Hi Rob,
I used parellel hdf5 in HP linux and cluster which is NFS file system. And I have already set
in MPIO mode.In cluster, parellel hdf5 showed extremely low efficient. I think this is due to the cluster setting.In linux, it performed much more better, but is still not as quickly as 1 process when used 4 process.
Thanks for your answers.
Tony
在2009-08-27，robl发给您的邮件内容如下
···
------------------------------------------------------------------------------
From: "Rob Latham"
To: hdf-forum@hdfgroup.org
Date: 2009-08-27 00:25:20
Subject:
On Wed, Aug 26, 2009 at 11:10:46PM +0800, 4 wrote:
When I used parellel hdf5 write speed with 4 process(collective) and 1 process , the results surprised me, the speed of 1 process is even faster than 4 process. I have noticed that when writing data with 4 process, the cpu use ratio drop to about 70% ~ %10, the last cpu seems to be has the minimal efficiency.
It is obviously reasonale, any one has expericece about this, Could you give me some suggestion, and I will check my code carefully.
Hi. What file system are you writing to? If you are writing to NFS,
you won't see much speedup with more clients. Are you using HDF5 in
MPI-IO mode, or are you using posix? I think MPI-IO will give you
better performance, especially if you are using hyperslabs.
==rob
--
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA
_______________________________________________
Hdf-forum is for HDF software users discussion.
Hdf-forum@hdfgroup.org
http://mail.hdfgroup.org/mailman/listinfo/hdf-forum_hdfgroup.org