We are using the Netview MLM on our system, and whenever we start up
midmand, it destroys NFS performance on the system. We've traced it
to one possibility: a possible conflict between 'portmap' and
'midmand'. Using 'lsof', we found portmap and midmand both using
UDP:*.* (whatever port *that* is...), and when this happens, the
biods (NFS client daemons) don't use any CPU, and NFS performance is
terrible (several seconds for a 'df', for example). Killing midmand
clears up the problem.
I suspect that the UDP:*.* port is being fought over by midmand and
portmap (which acts as a broker for rpc services, including NFS).
Then, when packets get grabbed by midmand, the portmap never sees
them, and we're in a timeout condition for the NFS application. The
bad performance seems random (sometimes it's OK), which might indicate
that the midmand was off doing something else, and the portman had a
chance to get at the socket.
Has anyone else seen this problem? Any ideas for a work-around?
FYI: AIX 126.96.36.199, mlm 188.8.131.52.U445257, RS/6k-380/192MB RAM/584MB Swap,
very lightly loaded (just us chickens!).
Steve_Casagrande AT be.ibm DOT com