- From: Jeffrey Mogul <mogul@pa.dec.com>
- Date: Wed, 23 Aug 95 13:56:07 MDT
- To: Dave Kristol <dmk@allegra.att.com>
- Cc: http-wg%cuckoo.hpl.hp.com@hplb.hpl.hp.com
DMK wrote: Jeffrey Mogul <mogul@pa.dec.com> wrote (Fri Aug 18 21:11:10 1995): > [dmk wrote:] > 1) If there's a Connection: keepalive request header, the server will > hold the connection open for 10 seconds. > > This might be a bit short. My trace-based curves show a fairly > sharp "knee" in mean requests/connection at somewhat higher timeouts, > around 1-2 minutes. As you mentioned in your paper there are two sources of such requests: requests for inlined images and subsequent hits on links by a user. I'm shooting for the first batch, I admit. In the future, with authentication and payment additions to HTTP, the short keepalive will address a larger proportion of the same-server traffic. Do you have any measurements for which of the two sources produces more of the same-server hits? I did not break down the traces by file name, so I don't have explicit results for inlined images vs. subsequent clicks. But in the paper I plot "requests arriving for already-open connections" vs. idle timeout. Generally, a large fraction of the gain (almost half) comes with timeouts greater than 10 seconds. This implies either of two things: (1) Many clients were delaying ca. 10 seconds before retrieving inlined images. (2) there were a lot of subsequent hits in the 10-100 second range. Hypothesis #1 seems rather unlikely, so I'd bet on #2. -Jeff
Received on Wednesday, 23 August 1995 14:08:59 UTC