Abstract

$Id: wlfu.tex,v 1.28 1999/01/22 23:33:22 tpkelly Exp $ Disk space in shared Web caches can be diverted to serve some sys-tem users at the expense of others. Cache hits reduce server loads, and if servers desire load reduction to dierent degrees, a replacement policy which prioritizes cache space across servers can provide dieren-tial quality-of-service (QoS). We present a simple generalization of least-frequently-used (LFU) replacement that is sensitive to varying levels of server valuation for cache hits. Through trace-driven simulation we show that under a particular assumption about server valuations our algorithm delivers a reasonable QoS relationship: higher byte hit rates for servers that value hits more. We furthermore adopt the economic perspective that value received by system users is a more appropriate performance metric than hit rate or byte hit rate, and demonstrate that our algorithm delivers higher \social welfare " (aggregate value to servers) than LRU or LFU.