From: Paul Reilly (pareilly@tcd.ie)
Date: Thu Jul 03 2003 - 08:47:52 EDT
Since creating a new file domain with 26 filesets (a-z for home dirs)
on a clustered system, the system has been experiencing complete hangs
during busy periods. It fails to respond to network activity, processes
never return, and even typing at the console is not echoed back. I see
console errors like
icscli_handle_get[low_mem]: th=0xfffffc003fd7d180
and so it looks like memory is the issue. The cluster uses a 128 MB
memory channel interconnect. Each node of the 2 node cluster has 1024 MB
RAM, or about 890 MB after bootup claims some wired pages.
My kernel parameters are:
ubc_maxpercent = 70
ubc_minpercent = 10
ubc_borrowpercent = 20
Are the icscli_handle_get messages referring to MC RAM or system RAM?
Do my ubc values look OK, for ensuring there's enough RAM for the system
processes?
HP/Compaq suggested we split the cluster filesystem load between both
cluster members using cfsmgr, but his can only be done a per domain basis,
and not a per fileset basis. Even when split, the situation recurs.
Does anyone know the memory implications of creating extra filedomains and
filesets in a clustered filesystem setup. Would it be better to split my
26 filesets between two filedomains, or should I revert to having a single
file domain and single fileset?
Paul
This archive was generated by hypermail 2.1.7 : Sat Apr 12 2008 - 10:49:25 EDT