Re: Two GPU-based workstation

From: James Starlight (jmsstarlight_at_gmail.com)
Date: Tue Jan 14 2014 - 04:45:38 CST

Dear NAMD users!

This time I'd like to upgrade my GPU-based workstation equiped with 2
TItans. I've mentioned previously that there were no increasing in
performance using both titans simultaneosusly with the single i7 processor
and 32 gram). Consequently I'd like to change MB and processor on this
machine thinking about server-like MB with two XENONS cpu's. Could some one
provide me with the example of such configuration (with exact models of
MB's and CPUs) which could give me best performance on the modelling on TWO
GeForce Titans (I'm modelling water soluble and membrane proteins in
explicit solvent ~ overall systems is 100k atoms). Also additional advises
are welcome :)

James

2013/11/11 James Starlight <jmsstarlight_at_gmail.com>

> Norman,
>
>
> 1) I dont know why but I have this error with the PBC vectors in case
> when I change fullElectFrequency value only (not any changing in barostat
> options etc)
>
> 2) some benchmarks for the system of membrane receptor (60k atoms)
>
> using use +p6 +ppn3 +devices 0,1
>
> Info: Benchmark time: 3 CPUs 0.0472163 s/step 0.273242 days/ns 290.363 MB
> memory
>
> using +ppn6 +devices 0,1
>
> Info: Benchmark time: 6 CPUs 0.0288685 s/step 0.167063 days/ns 325.297 MB
> memory
>
> However n the second case I have still bad performance (~1ns/4 hour). For
> comparison in Gromacs with the same setups I have twice better of the
> performance (although it can be possible also to obtain gain for the usage
> of the both GPUs simultaneously)
>
>
> Does anybody else have some experience with the GeForce titans ?
>
> James
>
>
> 2013/11/11 Norman Geist <norman.geist_at_uni-greifswald.de>
>
>> Read something about „Intel HT-Hyperthreading“.
>>
>>
>>
>> Yes.
>>
>>
>>
>> Norman Geist.
>>
>>
>>
>> *Von:* owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu] *Im
>> Auftrag von *James Starlight
>>
>> *Gesendet:* Montag, 11. November 2013 09:49
>> *An:* Norman Geist; Namd Mailing List
>>
>> *Betreff:* Re: namd-l: Two GPU-based workstation
>>
>>
>>
>> 2) about cores
>>
>> so If my physical number is 6 ( I dont really know why debian recognize
>> 12 cores)
>>
>> I should use +p6 +ppn3 +devices 0,1 do adjust each 3 cores for each gpu
>> shouldn't it ?
>>
>>
>>
>> 2013/11/11 James Starlight <jmsstarlight_at_gmail.com>
>>
>> Norman,
>>
>> using
>>
>> pmegridspacing 1;
>> fullElectFrequency 2;
>>
>> instead of
>>
>> PMEGridSizeX $fftx; # should be close to the cell size
>> PMEGridSizeY $ffty; # corresponds to the charmm input
>> fftx/y/z
>> PMEGridSizeZ $fftz;
>> pmegridspacing 1;
>>
>> have crashed my simulation at the beginning
>>
>> namd2 +idlepoll +p12 +ppn6 +devices 0,1 ./aMD.conf >>
>> b2ar_p0gDiheBoostlog_20000
>> ------------- Processor 0 Exiting: Called CmiAbort ------------
>> Reason: FATAL ERROR: Periodic cell has become too small for original
>> patch grid!
>> Possible solutions are to restart from a recent checkpoint,
>> increase margin, or disable useFlexibleCell for liquid simulation.
>>
>>
>> (this time I've start from the checkpoint of the previous run, performed
>> with the defined XYZ of the PME. Does it means that I should to begin
>> simulation from the beginning of the equilibration phase with
>> fullElectFrequency >1 or there are alternative solutions?
>>
>> James
>>
>>
>>
>>
>>
>> 2013/11/11 Norman Geist <norman.geist_at_uni-greifswald.de>
>>
>> Why not simply setting „pmegridspacing 1“?
>>
>>
>>
>> I usually use numcpus/numgpus for each gpu. And really believe me, you
>> only got 6 cores not 12. And running 2 simulations on only one cpu socket
>> is inefficient.
>>
>>
>>
>> Norman Geist.
>>
>>
>>
>> *Von:* owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu] *Im
>> Auftrag von *James Starlight
>>
>>
>> *Gesendet:* Samstag, 9. November 2013 15:53
>>
>> *An:* Namd Mailing List
>>
>>
>> *Betreff:* Re: namd-l: Two GPU-based workstation
>>
>>
>>
>> by the way increasing of the fullElectFrequency > 1 has end simulation
>> with the errors about unproperly set XYZ of the PME boundaries ( with
>> fullElectFrequency 1 I use 80 80 120 simulating membrane protein and have
>> not any errors. How could I change pme options ?
>>
>> also my question is the optimal balancing of the number of CPUs for each
>> GPU. Is there some impirical relationships showing what amount of CPUs is
>> needed for each GPU ?
>>
>> assuming that I obtained best performance using
>> namd2 +idlepoll +p12 +devices 0 ./aMD.conf
>>
>> I'd like to share some CPUs between both available GPUs for the 2
>> parallel simulations.
>>
>> James
>>
>>
>>
>> 2013/11/8 James Starlight <jmsstarlight_at_gmail.com>
>>
>> Could fullelectfrequency 4 increase performance exactly dual-gpu regime ?
>>
>> In case of running two simulations will it be enough to provide each gpu
>> with the 6 cores ? ( I suppose that I have not obtain good performance in 2
>> gpu regime exactly due to small number of cores for each gpu)
>>
>> James
>>
>>
>>
>> 2013/11/7 Ajasja Ljubetič <ajasja.ljubetic_at_gmail.com>
>>
>> On 7 November 2013 06:32, James Starlight <jmsstarlight_at_gmail.com> wrote:
>>
>> I've gone to conclusion that using 2 GPUs simultaneously gave me the
>> same performance as 1 GPU like
>>
>>
>>
>> Yes, this is expected, for such small systems there is too little work
>> at each step to scale efficiently. You can however run one (or two or
>> three) independent simulations on each GPU.
>>
>>
>>
>> Regards,
>>
>> Ajasja
>>
>>
>>
>>
>>
>>
>>
>>
>>
>
>

This archive was generated by hypermail 2.1.6 : Wed Dec 31 2014 - 23:22:02 CST