- From: Jeffrey Mogul <mogul@pa.dec.com>
- Date: Fri, 16 Dec 94 14:04:33 PST
- To: Simon E Spero <ses@tipper.oit.unc.edu>
- Cc: http-wg%cuckoo.hpl.hp.com@hplb.hpl.hp.com
What value were you using for 2MSL in the traces?
The "legal" value, 240 seconds. Although my simulator allows me
to change this. And (shhh, don't tell anyone) I think DEC OSF/1,
which was what we were actually running, uses 120 seconds.
If you want a few simulations done with different values, let me
know. I suspect the main difference is that the number of TIME_WAIT
entries depends approximately linearly on this value.
Also, did you assume that the effective bandwidth for a given path
remained the same between the traces and the simulated run?
I assumed that the "request" durations would be the same. Since
the server CPU and elapsed times were a tiny component of the overall
elapsed times, server loading should not matter. On the other hand,
you're right to point out (I assume this is what you mean) that the
actual durations would likely be shorter, since the requests would
go faster without the cost of connection setup or slow-start.
Shorter connections should generally mean fewer active ones at once,
which is "good". So my failure to account for this effect is a
conservative error, given that my goal is to argue in favor of
persistent connections.
-Jeff
Received on Friday, 16 December 1994 14:10:45 UTC