- From: Jeffrey Mogul <mogul@pa.dec.com>
- Date: Thu, 21 Nov 96 13:53:20 PST
- To: Ingrid Melve <Ingrid.Melve@uninett.no>
- Cc: http-wg%cuckoo.hpl.hp.com@hplb.hpl.hp.com
The default use includes both usage-limiting and hit-metering, servers are likely to try to take advantage of the combination. Yes, maybe. After all, a server that wants to do some form of usage-limiting has exactly one option today: cache-busting. In our proposal, it has the option to do usage-limiting without entirely breaking caching (by setting the limit above zero) and without eliminating the possibility of prefetching. So I believe our proposal at least gives the option of compromise. If usage-limiting and hit-metering are independent, why have an example that shows how to use usage-limiting in combination with hit-metering? Precisely because they are independent, they can be combined. We gave examples of various combinations of hit-metering and usage-limiting to show that they are, in fact, independent. If a origin server hands out usages only to caches that do hit-metering; and cache-busting documents to the rest (I read your draft as suggesting this is a neat way to do it), then the combination does influence which parent I fetch documents from. The security aspects of this are interesting, because it may force unaware caches to channel their traffic through one branch of the mesh (and possibly take down the top cache server as it gets overloaded). If the origin server keeps track of what usage-limiting happens at different cache servers (as your example suggest), this is a bigger problem. We certainly cannot solve all of the problems arising from trying to simultaneously optimize security, bandwidth, latency, and ease of management. Our proposal does not, however, force a cache to use any particular path through the mesh (i.e., if it has multiple paths, we don't force GETs to follow any specific path, although we would at least expect HEADs to return reports to the appropriate server.) Since we open up another possible dimension for optimization (i.e., one path allows metering, one does not) this makes the optimization problem harder, but the default solution is not any worse. > As I said in my response to Ted Hardie, our specification probably > ought to say explicitly that the proxy needs to record the identity > of the immediate source of a response, and this is another example > where that is important. How will this influence load-sharing? It should not. It only influences who needs to see the hit-meter reports, and it would be entirely acceptable for the proxy to store multiple source-identities if it is willing to do the bookkeeping according to the rules we defined. The implementation becomes somewhat more complex, but this is the tradeoff for trying to optimize things. -Jeff
Received on Thursday, 21 November 1996 17:02:20 UTC