Testing Fiber Implementations in JVM
posted at August 7, 2014 with tags akka, java, quasar

The heart of the problem lies in the question of mapping $m$ light-weight threads to $n$ physical processor cores ($m \gg n$) given that the operating system has no prior knowledge on what is going on at the application level. As an inevitable consequence of this problem, developers come up with the idea of providing application level constructs to hint the underlying process scheduling mechanism about the concurrency operating at the application level. Among many other attempts, coroutines and fibers are the two well-known examples emanated from this research. Quoting from Wikipedia,

Fibers describe essentially the same concept as coroutines. The distinction, if there is any, is that coroutines are a language-level construct, a form of control flow, while fibers are a systems-level construct, viewed as threads that happen not to run concurrently. Priority is contentious; fibers may be viewed as an implementation of coroutines, or as a substrate on which to implement coroutines.

If I would try to restate the given description from a more naive perspective targeting the Java Virtual Machine, coroutines are language-level constructs to form a way of cooperative execution of multiple tasks, whereas fibers are light-weight threads scheduled preemptively by the virtual machine. There is a set of coroutine implementations available for JVM, but in this post I will try to stick with two particular fiber implementations: Akka from Typesafe and Quasar from Parallel Universe.

# Akka and Quasar

Last year I had a plenty amount of time to play around with Akka during the The Principles of Reactive Programming course. Claiming that Akka is just a toolkit to realize fibers on the JVM would be a fairly incomplete definition – it is totally much more than that. Designed for resiliency from top to ground, a gigantic toolbox of common messaging pattern shortcuts, adaptive load balancing, routing, partitioning, and configuration-driven remoting, etc. just to name a few. Additionally, one should note that it is battle-tested with real-world work loads by means of both as a standalone application and a compound to the Typesafe Platform.

On the other hand, Quasar is a relatively new player in this league. That being said, it is implemented as a compact library with – in my opinion – a relatively sufficient set of features. Though it has a lot to go to catch Akka in terms of out of the box features. In addition, while Akka just provides actor abstractions at the programming language level, Quasar ships both actors and fibers. (Actually, Quasar actors are built on top of fibers.) Quoting from Quasar core developer Ron Pressler’s post:

Quasar and Akka provide completely different abstractions. Quasar provides fibers while Akka provides an asynchronous-programming framework based on callbacks. Both libraries then implement actors using their respective abstractions. […] fibers are lightweight threads, which means that their most important property – even before being light-weight – is being threads. The thread abstraction is a series of sequential computations with the ability to block – on IO or synchronization. Callbacks, on the other hand, are a completely different abstraction. Akka’s callbacks (and therefore actors), are not allowed to block, and therefore do not provide a thread abstraction; hence they’re not fibers.

Quasar’s main design goal was the desire to make the simple, familiar, threading abstraction (i.e., synchronous, blocking code) cheap, as kernel threads are expensive, and blocking a kernel thread carries a lot of overhead. Quasar is used to run servlets that serve hundreds of thousands of concurrent requests – instead of mere thousands – all without changing your existing Java code. We want developers get the full performance benefits of asynchronous code – which Akka also offers – while keeping their APIs and synchronous programming model – which Akka certainly doesn’t. Quasar makes threading cheap, while Akka abandons the thread abstraction altogether. In that respect, Quasar is a lot more similar to Erlang and Go (which, of course, provided the inspiration to Quasar), which also use the thread abstraction but provide a lightweight thread implementation.

Akka’s design goals are different, and we think Quasar is far simpler to use than Akka, requires less mental overhead and far-less re-engineering, while at the same time being more feature-rich.

# The Benchmark

In order to have an idea on how two implementations compare to each other in terms of performance, I put together a small benchmark suite (fiber-test) that employs both Akka (2.3.5) and Quasar (0.6.0) for the well-known thread-ring problem. ($n$ threads are spawned and connected as a ring structure. Through this ring a message – an integer comprising 4 bytes – is circulated involving $m$ message passings.) In addition, I included a version of the benchmark that uses Java Threads to establish a base line.

In fiber-test, you will observe that benchmark.sh employs taskset to bond the JVM to a given set of CPUs on the system. However, for small number (1-2) of CPUs, Quasar freezes randomly and Java Threads become totally unresponsive. Hence, I get rid off of the taskset for the figures presented here.

Tests are deployed on a Ubuntu GNU/Linux 14.04 (LTS) system running on a 6 physical core Intel(R) Xeon(R) E5645 2.40GHz processor, where 64-bit Java HotSpot VM (build 1.8.0_20-ea-b05) is used. JMH is set to run 5+5 warmup and regular iterations using 3 JVM forks. There are 503 threads in each benchmark and 1e7 message passings. Results are in milliseconds per benchmark.

Fiber Impl. Min. Avg. Max. Stddev. Confidence Interval
Quasar Fibers 695.613 29x 721.457 758.231 20.796 99.9% 699.226 743.689
Akka Actors 856.807 23x 911.295 963.403 34.345 99.9% 874.578 948.012
Quasar Actors 1553.224 13x 1606.718 1660.329 35.756 99.9% 1568.492 1644.943
Java Threads 15730.028 1x 20709.233 33084.117 4338.423 99.9% 16071.196 25347.270

Results point out that even in the worst case (which corresponds to Quasar Actors in the figures) fiber implementation on the average performs 13 times better compared to native Java threads. In addition, Akka Actors and Quasar Fibers improve this level up to 23x and 29x, respectively. All that being said and done, I could not reason about the performance difference between Quasar Fibers and Actors, given that actors are implemented using fibers in Quasar. In order to further investigate the problem, I profiled the benchmarks using -agentlib:hprof=cpu=samples,interval=20,depth=3. (See HPROF documentation for further information on the parameters.)

rank self accum count trace method
Quasar Fibers
1 47,55% 47,55% 67286 300366 co.paralleluniverse.fibers.Fiber.run1
2 47,48% 95,04% 67186 300364 com.github.vy.fibertest.QuasarFiberRingBenchmark$InternalFiber.run 3 3,37% 98,40% 4765 300372 com.github.vy.fibertest.QuasarFiberRingBenchmark$InternalFiber.run
Quasar Actors
1 22,17% 22,17% 77787 300526 co.paralleluniverse.fibers.Fiber.park1
2 18,53% 40,69% 65011 300508 co.paralleluniverse.fibers.Fiber.run
3 18,32% 59,01% 64292 300527 co.paralleluniverse.strands.channels.SingleConsumerQueueChannel.receive
4 18,27% 77,28% 64114 300517 com.github.vy.fibertest.QuasarActorRingBenchmark\$InternalActor.doRun
5 17,39% 94,67% 61009 300528 co.paralleluniverse.actors.ActorRunner.run
Akka Actors
1 64,09% 64,09% 5158 300307 scala.concurrent.forkjoin.ForkJoinPool.runWorker
2 12,51% 76,60% 1007 300298 sun.misc.Unsafe.park
3 10,69% 87,29% 860 300309 sun.misc.Unsafe.unpark
4 2,65% 89,94% 213 300300 scala.concurrent.forkjoin.ForkJoinPool.scan
5 1,12% 91,05% 90 300316 akka.dispatch.Dispatcher.dispatch