[ofa-general] ipoib performance (and xplot)

[Talpey, Thomas]

Interesting data. What app are you using to generate the TCP flow,
and what options are you using on it? Also, what are the scales on
the x- and y-axes (seconds and decimal kilobytes)? I have some
comments but they are only speculation without knowing this.

By "800Mb/s" do you megabytes or megabits? For ipoib, 800 megabytes/s
(MB/s) seems very high and 800 megabits/s (Mb/s) seems very low. In my
experience it gets 200-300 megabytes before running out of cpu (checksum
calculations mainly). But I haven't looked at it in a while.

Tom.

At 03:07 PM 3/6/2007, chas williams - CONTRACTOR wrote:
>while looking at some ipoib performance i had a chance to graph the
>tcp flow in xplot (see http://www.xplot.org/).  the graph appears very
>strange and is attached to this message.
>
>the lower solid line represent acks coming back from the tcp server, the
>up line represent the window size (i disabled window scaling btw, this
>doesnt affect performance).  the up/down arrows (they look like diamonds
>due to scale) represent packets.  this is a view from the tcp client.
>
>the initial part of the graph is tcp slow-start/congestion.  what is
>curious to how the returning acks (after slow start is finished) seem
>to get quiet periodically.  then the next ack that returns, then acks
>the entire window.  this seems to be leading to a very bursty behavior.
>
>i would normally expect to see two data packets followed by an
>ack as can be seen between the 'burst' regions.
>
>i see aboue 800Mb/s (good put) between the hosts which i understand
>to get typical for ipoib.  there dont appear to be any link errors
>either.
>
>so why the long pauses?
>
>ftp://ftp.cmf.nrl.navy.mil/pub/chas/ipoib.jpg
>_______________________________________________
>general mailing list
>general at lists.openfabrics.org
>http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
>
>To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general