<html><head><style type='text/css'>p { margin: 0; }</style></head><body><div style='font-family: Arial; font-size: 12pt; color: #000000'><br>>On Dec 12, 2009 Håkon Bugge wrote:<br>><br>>On Dec 12, 2009, at 7:59 , Rahul Nabar wrote:<br>><br>>> I have seen a considerable performance boost for my codes by using<br>>> Jumbo Frames. But are there any systematic tools or strategies to<br>>> select the optimum MTU size? I have it set as 9000. (Of course, all<br>>> switiching hardware supports jumbo frames and no talking to the<br>>> external world required of the interfaces) Have you guys found<br>>> performance to be MTU sensitive?<br>><br>>Once (i.e. several years ago) I tested an application on Gbe and found <br>>that around 1/3rd of the MTU was the optimum. But I guess YMMV.<br><br><div>I would seem that a larger MTU would help in at least two situations,</div><div>clearly applications with very large messages, but also those that</div><div>have transmission bursts of messages below the MTU that could</div><div>take advantage of hardware coalescing. The common MTU of 1500 was</div><div>not chosen arbitrarily, but was probably not tuned to serve the "average"</div><div>HPC application. If one is going to run one application predominately,</div><div>then that application might prefer the default or not. If one is running</div><div>a suite of codes, the it will probably help some and hurt others and </div><div>should be "weighted average" tuned.</div><div><br></div><div>I think the first-order focal point for performance should be on writing</div><div>good code while choosing the "right" MTU size is a third order concern.</div><div>Here is a stupid question ... it is the >>Maximum<< Transmission Unit ...</div><div>Right? You don't get it all the time ... how is the run-time value set</div><div>and adjusted based on the quality of message traffic? Experts ... ??</div><div><br></div><div>rbw</div><div><br>_______________________________________________<br>Beowulf mailing list, Beowulf@beowulf.org sponsored by Penguin Computing<br>To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf<br></div></div></body></html>