Development of Methods for Reducing the Spins of Guest Multiprocessors

概要

論文の詳細を見る
Methods are presented for redusing the spins of operating systems (OSs) that are running in guest multiprocessors, which are virtual machines (VMs) that contain multiple logical processors sharing a main memory area. VMs are functional copies of a real host computer. Different OSs can be run in different VMs concurrently. A hypervisor is a program that allocates real resources to, controls, and schedules multiple VMs. Three methods are presented by which a hypervisor can inform a guest OS, which means an OS in a VM, of the processor allocation forms, which mean dedicated real processors or shared real processors by VMs. According to the information, the guest OS determines whether or not it should spin. Only the hypervisor knows the processor allocation forms, because it allocates real processors to guest OSs. Three other methods for reducing the spin of guest OSs are presented. The guest OSs call the hypervisor when they are going to spin, and the hypervisor schedules partner logical processors, which belong to the same VM. before scheduling the spinning logical processor again. These methods are different in the timing at which the hypervisor reschedules the spinning logical processor. The first, called WAPD, sets a spinning logical processor in a wait state until all its partners that are ready have been scheduled. The second, called WOPD, sets a spinning logical processor in a wait state until one of its partners that are ready has been scheduled. The third, called RSLP, requeues the spinning logical processor at the last position in the ready queue. These three methods have been experimentally implemented, and their effects have been measured and estimated quantitatively. For small workloads, the three methods have comparable performance. According to experiments, WOPD cannot suppress an excessive spin of a guest OS for large workloads that heavy spin activities. WAPD has a slightly (0.02-0.27%) larger overhead than the RSLP, and its performance is comparable to that of RSLP. For large workloads, WAPD has a slightly (3-4%) better performance than RSLP.
一般社団法人情報処理学会の論文
1995-03-15

Development of Methods for Reducing the Spins of Guest Multiprocessors

スポンサーリンク

概要

著者

関連論文

スポンサーリンク