Seaborg Update - I/O Performance

From: James M. Craw (craw_at_nersc.gov)
Date: 09/16/2004


NERSC Users:

A quick update on Seaborg's degraded I/O performance (for some 
operations).  Since earlier this week, NERSC and IBM have been working 
around the clock together to find the root cause of the I/O slowness. 
We believe we have found a major reason for the slowness in the "ml0" 
code.  This code "load balances" switch communication across each node's 
two css adapters.

An initial fix was generated last night and tested on a IBM test system, 
the NERSC development system and on a few nodes of Seaborg today 
successfully.  The E-fix was then pushed out across the entire system 
early this afternoon.

Tests to date, show positive results.  Additional testing will continue 
through the night.  Users are encouraged to run their normal 
applications that can stress the GPFS filesystems (scratch and homes). 
As always, please report any problems to the NERSC consulting staff at 
consult@nersc.gov as soon as you can.

Thank you, again, for your patience.

Jim Craw
NERSC Computational Systems Group Leader

This archive was generated by hypermail 2.1.6 : 08/21/2008 PDT