[gmx-users] pp/pme ratio differences between gromacs 5.0.x and 5.1.x
johannes.wagner at h-its.org
Tue Feb 9 12:10:15 CET 2016
thanks, I will look a bit deeper into this! My last reply can then be disregarded.
> On 9 Feb 2016, at 9:20 AM, gromacs.org_gmx-users-request at maillist.sys.kth.se wrote:
>> hey guys,
>> came across an issue with pp/pme ratio difference from 5.0 to 5.1. For a
>> system running on 2 nodes with a pp/pme ratio of 80:16 on 5.0, the same
>> thing starts with 72:24 on gromacs 5.1. The 5.1 run has a much higher load
>> imbalance and is hence slower. Are there known changes from 5.0 to 5.1 in
>> the pp/pme ratio that cause this?
>> cheers, johannes
> Yes and no. Even on only two nodes, your observations can be affected by
> network congestion, so I would investigate whether you can reproduce the
> observations. The changes that I identified that had the capacity to change
> performance are noted at
> http://manual.gromacs.org/documentation/5.1/ReleaseNotes/performance.html <http://manual.gromacs.org/documentation/5.1/ReleaseNotes/performance.html>.
> For example, if you're using LINCS then the SIMD improvements there affect
> only the PP ranks so it might seem to the auto-tuner that it's more
> efficient to use fewer PP ranks. Then later some kind of load imbalance has
> a stronger effect. The interactions of DD DLB and PME tuning are complex,
> and we can't rule out that there's a problem that we should fix. But we
> can't say much without being able to compare some (reproducible) log files.
More information about the gromacs.org_gmx-users