[ofa-general] problems with epoll and ipoib (memory leak and cpu creep)

[murray smigel]

Hi,
We have an application using epoll to listen to a group of udp multicast 
broadcasts
that come in over a set of ethernet ports built into a Voltaire ISR2004 
switch via the
IPR module.  The IPR ports are assigned ethernet addresses and routing 
is set up
to listen to the multicasts over the appropriate IPR port.  There are a 
large
number of multicast groups involved ~100 distributed over two IPR ports.

When we run the application using poll, things work fine (except for the 
occasional dropping
of packets due to the large set of fds passed to poll).  To try to 
remedy the problem,
we switched to epoll.  Now, as the program runs, the cpu utilization 
rises over time towards 100% and
the memory usage grows as well.

The same epoll based program runs fine when it is on a machine with 
physical ethernet ports
(eth1 eth2) rather then ipoib mapped IPR ports (ib0.8200, ib0.8600).

Hardware is x86-64 based dualxdual core intel processors.
We are running Debian Etch with a vanilla 2.6.26.5 kernel and the ofed 
stack that is part of the standard
kernel distro.

Any analysis or suggestions would be appreciated.

thanks,
murray smigel