good if you throw more exceptions.  Caching exceptions doesn't help, either, but I digress.  The bottom line is that loop is pretty inefficient in the standard actors implementation.

State-based Actor loop

In my state-based actors, there are two states to handle looping, both of which are subclasses of Running.  The first is Looping, which is initially entered when the actor starts looping.  The second is ResumeLooping, which is used to resume looping after the actor has been suspended.  Both run the body of the react in a tail-recursive loop until their are no more messages to process, at which time they suspend the actor.  This means that many message can potentially be processed before any exception-based signaling is used.  The following is a snippet from a profile:

     Compiled + native   Method                        
 67.5%  3233  +     0    ExceptionBlob
  6.4%   306  +     0    scala.actors.Actor$class.withLock
  4.9%    44  +   189    actorbench.LoopPingSender$$anonfun$act$1.apply
  4.8%    97  +   135    scala.actors.Actor$$anonfun$send$1.apply
  4.5%   116  +    99    scala.actors.Actor$Running.tr$1
  3.8%    36  +   145    scala.actors.Actor$ResumeLooping.enter
--- snip ---
         Stub + native   Method                        
  0.2%     0  +     9    java.lang.Thread.isInterrupted

Again, ExceptionBlob dominates the execution time, although the total time it consumes is lower.  Unfortunately, despite using significantly fewer exception signals, my state-based actors tend to build up significantly deeper stacks than the actors in the standard library.

Standard Actor recursive react

Recursive react seems to be the best performing looping pattern for the standard actor library.  It's a pretty simple pattern:

class PingSender(val totalPings: Int, val pingReceiver: PingReceiver) extends Actor {
  def act() {
    def recurse(i: Int): Unit = {
      if (i > 0) {
        pingReceiver ! Ping(i)
        if ((i % 1000) == 0) {
          react {
            case More => recurse(i - 1)
          }
        } else {
          recurse(i - 1)
        }
      } else {
        println("Sender Done")
      }
    }
    recurse(totalPings)
  }
}

As you can see, the function containing the react block is simply recursively called as needed.  Note that despite the fact that the recursive call appears to be in a tail position, it is not a tail recursive call and will not be optimized by the compiler.  The PartialFunction defined within the react block is actually a separate object with an apply method, so while recurse and the PartialFunction are mutually recursive, they are not tail recursive.

     Compiled + native   Method                        
 19.1%   456  +     0    ExceptionBlob
 13.0%    89  +   221    scala.actors.Reaction.run
  8.6%    53  +   152    scala.actors.Actor$class.send
  7.9%    61  +   127    scala.actors.Actor$class.react
  5.3%    59  +    67    scala.Iterator$class.toList
--- snip ---
 68.1%   862  +   765    Total compiled

         Stub + native   Method                        
 26.4%     0  +   631    java.lang.Thread.isInterrupted
--- snip ---
 26.9%     0  +   643    Total stub

This time Thread.isInterrupted dominates, with ExceptionBlob not too far behind.

Actor States recursive react

It turns out that the "hybrid trampoline" I implemented to facilitate recursive reacts without giving up the thread provides pretty bad performance.

 64.9%  3714  +     0    ExceptionBlob
 11.4%   568  +    85    scala.actors.Actor$class.withoutLock
  8.4%   478  +     1    scala.actors.Actor$class.withLock
  5.1%    90  +   201    scala.actors.Actor$$anonfun$reactWithin$1.apply
  4.2%   115  +   125    scala.actors.Actor$$anonfun$send$1.apply
--- snip ---
 97.8%  5006  +   595    Total compiled
         Stub + native   Method                        
  0.1%     0  +     6    java.lang.Thread.currentThread
--- snip ---
  0.2%     0  +    12    Total stub

The problem is that when I wrote it, I thought that the cost of throwing exceptions was primarily proportionate to the number of exceptions thrown rather than the amount of stack that is unwound. Consequently, I made the trampoline only bounce once every n invocations in order to reduce the number of exceptions being thrown. As it turns out, what I should do is find a place where the stack is likely to be the shallowest, and throw an exception every time.

Actor States reactWhile

reactWhile addresses the pitfalls presented in the previous versions and provides roughly twice the performance of next best performer, recursive react on the standard actor library.

     Compiled + native   Method                        
 30.0%   171  +   331    scala.actors.Actor$Running.tr$1
 24.4%   263  +   146    actorbench.ReactWhileReceiver.withLock
 12.4%    26  +   181    actorbench.ReactWhileSender.sendPings
  8.5%   142  +     0    scala.actors.Actor$class.withoutLock
--- snip ---
 79.0%   660  +   661    Total compiled

         Stub + native   Method                        
 15.1%     0  +   252    java.lang.Thread.isInterrupted
  0.9%     0  +    15    java.lang.Thread.currentThread
 16.0%     0  +   267    Total stub

ExceptionBlob has completely disappeared from the profile. That is because reactWhile makes minimal use of exceptions for signaling.  Consequently, its performance is dominated by some of the complicated for my (worthless) hybrid trampoline, my completely unoptimized lock implementation, and calls to Thread.isInterrupted that are made when the bound thread is returned to the pool when the actor suspends.  I'll write a blog on how reactWhile works at a later date, after I cleanup some of the other problems that I've listed.

Parting Matter

It really annoys me when people post benchmark results without providing details of their setup along with all of the code required to duplicate them, and that's just what I've done.  I've also benchmarked a very narrow set of behavior, failed for warm up the JVM, and not run multiple trials.  Developing a lightweight framework for benchmarking actor performance, along with a more comprehensive set of benchmarks is on my to-do list.  Once I have that, I'll start doing more rigorous benchmarking.  But I don't think I quite need it yet because bottlenecks are showing up quickly and consistently in my profiling, and in most cases they seem to have reasonably straight forward solutions.  All the above profiles against my library were done against revision 49658b3475e2 in my bitbucket repository, so you can at least access that code.

Sphere: Related Content