Re: [Starpu-devel] bad performance scaling behavior of cholesky decomposition

[Mario Tacconi <mtacco@caspur.it>]

> > During the benchmarks I found that the dmdas policy actually gives the
> > best performance.
> Good, that's what is expected :) though heft is supposed to achieve the
> same. Is the difference very big?
>
> Samuel

well, actually dmdas and heft seem to battle head-to-head for the
performance crown:

size     dmdas heft      dmdar  dmda (GFLOPS)
4096   116.55 131.61 143.64 128.89
8192   328.19 320.36 302.08 278.37
24000 638.11 649.57 608.03 523.48
32000 744.76 703.07 679.96 606.96
36000 726.29 741.33 709.56 644.47
42000 749.97 750.38 707.70 655.39
48000 769.58 771.42 753.85 711.83
52000 776.10 761.46 763.27 724.30
58000 745.42 722.25 757.51 715.01
64000 710.51 728.24 755.58 746.31

above I'm using 2 CPU workers + 2 CUDA workers.
Below I'm reporting the data for dmdas policy using 10 cpu worker + 2 
cuda worker
(basically, a whole compute node of our cluster: 2 Xeon 
X5650@2.67GHz+2Tesla C2050):

4096 205.16
8192 359.04
16000 563.40
24000 672.37
32000 799.73
36000 790.88
42000 798.96
48000 831.15
52000 811.96
58000 817.33
64000 863.52

Are these results reasonable? Do you think that StarPU is correctly 
behaving and performing on our system?

Mario