Scalable Performance for Scala Message-Passing Concurrency

Transcription

1 Scalable Performance for Scala Message-Passing Concurrency Andrew Bate Department of Computer Science University of Oxford cso.io

2 Motivation Multi-core commodity hardware Non-uniform shared memory Expose potential parallelism Correctness and formal verification int arr[x][y]; Compatibility

3 Embedded DSL Bytecode rewriting Channels Scheduler Deadlock detection EMBEDDED DOMAIN-SPECIFIC LANGUAGE

4 Why an Embedded DSL? Ease of implementation Leverage existing tools Leverage known syntax Higher-order functions Rich type system Lightweight syntax Compile-time macros

5 Examples v in map ff out ff(v) def map[i, O](f: I => O)(in:?[I], out:![o]) = proc { } repeat { out! (f(in?)) } run (proc { in.closein } proc { out.closeout })

6 Examples out 1 v out 2 v v in tee out n v def tee[@specialized T](in:?[T], outs: Seq[![T]]) = proc { } var v = null val outputs = ( (out <- outs) proc { out! v })) repeat { v = in?; run outputs } run (proc { in.closein } ( (out <- outs) proc { out.closeout }))

7 1 Embedded DSL 2 Bytecode rewriting 3 Channels 4 Scheduler 5 Deadlock detection SCALABLE PERFORMANCE through bytecode rewriting

8 CPS Transformation Init Prelude Init rewinding Pre-call Call n() Call n(f) Post-call pausing Return Return

9 Analysing the call graph Transform these methods x() do() y()?() z()

10 Engineering Live variable analysis Lazy load and store Constant inlining

11 Functional Expressions for (i <- 0 until n; j <- i until n) println(i) Compiles to intwrapper(0).until(n).foreach( ) i: Int => intwrapper(i).until(n).foreach(j: Int => println(i)) Transforms to var i = 0 while (i < n) { var j = i while (j < n) { println(i); j += 1 } i += 1 }

12 More Features Tail call optimisations Shared memory SBT plugin support

13 Embedded DSL Bytecode rewriting Channels Scheduler Deadlock detection CHANNELS

14 More Features Generalised alt Specialization for primitives Optimised extended rendezvous

15 Embedded DSL Bytecode rewriting Channels Scheduler Deadlock detection SCHEDULER

16 Scheduler States Created Terminated Running Waiting Paused

17 Scheduling: Central FIFO Scheduler PP 1 PP 2 PP 3 PP nn Thread 1 Thread 2 Thread mm

18 Scheduling: FIFO per thread Scheduler PP 1 Scheduler PP 3 Scheduler PP nn Thread 1 Thread 2 Thread mm

19 Scheduling: Batches per thread Scheduler? Scheduler Scheduler?? Thread 1 Thread 2 Thread mm

20 Scheduling: Batches per thread Scheduler Dispatch Count = max const Batch Length, Dispatch Limit PP 1 PP 2 PP nn QQ 1 QQ mm RR 1 RR kk

21 Embedded DSL Bytecode rewriting Channels Scheduler Deadlock detection DEADLOCK DETECTION

22 Example Tee Tee x5 x2 x3 Merge Merge Tee Prefix 1 Console

23 Example Tee!! Tee x5! x2?? x3 Merge!! Merge Tee? Console! Prefix 1! Deadlock detected! The cycle of ungranted requests is: Prefix1 -!-> Tee1 Tee3 -!-> x5 Tee1 -!-> Tee2 x5 -!-> Merge2 Tee2 -!-> Tee3 Merge2 -!-> Prefix1

24 Embedded DSL Bytecode rewriting Channels Scheduler Deadlock detection PERFORMANCE EVALUATION

25 Ring topology Time to pass a message 300 times around an n process ring (ms) Java primitives CSO 2 FIFO Scheduler CSO 2 Batch Scheduler 100 Number n of processes spawned

26 Ring topology Java primitives Time to pass a message 300 times around an n process ring (ms) CSO 2 FIFO Scheduler CSO 2 Batch Scheduler 0 Number n of processes spawned

27 Fully connected topology Time to pass n 2 messages (ms) Erlang Scala Actors JCSP Java Primitives Occam CSO2 2 FIFO Scheduler CSO2 2 Batch Scheduler Go Number n of processes / actors spawned

28 Fully connected topology Time to pass n 2 messages (ms) Erlang Scala Actors JCSP Java Primitives Occam CSO2 2 FIFO Scheduler CSO2 2 Batch Scheduler Go Number n of processes / actors spawned

29 Fully connected topology JCSP Time to pass n 2 messages (ms) Occam CSO2 2 FIFO Scheduler CSO2 2 Batch Scheduler Go Number n of processes / actors spawned

30 Fully connected topology CSO 2 FIFO Scheduler Time to pass n 2 messages (ms) Go CSO 2 Batch Scheduler 0 Number n of processes / actors spawned

31 Summary High performance library for building massively concurrent systems on the JVM Deadlock detection Outperforms Java primitives, JCSP, Scala Actors, Occam, and very close to Go