David J N Begley writes:
>When Squid requests data from a remote site, surely it's up to the remote site
>(and the intervening network) to determine how fast the data will be
>delivered? After all HTTP 1.0/1.1 isn't a windowed/packet protocol (like
>Zmodem file transfers, for example) so Squid has no choice but to sit
>listening for data as quickly as it can be delivered across the network.
Yes, but it doesn't have to read from the socket. The data
will get queued up in the TCP stack, up to SO_RCVBUF
bytes.
Squid-2 doesn't read more than READ_AHEAD_GAP (16k)
bytes ahead of the fastest client.
If the "quick_abort" settings allow downloads to
continue after clients abort, then Squid reads from the
server-side as fast as the data comes in. This could explain
the spikes.
I don't remember hearing what the "quick_abort" settings were
on the caches that saw the spikes.
Duane W.
Received on Sun Jan 31 1999 - 11:24:03 MST
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:44:11 MST