>From looking over the code, running some experiments, and from the mailing
list archive, it looks like packets simply get dropped inside the kernel
when the queues overflow.
This seems to interact badly with TCP. If the initial retransmit time out
is 3 seconds, and one of the first few packets gets dropped (before the
rtt estimate is updated), then the connection is stalled for 3 seconds.
(TCP packet gets dropped silently, then kernel waits an RTT for a response
before retransmitting.)
Two questions:
1. Can this really happen, or have I overlooked something? (Our
experiments suggest that it can happen.)
2. If this can happen, is it worth changing? In particular, we might not
want to wait the entire RTT before retransmitting the packet if it was
dropped inside the kernel.
Thanks.
|