<div dir="ltr"><br><div class="gmail_extra"><br><br><div class="gmail_quote">On Sun, May 26, 2013 at 9:50 AM, Matt Mathis <span dir="ltr"><<a href="mailto:mattmathis@google.com" target="_blank">mattmathis@google.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="im"><br>
</div></div><div class="gmail_extra"><br></div><div class="gmail_extra">As for link layer flow control, it doesn't work. You can buy it today in many technologies, but if you turn it on, you will rediscover head-of-line blocking and global resource contention. e.g. If you have one overloaded exit, the cascaded backlogs at each stage through the fabric guarantee congestion and/or resource starvation everywhere, even if the rest of the fabric is otherwise lightly loaded.</div>
<div class="gmail_extra"><br></div></div></blockquote><div style>I don't know that I would say that LLFC doesn't work. It works for what is it designed to do...operate within usec to very-short term congestion to guarantee a near lossless link network. It is definitely not designed to handle sustained overloads...end-to-end flow control works best there.</div>
<div style><br></div><div style>This presentations discusses some of these issues on slides 16-23.</div><div style><a href="http://www.c3.lanl.gov/cac2007/presentations/hendel.pdf">http://www.c3.lanl.gov/cac2007/presentations/hendel.pdf</a><br>
</div><div style><br></div><div style>Anoop</div></div></div></div>