Thread Scheduling Strategy Research Of Multiprocessor Systems

In a SMP system which is applied with Chip Multi-Processor (CMP),Simultaneous Multithreading (SMT) technology, unreasonable shared mechanism maycause uncontrollable use of simultaneous multithreading for public resources, whichwill reduce the system throughput and resources utilization. Therefore, it has alwaysbeen the special emphasis of operating system research to design and implement aresource utilization considered SMP scheduling mechanism.This thesis firstly introduces the evolution and characteristics of themulti-processor and multi-core architecture technology. To review and summarize thecorresponding framework thread scheduling strategy is the next step. Thirdly, this papersummarizes the development of the Linux scheduler and the corresponding versionscheduler characteristics, and analyzes CFS scheduler framework of the Linux V3.0times and loading balance strategy which is based on the domain, and points out thedeficiency of the Linux scheduler.Current linux scheduler just balances the load weightbetween CPUs based on the priority values and numbers of the processes withoutconsidering the utilization of system bus bandwidth, which is likely to cause the declineof the utilization ratio of bus bandiwidth.Specifically speaking, current loading balanceoperation may aggravates the tension of bus bandwidth use for migrating a process withexcess bandwidth needs;or idles the bandwidth resources for not being able to migrate aprocess with morderate bus bandwidth need from the busiest CPU in time.Finally, this thesis proposes the optimization scheduling strategy and the improvingmethods based on current Linux scheduler. Optimization design is to on-line analyze theperformance counts through sampling the running threads, and assesse their busbandwidth usage during the recent sampling window, which provides direct basis fordecision-making. By using processor’s built-in performance counters, the number ofexecuted instructions and hierarchy-cache misses are available to calculates the threads’average bandwidth usage within the sampling window. Due to the dicision-making isbased on history data analysis, the period of sampling and the size of window are ofvital importance. On one hand, if the size of sampling window is too small, the recent average usage of bandwidth cannot be assessed presicisely;on the other hand, bigwindow size may leads to miss the proper timing of avoiding the fierce contention forSMP system bus bandwidth.In the concrete implementation, the appropriate value ofsampling window is determined by repeated tests. To optimize the original SMP loadingbalance algorithm, both the CPU/Cache affinity of and bus bandwidth usage should beincluded for selecting the migration process According to the results of three tests basedon STREAM Benchmark, the optimization improves the utilization of SMP system busbandwidth without compromising the effect produced by original algorithm.