In this study, we evaluate the effect of Inter-Core Time Aggregation Scheduler (ITAS) on a database server workload. ITAS is a kernel-level thread scheduler to utilize the locality of references between threads on Chip Multi-Processor (CMP) platforms. To detect the locality of reference between threads without the overhead of sampling the behavior of thread information, we focus on sibling threads, which are kernel-level threads sharing the same memory address space. We consider that we can utilize computing resources such as cache by dynamically aggregating sibling threads. We have investigated the effect of Time Aggregation Scheduler (TAS) on a single processing Core. TAS utilizes the locality of references and reduces the overhead due to context switching by executing sibling threads collectively in a group. We extend TAS into ITAS, which aggregates sibling threads on different Cores at the same time so that we can expect the effect of aggregation on CMPs. As the number of Cores increases, it is likely to run multiple multithreaded programs simultaneously to utilize all Cores on a chip. Thus, we consider that ITAS will be applicable to many situations. In this study, we show the effect of ITAS on a realistic workload of a multi-threaded database server running with a simple application server on a commodity CMP platform in terms of throughput, the number of cache misses and resource stalls and influence on the target program and other non-aggregated background programs. The experimental result indicates that ITAS enhances the performance of the database server without degrading that of non-aggregated background programs.
展开▼
