Distributed-thread scheduling methods for reducing page-thrashing

Abstract

Although distributed threads on distributed shared memory (DSM) provide an easy programming model for distributed computer systems, it is not easy to build a high performance system with them, because a software DSM system is prone to page-thrashing. One way to reduce page-thrashing is to utilize thread migration, which leads to changes in page access patterns on DSM. In this papel; we propose thread scheduling methods based upon page access information and discuss an analytical model for evaluating this information. Then, we describe our implementation of distributed threads, PARSEC (Parallel sofhyare environment for workstation cluster}. Using user-level threads, PARSEC implements thread migration and thread scheduling based upon the page access i n f o mtion. We also measure the peqormance of some applications with these thread scheduling methods. These measurements indicate that the thread scheduling methods greatly reduce page-thrashing and improve total system performance.

Thread migration and communication minimization in DSM systems

“…Two threads that frequently access the same shared pages can be pr esumed to share data. We define a density function as the access rate of thread i to page p. The correlation of two threads over page p can be computed as the product of the density function of the two threads for page p. The overall correlation of the two threads, then, is the sum of the correlations for each page in the system [18]. Unfortunately, page-based DSMs have no efficient way of deriving density functions because they can not track ind ividual accesses.…”

“…Thread migration has also been studied in the Millipede [22] and PARSEC [18] DSMs. Both systems implement thread migration in the context of sequential consistency rather than a relaxed consistency model.…”

Thread Migration and Communication Minimization in DSM Systems

“…Two threads that frequently access the same shared pages can be presumed to share data. We define a density function as the access rate of thread i to page p. The correlation of two threads over page p can be computed as the product of the density function of the two threads for page p. The overall correlation of the two threads, then, is the sum of the correlations for each page in the system [18]. Unfortunately, page-based DSMs have no efficient way of deriving density functions because they can not track individual accesses.…”

Adaptive sampling-based profiling techniques for optimizing the distributed JVM runtime

“…Such a profile is however difficult to obtain without high overheads. Passive correlation tracking (used in [10,11]) that relies on remote page faults to activate access logging can only capture partial sharing behavior because access to a validated page by other local threads is missed logging. Active correlation tracking [12] was proposed to track the sharing information.…”