Checkpoint synchronization
The checkpoint synchronization is a problem of synchronizing multiple tasks. Consider a workshop where several workers (tasks) assembly details of some mechanism. When each of them completes his work they put the details together. There is no store, so a worker who finished its part first must wait for others before starting another one. Putting details together is the checkpoint at which tasks synchronize themselves before going their paths apart.
You are encouraged to solve this task according to the task description, using any language you may know.
The task
Implement checkpoint synchronization in your language.
Make sure that the solution is race condition-free. Note that a straightforward solution based on events is exposed to race condition. Let two tasks A and B need to be synchronized at a checkpoint. Each signals its event (EA and EB correspondingly), then waits for the AND-combination of the events (EA&EB) and resets its event. Consider the following scenario: A signals EA first and gets blocked waiting for EA&EB. Then B signals EB and loses the processor. Then A is released (both events are signaled) and resets EA. Now if B returns and enters waiting for EA&EB, it gets lost.
When a worker is ready it shall not continue before others finish. A typical implementation bug is when a worker is counted twice within one working cycle causing its premature completion. This happens when the quickest worker serves its cycle two times while the laziest one is lagging behind.
If you can, implement workers joining and leaving.
Ada
<lang Ada>with Ada.Calendar; use Ada.Calendar; with Ada.Numerics.Float_Random; use Ada.Numerics.Float_Random; with Ada.Text_IO; use Ada.Text_IO;
procedure Test_Checkpoint is
protected Checkpoint is entry Deliver; entry Join (Label : out Character); entry Leave; private Signaling : Boolean := False; Ready_Count : Natural := 0; Worker_Count : Natural := 0; Unused_Label : Character := 'A'; entry Lodge; end Checkpoint;
protected body Checkpoint is entry Join (Label : out Character) when not Signaling is begin Label := Unused_Label; Unused_Label := Character'Succ (Unused_Label); Worker_Count := Worker_Count + 1; end Join; entry Leave when not Signaling is begin Worker_Count := Worker_Count - 1; end Leave;
entry Deliver when not Signaling is begin Ready_Count := Ready_Count + 1; requeue Lodge; end Deliver;
entry Lodge when Ready_Count = Worker_Count or Signaling is begin Ready_Count := Ready_Count - 1; Signaling := Ready_Count /= 0; end Lodge; end Checkpoint; task type Worker;
task body Worker is Dice : Generator; Label : Character; Shift_End : Time := Clock + 2.0; -- Trade unions are hard! begin Reset (Dice); Checkpoint.Join (Label); loop Put_Line (Label & " is working"); delay Duration (Random (Dice) * 0.200); Put_Line (Label & " is ready"); Checkpoint.Deliver; exit when Clock >= Shift_End; end loop; Checkpoint.Leave; end Worker;
Set : array (1..4) of Worker;
begin
null; -- Nothing to do here
end Test_Checkpoint;</lang> Sample output:
A is working B is working C is working D is working A is ready C is ready B is ready D is ready D is working C is working A is working B is working D is ready B is ready A is ready C is ready C is working B is working D is working A is working D is ready A is ready C is ready B is ready D is working A is working C is working B is working A is ready D is ready B is ready C is ready A is working D is working C is working B is working A is ready D is ready C is ready B is ready A is working C is working D is working B is working A is ready C is ready D is ready B is ready A is working D is working C is working B is working D is ready C is ready B is ready A is ready A is working D is working C is working B is working D is ready C is ready B is ready A is ready D is working C is working B is working A is working C is ready B is ready D is ready A is ready C is working B is working D is working A is working C is ready B is ready A is ready D is ready B is working A is working D is working C is working B is ready A is ready D is ready C is ready C is working D is working B is working A is working C is ready D is ready B is ready A is ready C is working D is working B is working A is working C is ready B is ready D is ready A is ready C is working B is working D is working A is working C is ready D is ready B is ready A is ready C is working D is working A is working B is working C is ready D is ready B is ready A is ready C is working D is working A is working B is working C is ready B is ready A is ready D is ready B is working A is working D is working C is working B is ready D is ready C is ready A is ready
E
The problem as stated is somewhat unnatural in E. We would prefer to define the control flow in association with the data flow; for example, such that the workers return values that are combined at the checkpoint; the availability of that result value naturally defines when the workers should proceed with the next round.
