[Starpu-devel] StarPU overhead for MPI

[Philippe SWARTVAGHER <philippe.swartvagher@inria.fr>]

Hello,

Attached to this mail, a plot resulting of the execution of 
mpi/tests/sendrecv_bench. This plot show latency of data transfers, 
according to data size. The reference curve is raw NewMadeleine 
nmad/examples/benchmarks/nm_bench_sendrecv. These two programs are the 
same, except that in StarPU bench, measured period is between just 
before starpu_mpi_send and just after starpu_mpi_recv. So this plot 
should show the StarPU-MPI overhead.

The overhead seems a little bit huge (a small comm lasts 10 times 
longer). Do you know what could be the origin of this overhead ?

StarPU was compiled without FXT and without --enable-debug.

I tried adding events in trace, to see which function could consume a 
lot of time. I tried this on dalton/joe01. A transfer of 0 byte lasts 
109µs in median and average. Nothing really relevant appeared, but 
between just before the call to starpu_mpi_recv and just after the call 
to nm_sr_recv_irecv, 14µs elapsed.


Any idea ?

--
Philippe SWARTVAGHER

Doctorant
Équipe TADaaM, Inria Bordeaux Sud-Ouest

Attachment: microbench_sendrecv_plafrim.png
Description: PNG image