Hi Sara,
One high-level point is that the performance advantage of asyncCopy
really only shows when running Native X10 on a x10rt implementation that
can exploit real network/hardware RDMA support. (PAMI on BG/Q and
Power775. MPI on some platforms with IB networks or better).
Without the underlying RDMA acceleration, asyncCopy gets implemented
in software on top of the same communication mechanism as at (but less
directly, because it is going through an emulation layer).
I suspect on x10rt_sockets, especially on Managed X10, copying a Rail
via an at will beat copying via asyncCopy by quite a bit. There is no
hardware acceleration and the software path length is shorter.
Ben, if I'm wrong please correct me.
--dave
Sara Salem Hamouda <sara.sa...@anu.edu.au> wrote on 11/05/2014 04:26:00 AM:
>
> I was trying to compare the performance of copying a Rail[Byte]
> between 2 places using Rail.async and using "at". Each place is
> running on a different host, and the network speed is 1000Mb/s.
>
> The first method:
> 1. A rail is created
> 2. A GlobalRef to that rail is created
> 3. Used At to move to the second place -> here I expect that "at"
> will need to copy only the GlobalRef, not the whole array.
> 4. Used Rail.asyncCopy from the second place to copy the array using
> the GlobalRef
>
> The second method:
> 1. A rail is created
> 2. Used At to move to the second place -> here I expect that "at"
> will implicitly copy the whole rail, that is why I expected it to be
> slower than the "at" in the first method.
> 3. Used Rail.copy from the second place to copy the array
>
> That is the used program code:
> import x10.util.Timer;
> public class TestRemoteCopy {
> public static def main(args:Rail[String]) {
> //Create Rail and Initialize it
> val size = Long.parse(args(0));
> val source:Rail[Byte] = new Rail[Byte](size);
> val sourcePlace = Place.places()(0);
> val destPlace = Place.places()(1);
> for (var i:Long = 0; i < size; i++){
> source(i) = (i%128) as Byte;
> }
>
> //Copy using Rail.asyncCopy
> val gr = GlobalRail[Byte](source);
> val timeBeforeAt = Timer.milliTime();
> at (destPlace){
> Console.OUT.println("At Time["+(Timer.milliTime()-
> timeBeforeAt)+"] ...");
> val dest:Rail[Byte] = new Rail[Byte](size);//1<
> val timeBeforeAsyncCopy = Timer.milliTime();
> finish Rail.asyncCopy(gr, 0, dest, 0, size);
> Console.OUT.println("AsyncCopy Time["+
> (Timer.milliTime()-timeBeforeAsyncCopy)+"] ...");
> }
>
> //Copy using at
> val timeBeforeSecondAt = Timer.milliTime();
> at(destPlace){
> Console.OUT.println("Second At Time["+(Timer.milliTime()-
> timeBeforeSecondAt)+"] ...");
> val dest:Rail[Byte] = new Rail[Byte](size);
> val timeBeforeSecondCopy = Timer.milliTime();
> Rail.copy(source, 0, dest, 0, size);
> Console.OUT.println("Second Copy Time["+(Timer.milliTime()-
> timeBeforeSecondCopy)+"] ...");
> }
> }
> }
>
> These are some results, time is printed in milliseconds:
> Rail Size: 1024 bytes
> Native X10:
> At Time[4305] ...
> AsyncCopy Time[1] ...
> Second At Time[4305] ...
> Second Copy Time[0] ...
>
> Managed X10:
> At Time[4521] ...
> AsyncCopy Time[31] ...
> Second At Time[4308] ...
> Second Copy Time[1] ...
>
> Rail Size: 1048576 bytes (1024*1024)
> Native X10:
> At Time[4306] ...
> AsyncCopy Time[10] ...
> --> using "at" hangs
>
> Managed X10:
> At Time[4534] ...
> AsyncCopy Time[106] ...
> Second At Time[4331] ...
> Second Copy Time[0] ...
>
> I have some questions based on these numbers:
> 1) Why "at" in the first method, which is supposed to copy only a
> GlobalRef, is taking as much time or even more time than the second
> method which copies the whole rail?
> 2) Is there a maximum limit on the data size that "at" can copy in native
X10?
> 3) The above numbers suggests that using "at" for copying rails is
> more preferable than using Rail.asyncCopy. Is this conclusion correct?
>
> ​Best Regards,
> Sara
>
------------------------------------------------------------------------------
> _______________________________________________
> X10-users mailing list
> X10-users@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/x10-users------------------------------------------------------------------------------
_______________________________________________
X10-users mailing list
X10-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/x10-users