That said, here is an implementation of the task as stated. We start by defining a 'flag set' data structure (which is hopefully also useful for other problems), which allows us to express the checkpoint algorithm straightforwardly while being protected against the possibility of a task calling deliver
or leave
too many times. Note also that each task gets its own reference denoting its membership in the checkpoint group; thus it can only speak for itself and not break any global invariants.
<lang e>/** A flagSet solves this problem: There are N things, each in a true or false
* state, and we want to know whether they are all true (or all false), and be * able to bulk-change all of them, and all this without allowing double- * counting -- setting a flag twice is idempotent. */
def makeFlagSet() {
# Each flag object is either in the true set or the false set. def trues := [].asSet().diverge() def falses := [].asSet().diverge() return def flagSet { /** Add a flag to the set. */ to join() { def flag { /** Get the value of this flag. */ to get() :boolean { } /** Set the value of this flag. */ to put(v :boolean) { def [del,add] := if (v) { [falses,trues] } else { [trues,falses] } if (del.contains(flag)) { del.remove(flag) add.addElement(flag) } } /** Remove this flag from the set. */ to leave() :void { trues.remove(flag) falses.remove(flag) } } falses.addElement(flag) return flag } /** Are all the flags true (none false)? */ to allTrue() { return falses.size().isZero() } /** Are all the flags false (none true)? */ to allFalse() { return trues.size().isZero() } /** Set all the flags to the same value. */ to setAll(v :boolean) { def [del,add] := if (v) { [falses,trues] } else { [trues,falses] } add.addAll(del) del.removeAll(del) } }
}
def makeCheckpoint() {
def [var continueSignal, var continueRes] := Ref.promise() def readies := makeFlagSet() /** Check whether all tasks have reached the checkpoint, and if so send the * signal and go to the next round. */ def check() { if (readies.allTrue()) { readies.setAll(false) continueRes.resolve(null) # send the continue signal def [p, r] := Ref.promise() # prepare a new continue signal continueSignal := p continueRes := r } } return def checkpoint { to join() { def &flag := readies.join() return def membership { to leave() { (&flag).leave() check <- () } to deliver() { flag := true check <- () return continueSignal } } } }
}
def makeWorker(piece, checkpoint) {
def stops := timer.now() + 3000 + entropy.nextInt(2000) var count := 0 def checkpointMember := checkpoint <- join() def stopped def run() { # Pretend to do something lengthy; up to 1000 ms. timer.whenPast(timer.now() + entropy.nextInt(1000), fn { if (timer.now() >= stops) { checkpointMember <- leave() bind stopped := true } else { count += 1 println(`Delivering $piece#$count`) when (checkpointMember <- deliver()) -> { println(`Delivered $piece#$count`) run() } } }) } run() return stopped
}
def checkpoint := makeCheckpoint() var waits := [] for piece in 1..5 {
waits with= makeWorker(piece, checkpoint)
} interp.waitAtTop(promiseAllFulfilled(waits))</lang>
Go
As of February 2011, Go has checkpoint synchronization in the standard library, with a type called WaitGroup in the package sync. Code below uses this feature and completes the task with the workshop scenario, including workers joining and leaving.
