Hal Rosenstock wrote:
Hi Yevgeny,

On Mon, 2007-05-14 at 09:02, Yevgeny Kliteynik wrote:
Hi Hal,

This patch optimizes fabric ranking.
All the leaf switches are marked with rank and added to the BFS list,
and only then ranking of rest of the fabric begins.

I actually thought that this is the way I've originally
implemented it, as I mentioned in the patch that was dealing with 8 and 16 bit integers :)

Similar optimization may be applicable to up/dn routing - the roots
should be marked with rank 0 and only then ranking of rest of the switches should begin, but frankly, it practically doesn't reduce the routing time, because ranking is only a small fraction of the routing runtime (I checked it on a 4K+ subnet).

It's still worth doing IMO. Can you look into this for up/down ?

Sure.

--Yevgeny


In case of fat-tree I'm going to need it anyway when I enhance
the routing to consider only subset of HCAs in the routing balancing
(compute nodes vs. management nodes).

Please apply to master.

-- Yevgeny

Signed-off-by:  Yevgeny Kliteynik <[EMAIL PROTECTED]>

>From dfa455f86d9ac48ff5cefd38a009718e5aeab1f9 Mon Sep 17 00:00:00 2001
From: Yevgeny Kliteynik <[EMAIL PROTECTED]>
Date: Mon, 14 May 2007 15:45:00 +0300
Subject: [PATCH] DELETE AFTER UPDATE: ranking

Signed-off-by: Yevgeny Kliteynik <[EMAIL PROTECTED]>

Thanks. Applied (to master only).

-- Hal



_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general

To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general

Reply via email to