[Beowulf] SC13: the Limulus boxen appear ...

[Douglas Eadline]

> On 12/09/2013 12:11 PM, Douglas Eadline wrote:
>> Responding to my own post, I recently tested the i7 haswell variant
>> of our Limulus and was 1.7 GFLOPS short of 0.5 TFLOPS (498.3 GFLOPS)
>> running HPL. I suppose with a little more fussing I could find
>> 2 extra GFLOPS, but I'm impressed with what AVX2 with FMA can do,
>> at least for HPL, I have not checked anything else yet.
>>
>> Latest results are here:
>> http://limulus.basement-supercomputing.com/wiki/LimulusBenchmarks
>>
>> Has anyone else seen any similar big jumps in performance?
>> (of course compiler support is needed)
>>
>>
>
> Any source code mods for this?  Did you use Intel or GCC compilers? Nice
> results!

No code mods, used Intel 13.1.3, MKL, Open MPI. Also running
kernel 3.12.2 for the Haswell support. It seemed with the newer
kernels I could use almost 92% of memory, older kernels (2.6.* kernels
seemed to limit at about 88%) Also, the best number, after properly sized,
came from rebooting everything and running with virgin page tables.

--
Doug

>
> --
> Joseph Landman, Ph.D
> Founder and CEO
> Scalable Informatics, Inc.
> email: landman at scalableinformatics.com
> web  : http://scalableinformatics.com
> twtr : @scalableinfo
> phone: +1 734 786 8423 x121
> cell : +1 734 612 4615
>
>
> --
> Mailscanner: Clean
>


-- 
Doug

-- 
Mailscanner: Clean