Endnotes

The numbers reported in this paper cannot be used to draw inferences about IBM's products, since (i) Shark (ESS Model 800) does not use any of the above three algorithms; (ii) Shark does not use the sequential prefetching algorithm described above; (iii) our software implementation on Shark hardware is experimental and academic; (iv) we use only one of the two available clusters on Shark; and (v) we have scaled throughput (IOPS) numbers. This paper is not intended to be an official SPC-1 submission, but is an academic study geared to demonstrate that SARC is better than LRU variants.