Related Research On Thread Lightweighting Based On SMP

In the course of parallel computing about 50 years, the potential of computing is greatly developed from ISA parallel inside processor to parallel in cluster on the basis of existing computing capability. As thread applications are being popular, higher parallel degree and finer parallel granularity are developed on system architecture, compile technology, programming model, and thread library, etc. Meanwhile, for massively concurrency fine-grained applications like network information service, it is hard to get speedup on SMP by traditional thread. And with the progress of CMP and SMT, optimization of thread overhead is brought into schedule especially.For this reason, to improve effective utilization of resources under fine-grained parallel, suitable granularity superior to traditional Posix thread and the corresponding high-efficient low-overhead scheduling technology are discussed in the thesis. The main work is as follows: Based on detail analysis and overhead test of pthread, the relationship between synchronization granularity and speedup on SMP under pthread microsecond level switch overhead is got by quantitative analysis with experiments. For synchronization granularity equivalent to tenfold level of thread switch overhead, the execution flow is hard to obtain effective performance speedup on SMP. The problem is pointed out that fine-grained parallel performance is limited by thread overhead itself.Aiming at thread overhead sensitive problem for fine-grained parallel, the principle on resource self closure and active scheduling is put forward, and high-efficient low-overhead scheduling module independent with operating system is designed and developed. To the scheduling module, it is of advantages on both real parallelism for kernel thread and low overhead for user thread. It can do scheduling according to application characteristics in user level, implement high efficient switch to frequency as ten thousands per second, and reduce switch overhead effectively. Thus, it conduces to lightweight on thread granularity and switch overhead tentatively, and overcome the problem that fine-grained parallel applications are unable to effectively obtain speedup on SMP by standard thread. Besides, the scheduling module implements SMP effective use without modifying operating system kernel, and can be widely used in fine-grained parallel applications with little dependency.By construction of resource self closure and the scheduling module further, the real performance of parallel simulator SandFox is improved and several times better than that on...