Sean> I'm not understanding the issue here. If connections are
Sean> being made faster than QPs are leaving the time wait state,
Sean> then the system will eventually run out of resources. But
Sean> this problem seems somewhat separate from the threading
Sean> model used to establish connections, unless that thread is
Sean> preventing other threads from executing.
Sorry I wasn't clearer. The problem I was trying to describe (which
incidentally has been seen with a real application) is that if MADs
like CM REQs are processed in one queue, and time wait expirations are
processed in a separate queue, then it's possible for the MAD queue +
application to starve the time wait queue. This means a larger and
larger backlog of time wait expirations accumulates and eventually the
system runs out of resources, even though the application only keeps a
constant number of QPs in use.
Sean> If that's the case, is it be worth considering exposing the
Sean> MAD work queue for use by not just the MAD layer, but also
Sean> specific clients, such as the CM and SA client code?
That would be another solution. However it seems reasonable to let
clients tune their work processing model to their needs (and avoid
having them clog up the MAD queue).
- R.
_______________________________________________
openib-general mailing list
[EMAIL PROTECTED]
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
