Re: [atlas-devel] OpenMP settings
Brought to you by:
rwhaley,
tonyc040457
From: R. C. W. <rcw...@ls...> - 2014-08-06 22:14:54
|
OK, I sped things up a bit by turning off PCA panel factorization in both LU and QR. I also installed 3.10.2, and it is even slower than 3.11.29 on the same systems. So, I guess my new stuff is working as well as it ever did with OpenMP, which is not well at all. I know some folks use OpenMP, and the question now is if they ever got good results that way. Please let me know if you use ATLAS wt OpenMP. Thanks, Clint On 08/06/2014 04:46 PM, R. Clint Whaley wrote: > Guys, > > I know some of you tend to compile ATLAS with OpenMP because you have > applications that use OpenMP. When you have done that in the past, what > environment variables do you typically set? > > The reason I ask is that I'm trying to re-enable OpenMP as an option in > the developer series, and I now have something that builds, but the > performance is often worse than serial. Here's a run comparing > cholesky, LU, and QR using pthreads on a 32-core opteron using my > pthreads blas: > > -- ORD UPLO N NRHS lda ldb TIME MFLOP > RESID > === ==== ====== ====== ====== ====== ========= ======== =========== > > C U 2000 1 2000 2000 0.074 36101.60 > 2.344654e-04 > C G 2000 1 2000 2000 0.207 25803.12 > 2.080718e-02 > C Q 2000 1 2000 2000 0.397 26863.33 > 6.125385e-03ORD > > > Now here's with my OpenMP blas: > UPLO N NRHS lda ldb TIME > > > MFLOP RESID > === ==== ====== ====== ====== ====== ========= ======== =========== > > C U 2000 1 2000 2000 0.108 24837.22 > 2.344654e-04 > C G 2000 1 2000 2000 16.977 314.51 > 4.030246e-02 > C Q 2000 1 2000 2000 41.308 258.32 > 5.260133e-03 > > > I've tried setting OMP_PROC_BIND to both true and false, and also > setting GOMP_CPU_AFFINITY to "0-31", and I still get these results. > > It's probably a bug in my OpenMP implementation; I'll install 3.10 on > the same system with OpenMP, and see what I get with that, but if you > have some environment variables I should set, please let me know. > > If it matters, the two machines I'm using are a 32-core AMD Opteron 6128 > and a 12-core Intel Xeon E5645. > > Thanks, > Clint > > ********************************************************************** > ** R. Clint Whaley, PhD * Assoc Prof, LSU * www.csc.lsu.edu/~whaley ** > ********************************************************************** > > ------------------------------------------------------------------------------ > Infragistics Professional > Build stunning WinForms apps today! > Reboot your WinForms applications with our WinForms controls. > Build a bridge from your legacy apps to the future. > http://pubads.g.doubleclick.net/gampad/clk?id=153845071&iu=/4140/ostg.clktrk > _______________________________________________ > Math-atlas-devel mailing list > Mat...@li... > https://lists.sourceforge.net/lists/listinfo/math-atlas-devel > -- ********************************************************************** ** R. Clint Whaley, PhD * Assoc Prof, LSU * www.csc.lsu.edu/~whaley ** ********************************************************************** |