GTX 280 benchmark please...

I’ll be presenting a poster at NVISION, and it would be great to have a GTX 280 benchmark to show off. The problem is, I don’t have one :(

I would greatly appreciate it if someone could run it for me. You can download the software from http://www.ameslab.gov/hoomd/ . Windows users will need to install Python 2.5 if they haven’t already. Run the lj_liquid_bmark.hoomd script which can be found under Start->HOOMD->benchmaks. Linux users should see the installation guide, it mentions how to run the benchmark from the command line.

The script will run for just a few minutes and print “Average TPS: ###” at the end. That number is all I need from the output.

Thanks in advance.

I tried, but using the precompiled build it came down to:

Time 00:02:00 | Step 34694 / 54000 | TPS 254.55 | ETA 00:01:15

***Error! unspecified launch failure after /home/joaander/hoomd-0.7.0/src/computes_gpu/BinnedNeighborListGPU.cc:485

Traceback (most recent call last):

  File "./lj_liquid_bmark.hoomd", line 17, in <module>

    run(50000)

  File "/home/joaander/hoomd-0.7.0/src/share/hoomd/../../lib/python-module/hoomd_script/__init__.py", line 93, in run

RuntimeError: CUDA Error

***Error! unspecified launch failure after /home/joaander/hoomd-0.7.0/src/updaters_gpu/NVTUpdaterGPU.cc:106

terminate called after throwing an instance of 'std::runtime_error'

  what():  CUDA Error

/afsuser/mbach2/tmp/hoomd-0.7.0-Linux-x86_64-Python2.5/bin//hoomd: line 43:  6615 Aborted                 (core dumped) ${HOOMD_ROOT}/lib/hoomd/hoomd $*

:argh: Looks like I have some testing to do once I get a GTX 280. This problem used to show up on G80 and G92, but I had worked around the issue and it has been stable for week+ runs… I guess I’ll need a different workaround on the GTX 280.

It looks like the benchmark made it most of the way through, though and “TPS 254.55” looks promising. Could you attach the entire output as a file so I can calculate an average TPS from the part of the simulation that did complete?

Thanks again.

Hope this is of any use to you.
benchmark.txt (2.26 KB)

Yep: it’s enough to see that the program ran fine up until the error, so I can put a “preliminary” benchmark number on my poster (~50% faster than 8800 GTX woohoo!)

I just got ahold of a GTX 280, and tried running the lj_liquid_bmark.hoomd as well. It fails at the same place, but on my system I only get a TPS around 197. This system has two GPUs in it, though device 0 is the GTX 280.

(A different server with a pair of identical 8800 GTX gets 175.)

Is there a way to check which GPU is being used by HOOMD?

Actually, my system has 3 GPUs, just for the record. However I somehow thought it would only use the first (GTX 280).

:(

It uses CUDA device 0 by default. You can change it on the command line:

hoomd lj_liquid_bmark.hoomd --gpu=#

(# is passed to cudaSetDevice)

When is your poster session? I may be going to the conference.

Tuesday, August 26, 2008 06:30 PM - 08:45 PM

SJCC, Room J3

I’ll also be at NVISION. I hope to meet a lot of you folks of the forum.

Obviously, I will be there too.

Somebody should start an NVISION thread…