On Wed, Jul 02, 2008 at 04:24:25PM +0300, Or Gerlitz wrote: > OK, I understand that to some extent its fine to do what others (BTW its > only 4 drivers under drivers/net that use LRO, so the sample size is kind > of small) do. I still don't see why 64 makes sense, for example ipoib > supports LSO on the xmit side, which means the network stack would submit > SKBs whose data size is up to 64K, assuming the common mtu is 2K-epsilon I > would use a value of 64K/2K = ~32 for the max_aggr value, why use 64 ?! > I don't see a connection to using or not using LSO -- the 64K comes from the fact that an IP datagram's max length is 64K. And since LRO involves generating at the receiver a hand made IP datagram, the max length is 64K. There is a connection to TCP MSS which could be up to MTU but could be smaller too. I don't see a problem keeping the current value. _______________________________________________ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
