Re: a vote for asymmetric connections in a session

[julian_satran@il.ibm.com]

I assume that many of us are aware of ECN and the additional requirements
to endpoint TCP stack. As iSCSI is destined for a wide variety of
environments (very much like TCP is)
I think that our position on it was that we can leave those requirements to
TCP and require only an "effective TCP support" for the specific physical
infrastructure.  Imagine only that for reasons of end-to-end connectivity
we will want to support iSCSI (over TCP) in an Inifiniband
network - where ECN is neither supported or required. Would we want to
require ECN?

Julo


Please respond to Fred Baker <fred@cisco.com>

To:   ips@ece.cmu.edu
cc:    (bcc: Julian Satran/Haifa/IBM)
Subject:  Re: a vote for asymmetric connections in a session




At 02:10 PM 9/6/00 -0400, Mike Kazar wrote:
>I think it might be illuminating to try to figure out some algorithms to
>keep the throughput high even when one of the connections drops a packet
>and cuts its TCP window size down significantly.

dumb question. Has anyone looked into the ECN proposal?

It seems like the issue here is the impacts of loss in the TCP state
machine, and those can be mitigated by managing the queues actively. You
could assert that you want ECN to be implemented in any iSCSI path, which
should have the effect of managing queue depths without incurring loss.

http://www.isi.edu/in-notes/rfc2481.txt