On 6/20/05, Grant Grundler <[EMAIL PROTECTED]> wrote: > On Mon, Jun 20, 2005 at 06:14:04PM +0300, Michael S. Tsirkin wrote: > > Reducing the rx ring size gives me bandwidth win of about 1%. > > Michael, > Doesn't this sound counter-intuitive to you? > > Can you share which configuration (chipset, CPU, IO bus, HCA) > you measured this on? > > I've also seen this to be true with gige - but only in one odd configuration. > > thanks, > grant
I'd be skeptical also. I've found this to be true in only one very contrived instance. When receiving UDP traffic on a slow host (no user level end to end flow control), you can improve receive throughput by reducing the number of outstanding receives. The CPU can be saturated receiving packets which are overflowing the socket (being discarded), and the user can be CPU starved (unable to empty the socket). In this case, you are better off having the HCA miss the packets than you are having the driver process each packet, only to discard it in the upper layers. This is not a real world situation, though. -- Bill Jordan SilverStorm Technologies _______________________________________________ openib-general mailing list openib-general@openib.org http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general