<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 3.2//EN">
<HTML>
<HEAD>
<META HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=us-ascii">
<META NAME="Generator" CONTENT="MS Exchange Server version 5.5.2654.89">
<TITLE>RE: [openib-general] ip over ib throughtput</TITLE>
</HEAD>
<BODY>
<BR>
<P><FONT SIZE=2>> -----Original Message-----</FONT>
<BR><FONT SIZE=2>> From: Woodruff, Robert J [<A HREF="mailto:robert.j.woodruff@intel.com">mailto:robert.j.woodruff@intel.com</A>] </FONT>
<BR><FONT SIZE=2>> Sent: Wednesday, January 05, 2005 3:53 PM</FONT>
<BR><FONT SIZE=2>> To: Diego Crupnicoff; Hal Rosenstock; Peter Buckingham</FONT>
<BR><FONT SIZE=2>> Cc: openib-general@openib.org</FONT>
<BR><FONT SIZE=2>> Subject: RE: [openib-general] ip over ib throughtput</FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> The trade off is that a fully connected model requires each </FONT>
<BR><FONT SIZE=2>> node to burn a QP for every node in the cluster and thus does </FONT>
<BR><FONT SIZE=2>> not scale as well as the UD model. My guess is that if </FONT>
<BR><FONT SIZE=2>> people need really high performance </FONT>
<BR><FONT SIZE=2>> socket access, they will use SDP instead.</FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> woody</FONT>
<BR><FONT SIZE=2>> </FONT>
</P>
<P><FONT SIZE=2>1 QP for every node in the cluster does not sound that bad. </FONT>
</P>
<P><FONT SIZE=2>SDP is a good alternative too. It has even further benefits as compared to IPoIB (built in HW reliability that eliminates the TCP/IP stack, potential for zero copy, etc). However, in terms of QP requirements, SDP would consume even more than what a connected mode IPoIB would (still not too bad given the IB HW capabilities).</FONT></P>
<P><FONT SIZE=2>Diego</FONT>
</P>
<BR>
<BR>
</BODY>
</HTML>