Ryan,
I did test the performance difference at one point and I think it is pretty
small -- just a couple percent. Only you can decide whether it is truly
worth all of the hassle.
I should say that I did that measurement quite some time ago, so probably
with AMBER12 or maybe even AMBER11. AMBER14 might take advantage of the
newer SSE instructions to a larger extent and maybe the performance gain is
more significant now.
Brent
On Wed, Jan 14, 2015 at 10:33 AM, Novosielski, Ryan <
novosirj.ca.rutgers.edu
> wrote:
> I had thought of that, but I would have the same question there though:
> wouldn't the code compiled with SSE4.2 run faster on the machines that
> support it? After all, I am trying to get the maximum performance out of
> this stuff. But I am not advanced enough to know whether we are talking
> about any serious gain between those two instruction sets. Compiling with
> -xSSE4.1 was going to be my fallback. But I am happy to report that setting
> SSE_TYPES and the resultant -ax flags does seem to work fine provided it is
> not being defeated by -xHost. It is clear something should be changed here,
> because otherwise the behavior is counterintuitive.
>
> In any case, on our cluster, the machine with the least hardware features
> does not have a full build environment. It is just a diskless node.
>
> ____ *Note: UMDNJ is now Rutgers-Biomedical and Health Sciences*
> || \\UTGERS |---------------------*O*---------------------
> ||_// Biomedical | Ryan Novosielski - Senior Technologist
> || \\ and Health | novosirj.rutgers.edu<mailto:novosirj.rutgers.edu>-
> 973/972.0922 (2x0922)
> || \\ Sciences | OIRT/High Perf & Res Comp - MSB C630, Newark
> `'
>
> On Jan 14, 2015, at 07:27, Jan-Philip Gehrcke <jgehrcke.googlemail.com
> <mailto:jgehrcke.googlemail.com>> wrote:
>
> Let me propose another quite simple and yet efficient solution on a
> heterogeneous cluster:
>
> Compile on the machine with the least hardware features w/o changing any
> of the configure options (i.e. using xHost).
>
> The resulting build works and uses the intersection of advanced hardware
> features on all machines in your cluster (SSE 4.1 in your case). This is
> what I have been doing many times now, and it really saves a lot of work.
>
> JP
>
> On 13.01.2015 18:20, Novosielski, Ryan wrote:
> Wouldn't there be performance tradeoffs there? I'd think that at least
> supporting 4.1, in my case, would result in faster execution.
>
> ____ *Note: UMDNJ is now Rutgers-Biomedical and Health Sciences*
> || \\UTGERS |---------------------*O*---------------------
> ||_// Biomedical | Ryan Novosielski - Senior Technologist
> || \\ and Health | novosirj.rutgers.edu<mailto:novosirj.rutgers.edu
> ><mailto:novosirj.rutgers.edu>- 973/972.0922 (2x0922)
> || \\ Sciences | OIRT/High Perf & Res Comp - MSB C630, Newark
> `'
>
> On Jan 12, 2015, at 20:45, Daniel Roe <daniel.r.roe.gmail.com<mailto:
> daniel.r.roe.gmail.com><mailto:daniel.r.roe.gmail.com>> wrote:
>
> Hi,
>
> Personally, on clusters like that I usually just configure with the
> '-nosse' flag to avoid the issue altogether.
>
> -Dan
>
> On Monday, January 12, 2015, Novosielski, Ryan <novosirj.ca.rutgers.edu
> <mailto:novosirj.ca.rutgers.edu><mailto:novosirj.ca.rutgers.edu>>
> wrote:
>
> Hi all,
>
> I recently ran into the following error at runtime:
>
> "Please verify that both the operating system and the processor support
> Intel(R) SSE4_2 and POPCNT instructions."
>
> The reason for this is that we have some nodes with Nehalem chipsets and
> some with Harpertown chipsets on our cluster, and the master node is a
> newer machine that supports some of the newer instructions. Looking through
> the —full-help configure option, I saw that I could define SSE_TYPES to say
> which CPU instructions should be supportable at runtime. However, I tried
> building with SSE types set to SSE4.2,SSE4.1 and the reverse even though I
> doubted that would make any difference. It appears as if it does add
> -axSSE4.2,SSE4.1 to the appropriate places during the build, which from my
> read of the Intel documentation is exactly what it should do. However, when
> built, the code behaves the same way with the same error messages on the
> Harpertown nodes. I don’t see a way to disable POPCNT, but it appears as if
> SSE4.1 machines don’t support it, so I would assume it is not used when the
> runtime dispatcher selects SSE4.1.
>
> I was not aware of auto CPU dispatch before today, so I’m not sure if I’m
> doing something wrong somehow. Amber version is 14, Intel compilers are
> 15.0.1.
>
>
> ____ *Note: UMDNJ is now Rutgers-Biomedical and Health Sciences*
> || \\UTGERS |---------------------*O*---------------------
> ||_// Biomedical | Ryan Novosielski - Senior Technologist
> || \\ and Health | novosirj.rutgers.edu<mailto:novosirj.rutgers.edu
> ><mailto:novosirj.rutgers.edu> <javascript:;> - 973/972.0922
> (2x0922)
> || \\ Sciences | OIRT/High Perf & Res Comp - MSB C630, Newark
> `'
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org<mailto:AMBER.ambermd.org><mailto:AMBER.ambermd.org>
> <javascript:;>
> http://lists.ambermd.org/mailman/listinfo/amber
>
>
>
> --
> -------------------------
> Daniel R. Roe, PhD
> Department of Medicinal Chemistry
> University of Utah
> 30 South 2000 East, Room 307
> Salt Lake City, UT 84112-5820
> http://home.chpc.utah.edu/~cheatham/
> (801) 587-9652
> (801) 585-6208 (Fax)
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org<mailto:AMBER.ambermd.org><mailto:AMBER.ambermd.org>
> http://lists.ambermd.org/mailman/listinfo/amber
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org<mailto:AMBER.ambermd.org>
> http://lists.ambermd.org/mailman/listinfo/amber
>
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org<mailto:AMBER.ambermd.org>
> http://lists.ambermd.org/mailman/listinfo/amber
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
--
_______________________________________________
Brent P. Krueger.....................phone: 616 395 7629
Professor................................fax: 616 395 7118
Hope College..........................Schaap Hall 2120
Department of Chemistry
Holland, MI 49423
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Jan 14 2015 - 08:00:04 PST
