The proliferation of many-core architectures has led to the explosive development of parallel applications using programming models, such as OpenMP, TBB, and Cilk/Cilk++. With increasing number of cores, however, it becomes even harder to efficiently schedule parallel applications on these resources since current many-core runtime systems still lack effective mechanisms to support collaborative scheduling of these applications. In this paper, we study feedback-driven adaptive scheduling based on work stealing, which provides an efficient solution for concurrently executing a set of applications on many-core systems. To dynamically estimate the number of cores desired by each application, a stable feedback-driven adaptive algorithm, called SAWS, is proposed using active workers and the length of active deques, which well captures the runtime characteristics of the applications. Furthermore, a prototype system is built by extending the Cilk runtime system, and the experimental results, which are obtained on a Sun Fire server, show that SAWS has more advantages for scheduling concurrent parallel applications. Specifically, compared with existing algorithms A-Steal and WS-EQUI, SAWS improves the performances by up to 12.43% and 21.32% with respect to mean response time respectively, and 25.78% and 46.98% with respect to processor utilization, respectively.

Global computing system (GCS) harnesses the idle CPU resources of clients connected to Internet for solving large problems that require high volume of computing power. Since GCS scale to millions of clients, many projects usually adopt coarse-grained scheduling in order to reduce server-side contention at the expense of sacrificing the degree of parallelism and wasting CPU resources. In this paper, we propose a new type of client, i.e., a scheduling proxy that enables adaptive-grained scheduling between the server and clients. While the server allocates coarse-grained work units to scheduling proxies alone, clients download fine-grained work units from a relatively nearby scheduling proxy not from the distant server. By computation of small work units at client side, the turnaround time of work unit can be reduced and the waste of CPU time by timeout can be minimized without increasing the performance cost of contention at the server. In addition, in order not to lose results in the failure of scheduling proxies, we suggest a technique of result caching in clients.