You should look at the results for HPCC on the HPCC web site. HPL is
dense linear algebra, which maps very well to fast processors with
reasonably cache sizes. If you have lots of memory, you can hide the
cost of even slow networks like GigE. FFT, on the other hand, is not
well suited to TCP networks. Or any network, really, but that's
another story. You're never going to see the same FLOP/S with FFT as
HPL, which is why FFT is in the HPCC benchmark suite.
As for what PTRANS signifies, have a look at the HPCC home page:
http://icl.cs.utk.edu/hpcc/
Brian
On Dec 12, 2005, at 4:25 PM, Srinivasa Prade Patri wrote:
> HI!
> I ran HPCC benchmark for problem size 60000x60000 on 20 nodes
> with 4x5 processor grid for block sizes 64,128,256,512. I have some
> trouble evaluating the results. The performance FFTE is very low,
> is the FLOPS evaluated for 20 processors or only single processor.
> IN either case why is the performance very low comapred to HPL.
> What does PTRANS results signify and how different are they from
> communication bandwidth results.
>
> HPL: max FLOPS = 62GFLOPS
>
> FFT: Minimum Gflop/s 0.399492
> Average Gflop/s 0.421822
> Maximum Gflop/s 0.431742
>
> PTRANS results:
> TIME M N MB NB P Q TIME CHECK GB/s RESID
> ---- ----- ----- --- --- --- --- -------- ------ -------- -----
> WALL 30000 30000 64 64 4 5 12.69 PASSED 0.567 0.00
> CPU 30000 30000 64 64 4 5 4.04 PASSED 1.782 0.00
> WALL 30000 30000 128 128 4 5 12.46 PASSED 0.578 0.00
> CPU 30000 30000 128 128 4 5 3.72 PASSED 1.935 0.00
> WALL 30000 30000 256 256 4 5 12.36 PASSED 0.583 0.00
> CPU 30000 30000 256 256 4 5 3.75 PASSED 1.920 0.00
> WALL 30000 30000 512 512 4 5 13.80 PASSED 0.522 0.00
> CPU 30000 30000 512 512 4 5 5.07 PASSED 1.420 0.00
>
> communication bandwidth latency results
>
> Max Ping Pong Latency: 0.047505 msecs
> Randomly Ordered Ring Latency: 0.048903 msecs
> Min Ping Pong Bandwidth: 72.532088 MB/s
> Naturally Ordered Ring Bandwidth: 34.848659 MB/s
> Randomly Ordered Ring Bandwidth: 34.597277 MB/s
>
>
> Thanks for ur help
>
> Regards
> Srinivasa patri
>
>
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
|
