Charm++ parallel programming system

[Sameer Kumar <sameerk AT us.ibm.com>]

The NAMD 2.7b1 binary from the NAMD
website should be optimized and perform optimally at least for the APoA1
benchmark.  It may be missing some optimization that your system might
need?  You should try to verify if you get the same APoA1 results
as mentioned by Abhinav below.   May be Abhinav has access to the
APoA1 input files and optimized config files.

Benchmark time in the output file is
the correct reflection of performance.  You should get the similar
performance by running atleast 30-50K time steps and reducing loadbalancing
frequency using the following option to the config file.

LdbPeriod  50000  #should
be sufficient for BG/P as there is very little system noise

-Sameer.

Aditya Devarakonda <aditya08 AT cac.rutgers.edu>
Sent by: ppl-bounces AT cs.uiuc.edu

06/28/2012 07:57 PM

To	Abhinav Bhatele <bhatele AT illinoisalumni.org>
cc	charm AT cs.illinois.edu
Subject	Re: [ppl] [charm]  NAMD scaling on BlueGene/P

Thanks for sending these performance numbers, as you suggested
I started building the latest version of NAMD. However, I am seeing some
issues when building charm. Our BG/P system doesn't use the same standard
library paths as some of the other P systems, so I went through the build
configurations and changed the compiler flags.

However, I am still having issues building charm where its throwing an
error when soft linking charm header files and then an "ambiguous
or undefined" reference error. I'm attaching the build log, for detailed
information.

The ambiguous reference error occurs inside of a BG/P
GNU header file so I'm not sure what to make of this and whether the issue
is symptomatic of a system configuration error. If this is just as strange
to you guys, I'll follow up with the BG/P admin to see what's wrong.

Regards,
Aditya

On Wed, Jun 27, 2012 at 5:21 PM, Abhinav Bhatele <bhatele AT illinoisalumni.org>
wrote:
Yes, but the application is sufficiently load balanced
by the time the "Benchmark time" is printed, so that is a good
point in the simulation to compare performance. Sample output for ApoA1
(92K atoms) on 4,096 cores of BG/P:

[bhatele@login2] results_20100720 $ grep Benchmark nopme.4k.output 
Info: Benchmark time: 4096 CPUs 0.00332897 s/step 0.0385298
days/ns 10.6055 MB memory
Info: Benchmark time: 4096 CPUs 0.00293718 s/step 0.0339952
days/ns 10.6055 MB memory
Info: Benchmark time: 4096 CPUs 0.00295251 s/step 0.0341726
days/ns 10.6055 MB memory

NAMD 2.9 has been released although I dont see precompiled
versions for Blue Gene/P on the download page. I would recommend installing
the latest released version. The commands for installing Charm and NAMD
on BG/P are fairly straightfoward:

Charm: ./build charm++ bluegenep xlc --with-production
NAMD: ./config BlueGeneP-xlC; cd BlueGeneP-xlC
&& make

I couldn't find a publication available online that has
performance numbers for ApoA1 on Blue Gene/P, so I am pasting some time
per step (ms) numbers here:

Number of cores      w/o PME    
 w/ PME
1                
           5148.56        5955.26
256                
           23.94        
   27.64
512                
           13.48        
   15.67
1024              
             7.73      
       8.82
2048              
             4.43      
       5.08
4096              
             2.89      
       3.39

The w/ PME numbers are with PME every 4 steps. We can
provide you with files for running the ApoA1 system.

- Abhinav


On Wed, Jun 27, 2012 at 10:37 AM, Aditya Devarakonda <aditya08 AT cac.rutgers.edu>
wrote:
Hey Abhinav,

My output files don't have the usual "Benchmark time:" lines,
but I'm assuming you're referring to the "day/ns", "Wall
time" and "CPU Time" values. I'm using these values to assess
the scalability of the system.
I'm working on sending you details about the system, however this depends
on whether I'm actually allowed to disseminate the info.

I did not build charm/namd, I'm using the pre-build BlueGene/P binaries
from NAMD version 2.7b1

If I remember correctly, you guys are running a specific molecular system
to benchmark various environments - could we have access to these configuration
files.
Any pointers to papers/tech reports with NAMD on BlueGene P would also
be helpful in narrowing down the list of potential issues. I vaguely remember
someone doing runs on Intrepid/Surveyor when I was around last summer.

Regards,
Aditya


On Tue, Jun 26, 2012 at 11:47 PM, Abhinav Bhatele <bhatele AT illinoisalumni.org>
wrote:
Hi Aditya,

Can you send us some performance numbers (grep for Benchmark
on the output file) for different number of cores. Details about which
molecular system you are running (number of atoms, PME frequency etc) would
also help.

Blue Gene/P cores are ~3 times slower than Hopper and
Ranger. Are you seeing performance that is worse than this ratio?

Also, did you build charm and namd yourself? If yes, what
options did you use for the respective builds?

- Abhinav


On Tue, Jun 26, 2012 at 8:00 PM, Aditya Devarakonda <aditya08 AT cac.rutgers.edu>
wrote:
Hi everyone,

I've been working on our new BlueGene/P system at Rutgers
and have been experimenting with NAMD. However, I'm seeing some scaling
issues while running on our system. The performance has been an order of
magnitude worse on the BGP than on Hopper and Ranger, in fact our beowulf
cluster achieves the same performance (the beowulf is actually a bit better).

I'm using the BlueGene P build from NAMD version 2.7b1.
We're trying to narrow down whether this is some issue with the system
or if NAMD on BGP require a lot more cores to show good scaling (I'm currently
scaling up to 8k). Any insights you might have on running NAMD on BGP would
be most helpful.

Thanks,
Aditya

_______________________________________________
ppl mailing list
ppl AT cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/ppl




-- 
Abhinav Bhatele, people.llnl.gov/bhatele1
Center for Applied Scientific Computing, Lawrence Livermore National
Laboratory




-- 
Abhinav Bhatele, people.llnl.gov/bhatele1
Center for Applied Scientific Computing, Lawrence Livermore National
Laboratory
[attachment "charm_buildlog.pdf"
deleted by Sameer Kumar/Watson/IBM] _______________________________________________
charm mailing list
charm AT cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/charm
_______________________________________________
ppl mailing list
ppl AT cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/ppl