I am glad it helps.
As I mentioned before, I used intel MPI 2.0 to parallel the calculation. Intel MPI 2.0 is an implementation of MPI 2, which recognizes SMP. So actually there is no communication at all. The data exchange in memory. As I found, the CPU occupation is 100%. I also tried MPICH-1.2.7 to parallel it, which can only give CPU occupation about 97% in this dual dual-core box.
But if you want to parallel between boxes, you should think about high speed connnection with low latency. Giganet is poor.
Price is another issue you should concern.
BTW, this XEON I used is of bensley core, and Woodcrest will be available in market soon, I guess. Woodcrest has a amazing performance over bensley.
======= 2006-05-12 20:46:08 Robert Duke wrote=======
>Ye -
>Thanks for posting your results. I presume this is a pme run of pmemd 9?
>What's the size of the benchmark? One note - If you do runs of different
>lengths, there will be slight variation in runtimes as the system settles.
>So it becomes hard to compare absolutely down to the level of 3%.
>Generally, also, I can find minor things to do that will have a 1-3% impact.
>So with slight differences you may be able to throw the numbers either way.
>An exciting thing to me, though, is the dual core/dual cpu config, because
>this basically gives you a relatively inexpensive 4 cpu node that is not
>doing mpi i/o over the interconnect. I would expect the scaling to be
>pretty good. The only experience I have with this is on a cray xd1, and the
>dual cpu/dual core nodes are indeed nice; I can't see the differential in
>scaling though because the xd1 has a good inter-node interconnect. Still, I
>think there is a future down this path. One other issue I have not been
>following, and that is how each of these architectures is "topping out".
>The xeon is cranking out around 3.6 GHz these days; I am mostly seeing 2.2
>GHz opterons. I don't know where intel takes the xeon next, but it would
>seem that highspeed clock heating is an issue. On the other hand, the
>opteron already has heat issues down around 2 GHz, so I don't know how much
>further it is going. I think we are in the fortunate position these days of
>having two good and relatively inexpensive alternatives, though.
>Best Regards - Bob Duke
>
>----- Original Message -----
>From: "Ye Mei" <ymei.itcc.nju.edu.cn>
>To: <amber.scripps.edu>
>Sent: Friday, May 12, 2006 4:04 AM
>Subject: Re: AMBER: Opteron vs. Xeon performance
>
>
>>I ever did some tests for IBM new Xeon Blade, and compared with my
>>"homemade" Opteron workstation.
>>
>> Molecular Dynamics
>> Software: Amber 8 with all patches
>> Compiler and LIB: Inter Fortran 9.0 with IMKL8.1
>> Parallel Environment: Intel MPI 2.0
>>
>> Intel Xeon Dual Dual-Core 3.2GHz
>> Simulation time: 100ps CPU time: 467min speed: 0.214ps/min
>>
>> AMD opteron 275 (2.2GHz) Dual Dual-Core
>> Simulation time: 5000ps CPU time: 22571min speed: 0.222ps/min
>>
>> Speed: Intel:AMD=1:1.03
>>
>> AMD is a little bit faster.
>>
>> Best regards,
>>
>> Ye Mei
>> ymei.itcc.nju.edu.cn
>> Institute of Theoretical and Computational Chemistry
>> Key Laboratory of Mesoscopic Chemistry
>> School of Chemistry and Chemical Engineering
>> Nanjing University
>> Nanjing 210093
>> P.R.China
>> 2006-05-12
>>
>>
>> ======= 2006-05-12 02:47:42 Jim Paugh wrote=======
>>
>>>I am looking for recent data comparing the performance between Opterons
>>>and Xeons using Amber, as well as interpretation of that data.
>>>
>>>Thanks,
>>>Jim
>>>-----------------------------------------------------------------------
>>>The AMBER Mail Reflector
>>>To post, send mail to amber.scripps.edu
>>>To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>>>
>>
>> = = = = = = = = = = = = = = = = = = = =
>>
>>
>>
>>
>>
>>
>> -----------------------------------------------------------------------
>> The AMBER Mail Reflector
>> To post, send mail to amber.scripps.edu
>> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>>
>
>
>-----------------------------------------------------------------------
>The AMBER Mail Reflector
>To post, send mail to amber.scripps.edu
>To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>
= = = = = = = = = = = = = = = = = = = =
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Sun May 14 2006 - 06:07:14 PDT
