- From: Jeffrey Mogul <mogul@pa.dec.com>
- Date: Fri, 16 Aug 96 13:50:33 MDT
- To: Paul Leach <paulle@microsoft.com>
- Cc: http-wg%cuckoo.hpl.hp.com@hplb.hpl.hp.com
Koen writes: But when removing the cache entry, _the proxy does not necessarily have an open persistent connection to the origin server in question_. I would be more happy if the draft explicitly allowed a proxy to report the hit count of a removed cache entry within, say, one day of removing it. This would allow the proxy to batch the HEAD requests for each server and send them at night, when the there is enough bandwidth left to make opening extra connections not hurt. Paul writes: Sure -- it already says it need not wait for the HEAD request to finish before removing the item; it wouldn't hurt to mention that it can defer (and batch) the requests as long as there is a high probability that it will get made. I see no problem with saying that the proxy can defer and batch the final-report HEAD requests but I do not believe that it makes sense to have this delayed until "at night" (which is a problematic concept, since it assumes that the proxy actually knows when more bandwidth will be available). I'm worried that long delays in updating this information will discourage origin servers from using the mechanism, and I'm also worried that any mechanism that causes all of the proxies within a particular time zone to spontaneously generate a large stream of requests at a particular point in time will cause the nasty synchronized congestion that Floyd and Jacobson warned about in Proc. SIGCOMM '93. (One could avoid this by some randomization, but they point out that this is much harder than one might suppose.) So my suggestion is that the spec says something like: A proxy MAY defer the final-report HEAD request to an origin server, and send a batch of these later, either piggybacked on its next request to the server or when sufficient resources are available. A proxy SHOULD NOT delay these reports longer than absolutely necessary, to avoid discouraging the use of hit-metering. A proxy SHOULD NOT delay these reports if there is a significant probability that it will be unable to deliver them after the delay. A proxy MUST NOT delay these reports for delivery at a non-randomly-chosen specific time of day; this is necessary to prevent unnecessary congestion. -Jeff
Received on Friday, 16 August 1996 13:59:51 UTC