Re: Re: namd2.7b1 slower with openmpi 1.3.2 than namd2.6 withopenmpi 1.2.6

From: Axel Kohlmeyer (akohlmey_at_cmm.chem.upenn.edu)
Date: Sun Jul 05 2009 - 11:38:37 CDT

On Sun, 2009-07-05 at 10:17 +0800, liupeng012 wrote:
>
> Dear alex:
> yes! i use the command: ompi_info -a, check the setting of opneib.
>
> there is a plenty of parameter.
>
> which one is significant ?

hard to say in an e-mail. the basic answer is:
it depends. it is best to check with the openmpi
folks and their documentation for details.

on our machine setting btl_openib_use_srq to 1
was helping. there are several others that are
not part of the openib module, but related, like
mpi_paffinity_alone (which requires compilation
with libnuma support), or coll_tuned_use_dynamic_rules
which reads in yet another file and allows to
tweak at which message size which collective
algorithm will be used.

the only way to find out, is to do systematic tests
and benchmarks with realistic inputs and job sizes.

cheers,
   axel.

>
>
>
>
> 2009-07-05
>
> ______________________________________________________________________
> liupeng012
>
> ______________________________________________________________________
> 发件人: Axel Kohlmeyer
> 发送时间: 2009-07-04 18:56:06
> 收件人: Vlad Cojocaru
> 抄送: NAMD list
> 主题: Re: namd-l: namd2.7b1 slower with openmpi 1.3.2 than namd2.6
> withopenmpi 1.2.6
>
>
> dear vlad,
>
> > Both openmpi and mvapich2 I compiled myself with infiniband support
> > using same compiler. the flags used for the compilation of openmpi 1.3.2
> > were exactly the same as those used for the old compilation of openmpi
> > 1.2.6.
> >
> > Is there a reason for the drop in performance when compiled with openmpi
> > 1.3.2 ?
>
> i've heard similar reports with other codes. perhaps openmpi 1.3
> and 1.2 have somewhat different default settings and efficiencies for
> the choice of some algorithms (particularly collectives). depending on
> hardware (OpenMPI does adjust some of its settings to the individual
> HCA) you may just run into less optimal algorithms for your particular
> usage and hardware. there are a large number of parameters in the
> openib layer of OpenMPI that can be used to tweak the algorithms.
> check out: ompi_info -a
>
> cheers,
> axel.
>
> > Best wishes
> > Vlad
> >
> >
>
> --
> =======================================================================
> Axel Kohlmeyer akohlmey_at_cmm.chem.upenn.edu http://www.cmm.upenn.edu
> Center for Molecular Modeling -- University of Pennsylvania
> Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
> tel: 1-215-898-1582, fax: 1-215-573-6233, office-tel: 1-215-898-5425
> =======================================================================
> If you make something idiot-proof, the universe creates a better idiot.
>
>
> __________ Information from ESET NOD32 Antivirus, version of virus signature database 4215 (20090704) __________
>
> The message was checked by ESET NOD32 Antivirus.
>
> http://www.eset.com
>
>
>

-- 
=======================================================================
Axel Kohlmeyer   akohlmey_at_cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:51:10 CST