Re: namd on AWS

From: Miro Astore (miro.astore_at_gmail.com)
Date: Wed Oct 02 2019 - 15:47:38 CDT

So you're getting a performance penalty too?

I spoke to some people from AWS and they seemed surprised and skeptical
that things were slower on their platform. Perhaps there is a system I
could benchmark to show them.

Best, Miro

On Thu., 3 Oct. 2019, 02:02 Sanjay Hari, <shari_at_universalsequencing.com>
wrote:

> That sounds about right. I get 0.06 days/ns on a system with a fourth as
> many atoms, on a p2.xlarge with 4 cpus. I think the main advantages are
> cost (overnight run on a spot instance was $3) and not having any overhead
> or maintenance for maintaining a cluster.
>
>
>
> *From:* owner-namd-l_at_ks.uiuc.edu <owner-namd-l_at_ks.uiuc.edu> *On Behalf Of
> *Miro Astore
> *Sent:* Tuesday, October 1, 2019 11:47 PM
> *To:* Namd Mailing List <namd-l_at_ks.uiuc.edu>
> *Subject:* Fwd: namd-l: namd on AWS
>
>
>
>
>
> ---------- Forwarded message ---------
> From: *Miro Astore* <miro.astore_at_gmail.com>
> Date: Wed, 2 Oct 2019 at 11:04
> Subject: Re: namd-l: namd on AWS
> To: Sanjay Hari <shari_at_universalsequencing.com>
>
>
>
> Thanks Sanjay,
>
>
>
> I have tried a few things and experimented with numactl. At the moment I'm
> using, on a p3.8xlarge
>
> charmrun namd +p8 +devices 3 smd.conf | tee smd.conf.log &
>
>
>
> and my benchmark is 0.141994 days/ns
>
>
>
> Best, Miro
>
>
>
> On Tue, 1 Oct 2019 at 20:44, Sanjay Hari <shari_at_universalsequencing.com>
> wrote:
>
> I really like the AMI. Some details would help here, such as your run
> command, number of processors invoked, and benchmark statistics. The gpu
> tutorial gives a good primer on benchmarking cuda-accelerated runs.
> ------------------------------
>
> *From:* owner-namd-l_at_ks.uiuc.edu <owner-namd-l_at_ks.uiuc.edu> on behalf of
> Miro Astore <miro.astore_at_gmail.com>
> *Sent:* Tuesday, October 1, 2019 1:02:14 AM
> *To:* Namd Mailing List <namd-l_at_ks.uiuc.edu>
> *Subject:* namd-l: namd on AWS
>
>
>
> Hello everyone.
>
>
>
> I am trying to run the namd AMI package and I'm not getting very good
> performance. I've tried a few types of p2 and p3 node types and I can't get
> comparable performance to my university's HPC clusters when running 2
> workloads on the same node. ~200000 atom system. The performance doesn't
> appear to be very different from copying over the precompiled binary from
> the website. Just wondering if anyone else has encountered this issue.
>
>
>
> https://aws.amazon.com/marketplace/pp/B06VSMMHRY
>
>
>
> Completely possible I'm just not running things in the correct way, so I
> would be very interested to see how you guys run things.
>
>
>
> Best, Miro
>
>

This archive was generated by hypermail 2.1.6 : Tue Dec 31 2019 - 23:20:58 CST