Also see the Go solution(s) to concurrent computing That is a much simpler task, and shown there are two different implementations of checkpoint synchronization. <lang go>package main
import (
"fmt" "rand" "sync" "time"
)
const nMech = 5 const detailsPerMech = 4
func main() {
assemble := make(chan int) var complete sync.WaitGroup
go solicit(assemble, &complete, nMech*detailsPerMech)
for i := 1; i <= nMech; i++ { complete.Add(detailsPerMech) for j := 0; j < detailsPerMech; j++ { assemble <- 0 } // Go checkpoint feature complete.Wait() // checkpoint reached fmt.Println("mechanism", i, "completed") }
}
func solicit(a chan int, c *sync.WaitGroup, nDetails int) {
rand.Seed(time.Nanoseconds()) var id int // worker id, for output for nDetails > 0 { time.Sleep(5e8 + rand.Int63n(5e8)) // some random time to find a worker id++ // contract to assemble a certain number of details contract := rand.Intn(5) + 1 if contract > nDetails { contract = nDetails } dword := "details" if contract == 1 { dword = "detail" } fmt.Println("worker", id, "contracted to assemble", contract, dword) go worker(a, c, contract, id) nDetails -= contract }
}
func worker(a chan int, c *sync.WaitGroup, contract, id int) {
// some random time it takes for this worker to assemble a detail assemblyTime := 5e8 + rand.Int63n(5e8) fmt.Println("worker", id, "enters shop") for i := 0; i < contract; i++ { <-a fmt.Println("worker", id, "assembling") time.Sleep(assemblyTime) fmt.Println("worker", id, "completed detail") c.Done() } fmt.Println("worker", id, "leaves shop")
}</lang> Output:
worker 1 contracted to assemble 2 details worker 1 enters shop worker 1 assembling worker 2 contracted to assemble 5 details worker 2 enters shop worker 2 assembling worker 1 completed detail worker 1 assembling worker 2 completed detail worker 2 assembling worker 3 contracted to assemble 1 detail worker 3 enters shop worker 1 completed detail worker 1 leaves shop worker 2 completed detail mechanism 1 completed worker 3 assembling worker 2 assembling ... worker 5 completed detail worker 7 completed detail worker 7 leaves shop mechanism 4 completed worker 6 assembling worker 5 assembling worker 6 completed detail worker 6 assembling worker 5 completed detail worker 5 leaves shop worker 6 completed detail worker 6 assembling worker 6 completed detail worker 6 leaves shop mechanism 5 completed
J
The current implementations of J are all single threaded. However, the language definition offers a lot of parallelism which I imagine will eventually be supported, once performance gains significantly better than a factor of 2 on common problems become economically viable.
For example in 1 2 3 + 4 5 6, we have three addition operations which are specified to be carried out in parallel, and this kind of parallelism pervades the language definition.
Java
<lang Java> import java.util.Scanner; import java.util.Random;
public class CheckpointSync{ public static void main(String[] args){ System.out.print("Enter number of workers to use: "); Scanner in = new Scanner(System.in); Worker.nWorkers = in.nextInt(); System.out.print("Enter number of tasks to complete:"); runTasks(in.nextInt()); }
/* * Informs that workers started working on the task and * starts running threads. Prior to proceeding with next * task syncs using static Worker.checkpoint() method. */ private static void runTasks(int nTasks){ for(int i = 0; i < nTasks; i++){ System.out.println("Starting task number " + (i+1) + "."); runThreads(); Worker.checkpoint(); } }
/* * Creates a thread for each worker and runs it. */ private static void runThreads(){ for(int i = 0; i < Worker.nWorkers; i ++){ new Thread(new Worker(i+1)).start(); } }
/* * Worker inner static class. */ public static class Worker implements Runnable{ public Worker(int threadID){ this.threadID = threadID; } public void run(){ work(); }
/* * Notifies that thread started running for 100 to 1000 msec. * Once finished increments static counter 'nFinished' * that counts number of workers finished their work. */ private synchronized void work(){ try { int workTime = rgen.nextInt(900) + 100; System.out.println("Worker " + threadID + " will work for " + workTime + " msec."); Thread.sleep(workTime); //work for 'workTime' nFinished++; //increases work finished counter System.out.println("Worker " + threadID + " is ready"); } catch (InterruptedException e) { System.err.println("Error: thread execution interrupted"); e.printStackTrace(); } }
/* * Used to synchronize Worker threads using 'nFinished' static integer. * Waits (with step of 10 msec) until 'nFinished' equals to 'nWorkers'. * Once they are equal resets 'nFinished' counter. */ public static synchronized void checkpoint(){ while(nFinished != nWorkers){ try { Thread.sleep(10); } catch (InterruptedException e) { System.err.println("Error: thread execution interrupted"); e.printStackTrace(); } } nFinished = 0; }
/* inner class instance variables */ private int threadID;
/* static variables */ private static Random rgen = new Random(); private static int nFinished = 0; public static int nWorkers = 0; } }</lang> Output:
Enter number of workers to use: 5 Enter number of tasks to complete:3 Starting task number 1. Worker 1 will work for 882 msec. Worker 2 will work for 330 msec. Worker 3 will work for 618 msec. Worker 4 will work for 949 msec. Worker 5 will work for 805 msec. Worker 2 is ready Worker 3 is ready Worker 5 is ready Worker 1 is ready Worker 4 is ready Starting task number 2. Worker 1 will work for 942 msec. Worker 2 will work for 247 msec. Worker 3 will work for 545 msec. Worker 4 will work for 850 msec. Worker 5 will work for 888 msec. Worker 2 is ready Worker 3 is ready Worker 4 is ready Worker 5 is ready Worker 1 is ready Starting task number 3. Worker 2 will work for 976 msec. Worker 1 will work for 194 msec. Worker 4 will work for 532 msec. Worker 3 will work for 515 msec. Worker 5 will work for 326 msec. Worker 1 is ready Worker 5 is ready Worker 3 is ready Worker 4 is ready Worker 2 is ready
Perl
The perlipc man page details several approaches to interprocess communication. Here's one of my favourites: socketpair and fork. I've omitted some error-checking for brevity.
<lang perl>#!/usr/bin/perl use warnings; use strict; use v5.10;
use Socket;
my $nr_items = 3;
sub short_sleep($) {
(my $seconds) = @_; select undef, undef, undef, $seconds;
}
- This is run in a worker thread. It repeatedly waits for a character from
- the main thread, and sends a value back to the main thread. A short
- sleep introduces random timing, just to keep us honest.
sub be_worker($$) {
my ($socket, $value) = @_; for (1 .. $nr_items) { sysread $socket, my $dummy, 1; short_sleep rand 0.5; syswrite $socket, $value; ++$value; }
exit;
}
- This function forks a worker and sends it a socket on which to talk to
- the main thread, as well as an initial value to work with. It returns
- (to the main thread) a socket on which to talk to the worker.
sub fork_worker($) {
(my $value) = @_; socketpair my $kidsock, my $dadsock, AF_UNIX, SOCK_STREAM, PF_UNSPEC or die "socketpair: $!";
if (fork // die "fork: $!") { # We're the parent close $dadsock; return $kidsock; } else { # We're the child close $kidsock; be_worker $dadsock, $value; # Never returns }
}
- Fork two workers, send them start signals, retrieve the values they send
- back, and print them
my $alpha_sock = fork_worker 'A'; my $digit_sock = fork_worker 1;
for (1 .. $nr_items) {
syswrite $_, 'x' for $alpha_sock, $digit_sock; sysread $alpha_sock, my $alpha, 1; sysread $digit_sock, my $digit, 1; say $alpha, $digit;
}
- If the main thread were planning to run for a long time after the
- workers had terminate, it would need to reap them to avoid zombies:
wait; wait;</lang>
A sample run:
msl@64Lucid:~/perl$ ./checkpoint A1 B2 C3 msl@64Lucid:~/perl$
PicoLisp
The following solution implements each worker as a coroutine. Therefore, it works only in the 64-bit version.
'checkpoints' takes a number of projects to do, and a number of workers. Each worker is started with a random number of steps to do (between 2 and 5), and is kept in a list of 'Staff' members. Whenever a worker finishes, he is removed from that list, until it is empty and the project is done.
