Posted by Evgeniy on November 28, 2009 at 10:14:38:
In Reply to: Re: I/O retardations (problems) in seward posted by Evgeniy on November 27, 2009 at 17:46:09:
: :
: : >I observe this on a big cluster which has a global scratch disk.
: : If this global scratch disk is mounted as NFS, of course, I/O will be incredibly slow.
: : Even if your system administrator claims that this global scratch area is fantastically fast - you MUST use local disks for WorkDir(s).
: : If on your cluster there is no local hard disks (which is typical for e.g. BG solutions) - I am sorry, you have to find another computing system.
:
: Thanks for your reply!
: Unfortunately there are no local disks on the cluster nodes. But the global scratch disk has average rate ~500 MB/s. I am not sure but I think the problem was in a too small buffer size for CD. When I set VBUFfer to 1.0 I got increase by ~50-80 times. This supports the conjecture that I/O was effected only as for the seward module (as I wrote before even caspt2 which is very sensitive to I/O runs pretty fast).
Sorry, I was wrong concerning the too small value for VBUFfer. The seward indeed finishes fast but produces something wrong in this case (the following scf does not work). I guess it does so because I use aCD (RICD option) in the gateway which precedes the seward (sorry I forgot to mention this in the previous posts). The problem thus remains :(.
Best regards,
Evgeniy