UC Santa Cruz

The use of non-volatile write caches is an effective technique to bridge the performance gap between I/O systems and processor speed. Using such caches provides two benefits: some writes will be avoided because dirty blocks will be overwritten in the cache, and physically contiguous dirty blocks can be grouped into a single I/O operation. We present a new block replacement policy that efficiently expels only blocks which are not likely to be accessed again and coalesces writes to disk. In a series of trace-based simulation experiments, we show that a modestly sized cache managed with our replacement policy can reduce the number of writes to disk by 􏰨􏰧 percent and often did better. We also show that our new policy is more effective than block replacement policies that take advantage of either spatial locality or temporal locality, but not both.