nemesis ext_procs optimization about mpich HOT 7 CLOSED

Originally by Dave Goodell on 2008-08-01 08:42:34 -0500

This message has 0 attachment(s)

from mpich.

Originally by thakur on 2008-09-15 14:34:01 -0500

Darius will check if this is already fixed.

from mpich.

Originally by buntinas on 2008-09-16 12:37:32 -0500

This is still an issue in 1.0 and 1.1, but since it's performance issue, I think we shouldn't hold up 1.0.8 for this.

When we add support for multiple netmods, we'll have a list of "active" netmods, and only call poll on those netmods. Doing that will resolve this issue, so we should just leave this until then.

-d

from mpich.

Originally by balaji on 2009-03-03 23:21:41 -0600

OSU reported nearly a 0.5us increase in latency here. The additional latency could be because of other reasons too, but this note is to make sure we do a compare against 1.0.8 for performance.

from mpich.

Originally by buntinas on 2009-05-07 14:39:55 -0500

Fixed in [888cb39].

The original plan was to poll the network only when there is an external process (not on this node), or while a port is open (by MPI_Open_port). The problem is that through some communicator creation magic, a process may end up belonging to a communicator with spawned or connected processes, even though it has never called spawn or connect. So keeping track of whether there are external processes ends up being pretty hairy.

Instead, we took a different approach, and reduced the polling frequency for the tcp module. If a process hasn't had any network activity (nothing from the listener socket and no connect requests), then the poll period is very large (1<<22 for now). As soon as some activity is detected, we reduce the polling period to something smaller (currently 128). Note that in this method, because we don't know whether another process might try to connect to us, we still need to poll once in a while, even if we haven't initiated a network connection ourselves.

-d

from mpich.

Originally by jayesh on 2009-05-07 14:44:44 -0500

Since the changes are in the tcp network module, I need to port the changes to wintcp.

Regards,
Jayesh

from mpich.

Originally by jayesh on 2009-05-11 14:38:19 -0500

Ported the changes to windows netmod in [56c991e]

-jayesh

from mpich.