I/O Workload Fingerprinting in the Genetic-Library
One great difficulty in writing an I/O scheduler is having one set of tunables which works well for every workload. If the I/O scheduler knew what kind of workload was occurring, it could modify its tunables for better performance. However, due to the I/O scheduler's depth in the kernel, it is very difficult to see this information. One method which can be used to obtain this information is to look at many small pieces of information, and then aggregate them to create a usable fingerprint.
This paper describes how to create an I/O workload fingerprint and its uses in both I/O schedulers, and in the genetic-library. The paper's main focus is on the application of the fingerprinting in the genetic library. By having a workload fingerprint, the genetic library can save genes which worked well for a particular workload, and reintroduce them back into the gene pool when that workload is seen again. This leads to faster convergence on an optimal tunable in an rapidly changing environment.