[parsec-users] Runtime problem with canneal and intel compiler

Christian Bienia cbienia at CS.Princeton.EDU
Mon Apr 21 12:58:43 EDT 2008


Hello Michael,

That is strange, I don't know of anything that might be the reason. In our 
experiments with icc we never had any problems. The hotspot of canneal is a 
small inner loop, so it shouldn't be too hard to find out what's going on. No 
external libraries are used there.

Could you check the following:
1. Use profiling to confirm that the hotspot is still the same and it's not 
hanging somewhere else
2. Generate annotated assembly output for both gcc and icc. The inner loop 
should have few instructions, so it shouldn't be too difficult to compare the 
two. Maybe there are noteworthy differences.

It seems this is more an icc problem than a canneal problem. Try contacting 
the compiler team at Intel, they might be interested in this behavior.

- Chris



On Monday 21 April 2008 04:36 am, Schilli, Michael wrote:
> Hi all,
>
>
>
> I have strange runtime behaviour of canneal in combination with the Intel
> compiler. When compiling canneal with gcc and running the native testset,
> it takes about 8 min on my machine to complete (single threaded). When
> using the Intel compiler and native testset it takes more than 2 hours!
> Both tests have been compiled with the included configuration files and for
> 64-Bit. Running both with simlarge and smaller testsets results in
> comparable runtimes, so this problem only arises with the native testset. I
> have already lowered optimization level in but nothing changed.
>
> Perhaps someone has any idea on this.
>
>
>
> Thanks
>
> Michael



More information about the parsec-users mailing list