All other things being equal, it is better not to put packets into the
network faster than it can drain them out. Large bursts increase delay
variation, and increase the probability that two or more packets in a
connection will be dropped within an RTT (not every box is implementing
AQM yet). New 10Gig-switch-on-a-chip devices like Fujitsu's MB87Q3140
(http://www.fujitsu.com/us/services/edevices/microelectronics/networkingassps/mb87q3140/)
have only limited on-chip buffer memory. So transmit packet pacing is
preferred; see
http://yuba.stanford.edu/~yganjali/research/publications/Very-Small-Buffers-CCR.pdf
for further arguments.
True. This is the first thing we learnt when dealing with 10Gbps.
Chelsio's TOE is capable of micro-second granularity pacing and that has
been crucial for high performance. Even then, it's very hard to fully avoid
packet loss. Hardware TCP and retransmission straight from the NIC turn out
to be very helpful in quickly recovering from loss. In fact. even interrupt
moderation delays show up in degraded performance.
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
