期刊名称：International Journal of Computer Science and Network
印刷版ISSN：2277-5420
出版年度：2012
卷号：1
期号：3
出版社：IJCSN publisher
摘要：Chip multiprocessors (CMPs) combine multiple processorson a single die, however, the increasing number of processorcores on a single chip increases the demand of two criticalresources, the shared L2 cache capacity and the off-chip pinbandwidth. Demand of critical resources are satisfied by thetechnique of cache compression. It is well known thatCompression technique, which can both reduce cache missratio by increasing the effective shared cache capacity, andimprove the off-chip bandwidth by transferring data incompressed form. However, decompressing cache lines alsoincreases cache access latency, potentially degradingperformance.to minimized performance degradation occursdue to decompression in this paper, we develop an dynamiccompression policy which dynamically adapts the cost andbenefit of cache compression, and block will be compressedonly when compression is beneficial, otherwise block will beuncompressed.We use two-level cache hierarchywhere the Level1cache holds uncompressed data and the Level2cachedynamically selects between compressed anduncompressedstorage.The effective size of an Level2 cache can beincreased by using a LZW dictionary-based compressionscheme. Variable-sized compressed blocks tend to increasethe design complexity of the compressed cache architecture.This paper suggests technique to reduce the decompressionoverhead and to manage the variable sized compressedblocks efficiently.We evaluate dynamic cache compressionusingsimulation and different workloads. Weshow thatcompression can improve performance formemory-intensivecommercial workloads. However, always using compressionhurts performancefor low-miss-rate workloads due tounnecessarydecompression overhead. By dynamicallymonitoring workload behavior,the dynamic compressionpolicy achieves comparable benefitsfrom compression, whilenever degrading performance for memory intensiveworkloads. Here also we compare performance of dynamiccache compression under Least Recently Used (LRU)replacement policy as well as Efficient Replacement Policy(ERP). Performance of dynamic compression scheme isobserved under both the replacement policies and it is findout that for most of the cases efficient replacement policywork well in dynamic cache compression technique
