My reading of rfc2001 is that during slow start cwnd should open
approximately exponentially (1->2->4->8->...), not linearly
(1->2->3->4->...). I agree that counting ACKs (instead of ACKed
packets) is an understandable simplification, even though I'm told that
Jacobson's original algorithm counted ACKed packets. The rfc2001
acknowledges that the "counting ACKs" does not produce exactly
exponentially opening cwnd because of delayed ACKs, but it states that
typically one ACK will be received for every two segments sent.
This may be true for large packets, but it takes a LOT of tiny packets
to make two segments. For example, our MSS=1500 and length of TCP
packet carrying 1 byte is 41 bytes, so 73 such packets can be sent and
*still* not reach the 2*MSS limit which would force an ACK.
> I don't know what 30s timer you're refering to (you don't mean the probe timer,
> do you?)
No. This timer appears in our kernel 2.0.36 but seems to be gone by
2.2.2. Check tcp_output.c (kernel 2.0.36) and look for the line No. 297
which reads
sk->partial_timer.expires = jiffies+30*HZ;
(This line has puzzled other people in the past, which may explain why
it is gone in 2.2)
> I will try to get sysctls to turn off delayed ack and a reasonable way to
> increase initial cwnd into 2.2/2.3 (reasonable for cwnd should at least
> include a recompile, because otherwise it is too easy for users to cause too
> much harm)
While the *initial* cwnd=3 would help a bit, please note that rfc2414
allows this only if there were *no* retransmits during the initial
"three-way handshake". Moreover, the initial cwnd is quickly adjusted
upwards, so a larger initial value helps only once.
A more significant benefit of rfc2414 is that a large cwnd, once
established, is not dropped all the way to 1 the moment the socket has
been idle for an RTO period. This can happen very often during a
session, forcing lots of slow starts from 1. The rfc2414 suggests
reducing the cwnd*MSS window to a more reasonable value of
min(4*MSS,max(2*MSS,4380)) bytes (which works out to cwnd=3 in our
case).
Actually, we do the following in tcp_output.c (kernel 2.0.36) after
being idle for longer than sk->rto:
if (sk->cong_window > 3) sk->cong_window = max(3, sk->cong_window >> 1);
This works just fine. For example, cwnd=2048 is reduced to 3 in only 10
steps, which fits well with our 10 times shorter RTO floor. Retransmits
could have forced cwnd=1, so we do not touch cwnd<3 after an idle
period. Of course, "3" above should be replaced by the value computed
for the particular socket's MSS.
Sincerely,
Josip
-- Dr. Josip Loncaric, Senior Staff Scientist mailto:josip@icase.edu ICASE, Mail Stop 132C http://www.icase.edu/~josip/ NASA Langley Research Center mailto:j.loncaric@larc.nasa.gov Hampton, VA 23681-2199, USA Tel. +1 757 864-2192 Fax +1 757 864-6134- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/