An initial Operating System adaptation heuristic for Swap Cluster Max (SCM)

One of the goals of an Operating System (OS) is to efficiently manage system resources so that they are available to applications when they are needed. In general, the operating system\u27s resource management policies and the parameters that influence the behavior of these policies are set at OS compilation time and are fixed irrespective of the type of workload being executed. Since different workloads have different resource-usage patterns, such a generalized management policy suits the needs of some workloads more than others, thereby degrading performance for some workloads. To help alleviate this problem, the Linux operating system includes customization features that allow system administrators to select the management policies and parameters associated with some system resources in order to best serve their workloads. One drawback of this solution is that the administrators need to be knowledgeable about both the workloads\u27 resource-usage patterns as well as the inner workings of the operating system. An alternate solution would be to let the OS dynamically detect resource-usage patterns at run time and adapt its own policies and parameters accordingly, thereby eliminating the need for the user to participate in OS adaptation. This thesis is a first step towards a dynamic adaptation heuristic for one such parameter of the Linux Virtual Memory Manager (VMM), namely, Swap Cluster Max (SCM). Kandiraju [11] and Portillo [2] have shown that SCM is a good candidate for dynamic adaptation. This thesis extends Portillo\u27s [2] work and makes two main contributions to the dynamic adaptation of SCM. First, this thesis develops a workload generator that can simulate phases of a class of applications with simple memory-access patterns, which are characterized by their Page Fault Rates (PFRs). The workload generator also refines Portillo\u27s [2] experimental methodology by creating natural memory pressure without the need for a RAMDISK to artificially induce memory pressure. Second, using the workload generator, this thesis presents experimental results that are used to study the correlation of SCM and PFR. This analysis is a first step towards a dynamic adaptation heuristic for SCM