'worker' takes a number of steps to perform. It "works" by printing each step, and returning NIL when done. <lang PicoLisp>(de checkpoints (Projects Workers)
(for P Projects (prinl "Starting project number " P ":") (for (Staff (mapcar '((I) (worker (format I) (rand 2 5))) # Create staff of workers (range 1 Workers) ) Staff # While still busy (filter worker Staff) ) ) # Remove finished workers (prinl "Project number " P " is done.") ) )
(de worker (ID Steps)
(co ID (prinl "Worker " ID " has " Steps " steps to do") (for N Steps (yield ID) (prinl "Worker " ID " step " N) ) NIL ) )</lang>
Output:
: (checkpoints 2 3) # Start two projects with 3 workers Starting project number 1: Worker 1 has 2 steps to do Worker 2 has 3 steps to do Worker 3 has 5 steps to do Worker 1 step 1 Worker 2 step 1 Worker 3 step 1 Worker 1 step 2 Worker 2 step 2 Worker 3 step 2 Worker 2 step 3 Worker 3 step 3 Worker 3 step 4 Worker 3 step 5 Project number 1 is done. Starting project number 2: Worker 1 has 4 steps to do Worker 2 has 3 steps to do Worker 3 has 2 steps to do Worker 1 step 1 Worker 2 step 1 Worker 3 step 1 Worker 1 step 2 Worker 2 step 2 Worker 3 step 2 Worker 1 step 3 Worker 2 step 3 Worker 1 step 4 Project number 2 is done.
PureBasic
PureBasic normally uses Semaphores and Mutex’s to synchronize parallel systems. This system only relies on semaphores between each thread and the controller (CheckPoint-procedure). For exchanging data a Mutex based message stack could easily be added, both synchronized according to this specific task or non-blocking if each worker could be allowed that freedom. <lang PureBasic>#MaxWorktime=8000 ; "Workday" in msec
- Structure that each thread uses
Structure MyIO
ThreadID.i Semaphore_Joining.i Semaphore_Release.i Semaphore_Deliver.i Semaphore_Leaving.i
EndStructure
- Array of used threads
Global Dim Comm.MyIO(0)
- Master loop synchronizing the threads via semaphores
Procedure CheckPoint()
Protected i, j, maxthreads=ArraySize(Comm()) Protected Worker_count, Deliver_count Repeat For i=1 To maxthreads With Comm(i) If TrySemaphore(\Semaphore_Leaving) Worker_count-1 ElseIf TrySemaphore(\Semaphore_Deliver) Deliver_count+1 If Deliver_count=Worker_count PrintN("All Workers reported in, starting next task.") Deliver_count=0 For j=1 To maxthreads SignalSemaphore(Comm(j)\Semaphore_Release) Next j EndIf ElseIf TrySemaphore(\Semaphore_Joining) PrintN("A new Worker joined the force.") Worker_count+1: SignalSemaphore(\Semaphore_Release) ElseIf Worker_count=0 ProcedureReturn EndIf Next i EndWith ForEver StartAll=0
EndProcedure
- A worker thread, all orchestrated by the Checkpoint() routine
Procedure Worker(ID)
Protected EndTime=ElapsedMilliseconds()+#MaxWorktime, n With Comm(ID) SignalSemaphore(\Semaphore_Joining) Repeat Repeat ; Use a non-blocking semaphore check to avoid dead-locking at shutdown. If ElapsedMilliseconds()>EndTime SignalSemaphore(\Semaphore_Leaving) PrintN("Thread #"+Str(ID)+" is done.") ProcedureReturn EndIf Delay(1) Until TrySemaphore(\Semaphore_Release) n=Random(1000) PrintN("Thread #"+Str(ID)+" will work for "+Str(n)+" msec.") Delay(n): PrintN("Thread #"+Str(ID)+" delivering") SignalSemaphore(\Semaphore_Deliver) ForEver EndWith
EndProcedure
- User IO & init
If OpenConsole()
Define i, j Repeat Print("Enter number of workers to use [2-2000]: ") j=Val(Input()) Until j>=2 And j<=2000 ReDim Comm(j) For i=1 To j With Comm(i) \Semaphore_Release =CreateSemaphore() \Semaphore_Joining =CreateSemaphore() \Semaphore_Deliver =CreateSemaphore() \Semaphore_Leaving =CreateSemaphore() \ThreadID = CreateThread(@Worker(),i) EndWith Next PrintN("Work started, "+Str(j)+" workers has been called.") CheckPoint() Print("Press ENTER to exit"): Input()
