热烈欢迎各位领导莅临指导.pptVIP

This figure shows the hit ratio of IOzone and TPC-C in combined and running-alone case respectively. The hit ratio of IOzone in running-alone case is about 11.9% and in the combined case is about 11.6%, which nearly has no degradation on hit ratio. In contrast, for TPC-C, the degradation on hit ratio is about 11% (from 31.1% to 20.1%), which is almost 37 times larger than IOzone. * Next, we will talk about our solution. * The below figure illustrates our approach’s architecture. Our approach is partition-based. We explicitly divide a cache into several partitions, one for each application. The scheduler is composed with three modules: utility predictor, candidate finder, and segment dispatcher. After receiving an allocation request, the scheduler will find the victim partition for the demand access. For each candidate partition, the scheduler invokes the utility predictor to estimate its change of performance and fairness by shrinking cache blocks. Then, the candidate finder iterates all partitions to find the appropriate partition. At last, the segment dispatcher removes the tail segment of the selected partition to the requested partition. * In our work, we use the maximum slowdown metric to evaluate fairness. The slowdown is defined as the ratio between the performance it achieves when running alone and running combined, which is also adopted by many other researches. In our work, the utility on performance refers to the change in overall performance by adding or shrinking a single cache block. The delta P(n) refers to the overall performance when the cache size is n. We classify the change of performance into two type. Delta P positive, means positive performance improvement by adding a cache block. Delta P negative, means performance decrease by removing a cache block. Both of these two terms are estimated based on the history of access requests. * When an access block missed in the cache and the cache does not have any free cache blocks left, an cach