No abstract

With recent declines m the cost of semiconductor memory and the mcreasmg need for high performance I/O disk systems, It makes sense to consider the design of large caches In this paper, we consider the effect of caching writes We show that cache sizes m the range of a few percent allow writes to be performed at neghglble or no cost and mdependently of locahty conslderatlons

No abstract

With recent declines m the cost of semiconductor memory and the mcreasmg need for high performance I/O disk systems, It makes sense to consider the design of large caches In this paper, we consider the effect of caching writes We show that cache sizes m the range of a few percent allow writes to be performed at neghglble or no cost and mdependently of locahty conslderatlons

Traditional mirrored disk systems provide high reliability by multiplexing disks. Performance is improved with parallel reads and shorter read seeks. However, writes must be performed by both disks, limiting performance. Doubly distorted mirrors increase the number of physical writes per logical write from 2 to 3, but performs logical writes more efficiently. This reduces the cost of a random logical write to 1/3 of the cost of a read. Moreover, much of the write cost can be absorbed in the rotational latency of the reads, performing under certain conditions all the writes for free. Doubly distorted mirrors achieves a 135% performance improvement over traditional mirrors in the TP1 benchmark. Although these techniques require a disk cache for writes, the cache need not be safe nor is recovery time impacted very much.