EndIf</lang>
Enter number of workers to use [2-2000]: 5 Work started, 5 workers has been called. A new Worker joined the force. A new Worker joined the force. A new Worker joined the force. A new Worker joined the force. A new Worker joined the force. Thread #5 will work for 908 msec. Thread #3 will work for 405 msec. Thread #1 will work for 536 msec. Thread #2 will work for 632 msec. Thread #4 will work for 202 msec. Thread #4 delivering Thread #3 delivering Thread #1 delivering Thread #2 delivering Thread #5 delivering All Workers reported in, starting next task. Thread #2 will work for 484 msec. Thread #4 will work for 836 msec. Thread #5 will work for 464 msec. Thread #3 will work for 251 msec. Thread #1 will work for 734 msec. Thread #3 delivering Thread #5 delivering Thread #2 delivering Thread #1 delivering Thread #4 delivering All Workers reported in, starting next task. Thread #3 will work for 864 msec. Thread #1 will work for 526 msec. Thread #5 will work for 145 msec. Thread #2 will work for 762 msec. Thread #4 will work for 283 msec. Thread #5 delivering Thread #4 delivering Thread #1 delivering Thread #2 delivering Thread #3 delivering All Workers reported in, starting next task. Thread #2 will work for 329 msec. Thread #4 will work for 452 msec. Thread #1 will work for 176 msec. Thread #5 will work for 702 msec. Thread #3 will work for 500 msec. Thread #1 delivering Thread #2 delivering Thread #4 delivering Thread #3 delivering Thread #5 delivering All Workers reported in, starting next task. Thread #5 will work for 681 msec. Thread #3 will work for 71 msec. Thread #2 will work for 267 msec. Thread #1 will work for 151 msec. Thread #4 will work for 252 msec. Thread #3 delivering Thread #1 delivering Thread #4 delivering Thread #2 delivering Thread #5 delivering All Workers reported in, starting next task. Thread #5 will work for 963 msec. Thread #3 will work for 378 msec. Thread #1 will work for 209 msec. Thread #4 will work for 897 msec. Thread #2 will work for 736 msec. Thread #1 delivering Thread #3 delivering Thread #2 delivering Thread #5 delivering Thread #4 delivering All Workers reported in, starting next task. Thread #2 will work for 44 msec. Thread #4 will work for 973 msec. Thread #1 will work for 700 msec. Thread #3 will work for 505 msec. Thread #5 will work for 256 msec. Thread #2 delivering Thread #5 delivering Thread #3 delivering Thread #1 delivering Thread #4 delivering All Workers reported in, starting next task. Thread #2 will work for 703 msec. Thread #4 will work for 296 msec. Thread #1 will work for 702 msec. Thread #3 will work for 99 msec. Thread #5 will work for 114 msec. Thread #3 delivering Thread #5 delivering Thread #4 delivering Thread #1 delivering Thread #2 delivering All Workers reported in, starting next task. Thread #3 will work for 97 msec. Thread #5 will work for 192 msec. Thread #2 will work for 762 msec. Thread #1 will work for 232 msec. Thread #4 will work for 484 msec. Thread #3 delivering Thread #5 delivering Thread #1 delivering Thread #4 delivering Thread #2 delivering All Workers reported in, starting next task. Thread #1 will work for 790 msec. Thread #5 will work for 602 msec. Thread #3 will work for 105 msec. Thread #2 will work for 449 msec. Thread #4 will work for 180 msec. Thread #3 delivering Thread #4 delivering Thread #2 delivering Thread #2 is done. Thread #4 is done. Thread #3 is done. Thread #5 delivering Thread #5 is done. Thread #1 delivering Thread #1 is done. Press ENTER to exit
Ruby
<lang ruby>require 'socket'
- A Workshop runs all of its workers, then collects their results. Use
- Workshop#add to add workers and Workshop#work to run them.
