[gmx-users] Gromacs on SGI Altix

[Mark Abraham]

Mingfeng Yang wrote:
> Erik Lindahl wrote:
> 
>> Hi,
>>
>>>>
>>> Thank you, Florian! I will try d.dppc. I have another question. Is 
>>> Gromacs specifically optimized for x86 machines, and the power of 
>>> Itanium2 is not fully used. Because  for Amber8, my Itanium2 is 
>>> almost two times faster than my opteron, but for Gromacs, opteron is 
>>> slightly (~1.2 times) faster than Itanium2. Other factors are almost 
>>> same. Just due to the reason that there is no assembly loops 
>>> optimization for IA64?
>>
>>
>> There are certainly assembly loops for IA64 (trust me, I handcoded 
>> them :-)

>> Erik

Yup, there was a big difference on my SGI Altix 3700Bx2 between 3.2.1 
and a 3.3 beta that had these routines in.

> Thank you, Erik! Anyway, I am impressed by the efficiency of Gromacs, 
> especially its performance on our opteron cluster. The scaling for the 
> SGI altix has come out; this time, with 2CPUs, it's 1.8 times faster 
> than single CPU. I don't care too much about altix, coz most of my job 
> will be run on opteron.

That doesn't sound good enough to me. Using a gromacs 3.3 beta, I 
reported scaling here back in August (See 
http://www.gromacs.org/pipermail/gmx-users/2005-July/016104.html) which 
didn't deviate from linear until somewhere between 8 and 16 processors 
(except for 2 processors which was super-linear!). Your configure line 
looked fine - my installation used an earlier version of icc and 
sgi-mpt, as you have.

If your 2 CPU benchmark is indicative of scaling to higher processor 
numbers, I'd guess your interconnects are not as good as mine, which is 
"SGI's NUMAlink4  interconnect both within and between partitions 
providing 3.2 Gbytes/s bidirectional bandwidth per link and < 2us MPI 
latency." However the Amber8 scaling suggests your connectivity should 
be good enough...

Mark