In this paper we investigate disk activity caused by accesses to large databases stored on modern storage systems. We show that read and write requests do not arrive according to a Poisson process but rather show a power law behaviour. Also, the aggregated rates of requests show signs of long range dependence through convincing 1/f noise in the power spectrum of the process. We then present a naive simulation model of read and write request performance that is parameterised by this empirical data, and compare observed and predicted response times. Finally we speculate on the origins of the observed behaviour and discuss its consequences for large systems.
pubs.doc.ic.ac.uk: built & maintained by Ashok Argent-Katwala.