AMBER Archive (2005)

Subject: Re: AMBER: amber8 parallel sander

From: Robert Duke (rduke_at_email.unc.edu)
Date: Tue Oct 04 2005 - 00:26:28 CDT


Okay, I presume everyone realizes this, but I have recently been amazed by a
few things, so:
1) PMEMD does the pme-type explicit solvent simulation subset of sander
simulations.
2) It scales much better than sander, up to 128+ cpu's, depending on
interconnect (gigabit ethernet does not work particularly well for anything;
anything serious - myrinet, infiniband, proprietary interconnects like ibm
hss, etc. and fast smp's do work well).
3) We can get up to 3+ nsec/day on 90K atoms, explicit solvent for pmemd 8,
depending on system; we have recently broken the 10 nsec/day barrier for
this system using a pmemd 9 prerelease version on large iron.
4) "We" here is me working with Lee Pedersen, Tom Darden, et al. at
UNC-Chapel Hill and NIEHS.
Regards - Bob Duke

----- Original Message -----
From: "MURAT CETINKAYA" <cetinkayam_at_psu.edu>
To: <amber_at_scripps.edu>
Sent: Monday, October 03, 2005 10:04 PM
Subject: Re: AMBER: amber8 parallel sander

> Hi Kathleen,
>
> I genereally use 16nodes (dual processor) for the optimum speed. I tried
> 32
> nodes, but there was no change in the computation speed, however I used
> twice
> as much resources. In my opinion you should find the optimum node# by
> trial.
> After a point I think, your system's speed will be counterbalanced with
> the
> time spent during the communication between nodes.
> I am not an expert in parallel computing, but this is the situation I had
> in my
> case.
>
> hope this helps.
>
> On Mon, 03 Oct 2005 14:37:47 +0000, "Kathleen Erickson" wrote:
>
>> Hi Carlos:
>>
>> If you don't mind my asking, is scaling efficiency a concern? I'm
>> curious
>> to know what % of n# of nodes/processors you are able to scale to in
>> order
>> to identify the satisfactory level of scaling efficiency. In other
>> words,
>> if you are getting only 20% scaling efficiency out of 160 nodes, is that
>> sufficient for your needs or do you need more efficiency out of your
>> existing nodes/processors?
>>
>> Anyone else have an opinion?- Kathleen
>>
>>
>>
>>
>>
>> Hi
>> Carlos:
>>
>> If you
>> don't mind my asking, is scaling efficiency a concern? I'm
>> curious to know what % of n# of nodes/processors you are able to scale to
>> in
>> order to identify the satisfactory level of scaling efficiency. In
>> other words, if you are getting only 20% scaling efficiency out of 160
>> nodes,
> is
>> that sufficient for your needs or do you need more efficiency out of your
>> existing nodes/processors?
>>
>> Anyone
>> else have an opinion?- Kathleen
>
> Murat CETINKAYA, M.Sc.
> Biomolecular Materials Lab,
> Dept. of Engr. Science and Mechanics,
> The Pennsylvania State University,
> University Park, PA 16802
> office: (814) 863 9967
> web: www.personal.psu.edu/muc176
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber_at_scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu
>

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu