Hello Vijay,
Thank you for your helpful answer.
I profiled the X10 C++ back-end as you suggested, and found that the GC
is heavy.
However, I also notice that for both parallel and sequential Fib, as
cores used to compute increases,
the time taken by following functions increase.
__lll_lock_wait
_L_lock_102
_L_unlock_56
__lll_unlock_wake
What are they? Are they related to the user-space scheduler?
The result of performance evaluation shows that, computing with 8-core
has best performance.
With more than 8 cores, it seems that the benefit of multi-cores is
trade off by the overhead of lock/unlock.
I attached the Fib program I used to test and one profile result. Please
check them out.
I hope the fib program can run with linear scalability. Is it possible?
Thank you very much.
Regards,
-Tetsu
On 07/24/2010 01:56 AM, Vijay Saraswat wrote:
Also, it would be good to see your Fib code, Tetsu.
Again: the main principle is, avoid creating throwaway objects if u
want good perf in C++. (Java is much better at dealing with this.)
Olivier Tardieu wrote:
Hi,
With Fib, memory allocation & garbage collection dominate the execution
time.
The runtime creates objects internally for each async in your code.
You probably also create a Fib object for each computation step.
The performance of the JVM GC in this context is much better than the
one
of the C++ backend.
You can verify this by profiling the generated binary.
I routinely do this using google performance tools.
Bottom line, the very fine granularity of asyncs in Fib cannot give you
good scalability with the current runtime.
You need to have *enough* sequential computation in each async, e.g.,
compute Fib(n) sequentially for n<=MIN for some constant MIN.
Regards,
Olivier
------------------------------------------------------------------------------
This SF.net email is sponsored by Sprint
What will you do first with EVO, the first 4G phone?
Visit sprint.com/first -- http://p.sf.net/sfu/sprint-com-first
_______________________________________________
X10-users mailing list
X10-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/x10-users
import x10.io.Console;
public class Fib {
var r:int;
var f:Boolean;
public def this(x:int, flag:Boolean) {
r = x;
f = flag;
}
public def run() {
if (r < 2) return;
val f1 = new Fib(r-1, f);
val f2 = new Fib(r-2, f);
if ( r > 12) {
finish {
async f1.run();
f2.run();
}
} else {
f1.run();
f2.run();
}
r = f1.r + f2.r;
}
public static def main(args:Rail[String]!) {
Console.OUT.println("# places = " + Place.MAX_PLACES);
val n = (args.length > 0) ? int.parse(args(0)) : 10;
val flag = (args.length > 1) ? Boolean.parse(args(1)) : true;
val f = new Fib(n, flag);
f.run();
Console.OUT.println("Fib("+n+") = "+f.r);
}
}
------------------------------------------------------------------------------
This SF.net email is sponsored by Sprint
What will you do first with EVO, the first 4G phone?
Visit sprint.com/first -- http://p.sf.net/sfu/sprint-com-first
_______________________________________________
X10-users mailing list
X10-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/x10-users
