[gmx-users] several runs
pall.szilard at gmail.com
Tue Feb 24 23:53:07 CET 2015
On Tue, Feb 24, 2015 at 10:54 PM, Victor Rosas Garcia
<rosas.victor at gmail.com> wrote:
> Perhaps I can be of help here.
> Except for the smallest systems, GROMACS simulations are very
> compute-intensive, so a single machine is needed for a single job. In
> many instances, several machines are required for a single job
> (provided you have a fast enough network).
> If you give a single machine more than one compute job, or if you
> require of it more computing power than if physically has (like asking
> for 8 cpus when the machine only has 6) that's called oversubscribing
> the machine. It causes severe performance degradation, as the machine
> has more work than it can handle. Not only in terms of CPU, but also
> because the input/output channels will saturate
> So, if a single simulation requires all the resources in your machine,
> running several will make it unusable.
That is true and most of the above is in general reasonable, but...
> It is a lot better to run a single job per machine.
this very much depends on the kind of simulations you're doing and the
machine you're using!
[Most often] Parallelizing isn't free and scaling isn't perfect!
Hence, if you don't have a big enough problem and/or you anyway have
multiple runs to do on a fixed amount of hardware, you may as well do
it in parallel and benefit from more efficient but "narrower" runs
with an end result of increased aggregate simulation throughput.
Think of single molecule solvation FE calculation with FEP. You'll
have 10-20 simulations to run, but the input will likely be quite
small (few thousand atoms). If you have a 4-socket 16-core machine,
you will definitely not want to parallelize this tiny system on 64
cores, but rather run e.g. 16 runs on 4 cores each. The same applies
even if you have only 8 cores! Running 8 single core runs will give
you better aggregate throughput than running eight runs sequentially
on 8-cores each.
So to conclude, it is *not always* better to run a single job per
machine - especially if you have multiple independent runs to do a
limited amount of resources.
> Hope this helps.
> 2015-02-24 13:40 GMT-06:00 mah maz <mahmaz71 at gmail.com>:
>> Hi Justin,
>> Thank you for your answer! If you may help with any of these questions I
>> would be greatly grateful;
>> How can I understand if they were interfered? How is pinning? Any other
>> ways you can recommend?
>> Thanks a lot
>> On Tue, Feb 24, 2015 at 10:14 PM, mah maz <mahmaz71 at gmail.com> wrote:
>>> Is running simulations in several terminals problematic?
>>> On Tue, Feb 24, 2015 at 8:42 PM, mah maz <mahmaz71 at gmail.com> wrote:
>>>> Dear all,
>>>> How can I perform several simulations simultaneously (in linux)?
>>>> thank you!
>> Gromacs Users mailing list
>> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting!
>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>> * For (un)subscribe requests visit
>> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-request at gromacs.org.
> Gromacs Users mailing list
> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting!
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-request at gromacs.org.
More information about the gromacs.org_gmx-users