- This implementation forks some processes to run the workers in
- parallel. Ruby must provide Kernel#fork and 'socket' library must
- provide UNIXSocket.
- Why processes and not threads? C Ruby still has a Global VM Lock,
- where only one thread can hold the lock. One platform, OpenBSD, still
- has userspace threads, with all threads on one cpu core. Multiple
- processes will not compete for a single Global VM Lock and can run
- on multiple cpu cores.
class Workshop
# Creates a Workshop. def initialize @sockets = {} end
# Adds a worker to this Workshop. Returns a worker id _wid_ for this # worker. The worker is a block that takes some _args_ and returns # some value. Workshop#work will run the block. # # This implementation forks a process for the worker. This process # will use Marshal with UNIXSocket to receive the _args_ and to send # the return value. The _wid_ is a process id. The worker also # inherits _IO_ objects, which might be a problem if the worker holds # open a pipe or socket, and the other end never reads EOF. def add child, parent = UNIXSocket.pair
wid = fork do # I am the child. child.close @sockets.each_value { |sibling| sibling.close }
# Prevent that all the children print their backtraces (to a mess # of mixed lines) when user presses Control-C. Signal.trap("INT") { exit! }
loop do # Wait for a command. begin command, args = Marshal.load(parent) rescue EOFError # Parent probably died. break end
case command when :work # Do work. Send result to parent. result = yield *args Marshal.dump(result, parent) when :remove break else fail "bad command from workshop" end end end
# I am the parent. parent.close @sockets[wid] = child wid end
# Runs all of the workers, and collects the results in a Hash. Passes # the same _args_ to each of the workers. Returns a Hash that pairs # _wid_ => _result_, where _wid_ is the worker id and _result_ is the # return value from the worker. # # This implementation runs the workers in parallel, and waits until # _all_ of the workers finish their results. Workshop provides no way # to start the work without waiting for the work to finish. If a # worker dies (for example, by raising an Exception), then # Workshop#work raises a RuntimeError. def work(*args) message = [:work, args] @sockets.each_pair do |wid, child| Marshal.dump(message, child) end
# Checkpoint! Wait for all workers to finish. result = {} @sockets.each_pair do |wid, child| begin # This waits until the child finishes a result. result[wid] = Marshal.load(child) rescue EOFError fail "Worker #{wid} died" end end result end
# Removes a worker from the Workshop, who has a worker id _wid_. # If there is no such worker, raises ArgumentError. # # This implementation kills and reaps the process for the worker. def remove(wid) unless child = @sockets.delete(wid) raise ArgumentError, "No worker #{wid}" else Marshal.dump([:remove, nil], child) child.close Process.wait(wid) end end
end
- First create a Workshop.
require 'pp' shop = Workshop.new wids = []
- Our workers must not use the same random numbers after the fork.
@fixed_rand = false def fix_rand
unless @fixed_rand; srand; @fixed_rand = true; end
end
- Start with some workers.
6.times do
wids << shop.add do |i| # This worker slowly calculates a Fibonacci number. fix_rand f = proc { |n| if n < 2 then n else f[n - 1] + f[n - 2] end } [i, f[25 + rand(10)]] end
end
6.times do |i|
# Do one cycle of work, and print the result. pp shop.work(i)
# Remove a worker. victim = rand(wids.length) shop.remove wids[victim] wids.slice! victim
# Add another worker. wids << shop.add do |j| # This worker slowly calculates a number from # the sequence 0, 1, 2, 3, 6, 11, 20, 37, 68, 125, ... fix_rand f = proc { |n| if n < 3 then n else f[n - 1] + f[n - 2] + f[n - 3] end } [j, i, f[20 + rand(10)]] end
end
- Remove all workers.
wids.each { |wid| shop.remove wid } pp shop.work(6)</lang>
Example of output:
{23187=>[0, 1346269], 17293=>[0, 1346269], 9974=>[0, 317811], 31730=>[0, 196418], 30156=>[0, 2178309], 25663=>[0, 832040]} ... {23187=>[5, 5702887], 17293=>[5, 832040], 31730=>[5, 514229], 17459=>[5, 2, 24548655], 18683=>[5, 3, 187427], 4494=>[5, 4, 1166220]} {}
Tcl
This implementation works by having a separate thread handle the synchronization (inter-thread message delivery already being serialized). The alternative, using a read-write mutex, is more complex and more likely to run into trouble with multi-core machines. <lang tcl>package require Tcl 8.5 package require Thread
namespace eval checkpoint {
namespace export {[a-z]*} namespace ensemble create variable members {} variable waiting {} variable event # Back-end of join operation proc Join {id} {
variable members variable counter if {$id ni $members} { lappend members $id } return $id
} # Back-end of leave operation proc Leave {id} {
variable members set idx [lsearch -exact $members $id] if {$idx > -1} { set members [lreplace $members $idx $idx] variable event if {![info exists event]} { set event [after idle ::checkpoint::Release] } } return
} # Back-end of deliver operation proc Deliver {id} {
variable waiting lappend waiting $id
variable event if {![info exists event]} { set event [after idle ::checkpoint::Release] } return
} # Releasing is done as an "idle" action to prevent deadlocks proc Release {} {
variable members variable waiting variable event unset event if {[llength $members] != [llength $waiting]} return set w $waiting set waiting {} foreach id $w { thread::send -async $id {incr ::checkpoint::Delivered} }
}
# Make a thread and attach it to the public API of the checkpoint proc makeThread Template:Script "" {
set id [thread::create thread::wait] thread::send $id { namespace eval checkpoint { namespace export {[a-z]*} namespace ensemble create
# Call to actually join the checkpoint group proc join {} { variable checkpoint thread::send $checkpoint [list \ ::checkpoint::Join [thread::id]] } # Call to actually leave the checkpoint group proc leave {} { variable checkpoint thread::send $checkpoint [list \ ::checkpoint::Leave [thread::id]] } # Call to wait for checkpoint synchronization proc deliver {} { variable checkpoint # Do this from within the [vwait] to ensure that we're already waiting after 0 [list thread::send $checkpoint [list \ ::checkpoint::Deliver [thread::id]]] vwait ::checkpoint::Delivered } } } thread::send $id [list set ::checkpoint::checkpoint [thread::id]] thread::send $id $script return $id
}
# Utility to help determine whether the checkpoint is in use proc anyJoined {} {
variable members expr {[llength $members] > 0}
}
}</lang> Demonstration of how this works.
<lang tcl># Build the workers foreach worker {A B C D} {
dict set ids $worker [checkpoint makeThread {
proc task {name} { checkpoint join set deadline [expr {[clock seconds] + 2}] while {[clock seconds] <= $deadline} { puts "$name is working" after [expr {int(500 * rand())}] puts "$name is ready" checkpoint deliver } checkpoint leave thread::release; # Ask the thread to finish }
}]
}
- Set them all processing in the background
dict for {name id} $ids {
thread::send -async $id "task $name"
}
- Wait until all tasks are done (i.e., they have unregistered)
while 1 {
after 100 set s 1; vwait s; # Process events for 100ms if {![checkpoint anyJoined]} {
break
}
}</lang> Output:
A is working C is working B is working D is working B is ready A is ready D is ready C is ready B is working A is working D is working C is working D is ready A is ready C is ready B is ready B is working D is working A is working C is working D is ready C is ready B is ready A is ready D is working C is working B is working A is working D is ready A is ready C is ready B is ready D is working C is working A is working B is working C is ready A is ready B is ready D is ready