Date: Fri, 7 Apr 2000 23:35:05 -0400 From: Abdelsalam 'Solom' Heddaya To: Alex Rousskov Cc: polyteam@ircache.net Subject: RE: Boulder meeting minutes Alex, Regarding the minimum hit ratio, I would like to reiterate our position: [1] Connection pass-through is an integral feature to our design of our transparent caches. [2] The current reporting format represents performance completely and accurately: peak throughput indicates the maximum req/sec rate that clients can submit to the system and obtain the associated behavior shown in the rest of the performance metrics. The resulting response time, hit ratio and associated derivative metrics, reflect the effects of any connection pass-through (or other load-shedding techniques employed by the system under test). [3] By setting the minimum hit ratio required to qualify a run at 25%, the benchmark rules out trivial solutions (taking the problem-solution conception of the benchmark). At the same time, the benchmark report rewards caches with higher hit ratios achieved by a cache. [4] The idea that "caches should cache (everything)" is anathematic. Caches (CPU, disk, etc.) have always varied in their hit ratio, response time, throughput, consistency levels, scalability, etc. It has always been an acceptable design practice to trade-off these characteristics against each other. E.g., sometimes it is better for a CPU cache to run at twice the clock rate, at the price of halving its size (and reducing its hit ratio). The benchmark shouldn't artificially exclude design dimensions, or, in the reporting of the results, pre-judge the applicability of different designs in different contexts. Solom... CTO, www.InfoLibria.com