Re: [Starpu-devel] Performance decreasing by adding empty tasks

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Xavier Lacoste, le Wed 08 Feb 2012 15:26:40 +0100, a écrit :
> I'm using 0.9.2 version of StarPU because I noticed regression with
> trunk version :
> shipsec_r5550_heft_8CPU_0CUDA:----- sopalin time 33.981822
> shipsec_r5550_heft_8CPU_0CUDA:----- sopalin time 33.579352

Did Nicolas tell you about trying to disable transfer overlap prediction
(see attached patch)?  That's the main thing I can think of concerning a
regression.

Otherwise, a bisection over svn would help to know exactly what brought
the regression.

Samuel
Index: src/sched_policies/heft.c
===================================================================
--- src/sched_policies/heft.c   (révision 5624)
+++ src/sched_policies/heft.c   (copie de travail)
@@ -157,6 +157,7 @@
        /* If there is no prediction available, we consider the task has a 
null length */
        if (!isnan(predicted_transfer))
        {
+#if 0
                if (starpu_timing_now() + predicted_transfer < 
exp_end[workerid])
                {
                        /* We may hope that the transfer will be finished by
@@ -169,6 +170,7 @@
                         * remainder into account */
                        predicted_transfer = (starpu_timing_now() + 
predicted_transfer) - exp_end[workerid];
                }
+#endif
                task->predicted_transfer = predicted_transfer;
                exp_end[workerid] += predicted_transfer;
                exp_len[workerid] += predicted_transfer;
@@ -193,6 +195,7 @@
        exp_end[best_workerid] += predicted;
        exp_len[best_workerid] += predicted;
 
+#if 0
        if (starpu_timing_now() + predicted_transfer < exp_end[best_workerid])
        {
                /* We may hope that the transfer will be finished by
@@ -205,6 +208,7 @@
                 * remainder into account */
                predicted_transfer = (starpu_timing_now() + 
predicted_transfer) - exp_end[best_workerid];
        }
+#endif
        exp_end[best_workerid] += predicted_transfer;
        exp_len[best_workerid] += predicted_transfer;