Firefly and PC GAMESS-related discussion club
Learn how to ask questions correctly
Re^2: Some problems with Windows HPC 2008 Cluster (MSMPI)
Some years ago I was testing Windows HPC 2008 on internal MS cluster.
Interestingly, I was primary testing MP2 gradient code.
I will share some things I remember.
- The cluster should be based on the IB fabric.
- It would be better if the command line you are using specify the
temporary directory for Firefly directly. Otherwise, if not specified,
this could be Windows root directory.
- For P2P to work normally, you need to completely disable
firewalls on all compute nodes. Note, just setting up rules does not help,
it should be completely disabled
You may also need to explicitly bind P2P to IB network etc...
- Use only local drives for scratch files. NOte you are most
likely using a single drive for I/O coming from four processes
so do not expect brilliantly fast I/O.
- Windows 2008 HPC is basically server version of Windows Vista
or Windows 7 (depending on the release you are using).
Thus, you need to apply all the I/O optimization related to MP2
code running under these operating systems. Yo can find them on
this forum (use search button on the left) and in the documentation
- Windows TCP sockets are a bit subtle, esp. if system
is heavily loaded. You may need to ask administrator to increase
the default number of retries (re-transmissions). As far as I
remember the default is 5 and setting it to 10 cures the problem.
This can be done via registry.
Hope this helps.
On Tue Apr 9 '13 2:38pm, Roman Kroik wrote
>I don't know exactly but it is likely that the Cluster Administrator has made something. And now my calculations finish correctly. But problem with performance remains now. What can You advice me to speed up my calculations on this Windows HPC 2008 Cluster? Is it some command line options or MPI parameters in input files?
>P.S. I'm attaching example ouput file which has finished correctly on the Cluster. Maybe, it need?
Wed Apr 10 '13 1:02pm
This message read 994 times