Throughout this paper, we evaluate each successive design proposal after describing that proposal in order to isolate its performance impact and motivate further refinements. We present our experimental setup in this section to assist this progressive unfolding of our design.
We evaluate our design proposals on an 866MHz Pentium III configured to use 64MB of memory, running our modified Linux kernel. Most modern desktop computers have much more memory. We restricted the memory size deliberately to facilitate comparison with prior work [4,3] and compensate for established I/O benchmarks [5] which use data sets that have not been updated to reflect growth in data set sizes. Our storage system consists of seven Compaq RZ1CB Ultra SCSI disks (12ms average access time). The file system is striped across four of the disks with a 64KB stripe unit. The central cylinders of the other three disks are designated as swap space. The maximum transfer rate supported by the disks and the SCSI interface is 40MB/s.
All our results are averages over three runs; however, the variance in execution time across these runs was always small (within a few percent of the calculated mean). All file- and swap-backed pages were flushed from memory before each run.