Studies On Shared-Memory Management And Optimization Technologies In Parallel And Distributed Operating Systems

In this thesis, we study the technologies of shared-memory management and itsoptimization in parallel and distribated operating systems. Our primary work focuses on thedesign and implementations of Shared-Memory Abstract Layef, Thread Consistency Modeland Instananeous Information-Based Dynamic Page Migraion PolicyFirst of all, according to the characteristics of memory sharing in different architectures,we bring forward the concePt of Shaxed-Memory Abstract Layer. Wth Shared-MemoryAbstract Layer, we are able to unite several memory sharing cases including mie sharing inSMP, hardware-supported distributed sharing in CC-N'UMA and virtUal sharing in clusters.Based on this, we discuss and design the logical framwork of paralel and distributedoperating systems.One of the key problems in shared-memory management is how tO maintain memeoryconsistency between multiprocessors. Most of current memory consistency models areconfronted with the deficiency of scalability and difficult to imPlement in operating systems.Aiming at this, we propose S3C Framework, which is a new framework of memory.. ~consistency models and their imPlemeniations founded on synchronization points. In S3CFramework, we define three synchronization poittts and one consistency maintaning pointwhich are used together to describe the correct event order in differeni models. Via S3CFramework, we can dive into and comPare different memory consistency models and theirimplemefltations so as to provide insmictions on designing and implementing new models.Wth the help of S3C Framework, we propose a new operating system-celltric memoryconsistency model, Thread Consistency Model. In this model, memory access event order isobserved and restricted by the transformation of threads' states ddring the execution ofprograms. It is relatively aPpropriate to implement this model in opefating systms sincememory and processes/thIeads management could be combined organically At the same time,Thread Consistency Model is able tO utilize the advanage of multithreaded architecture tooverlaP conunwhcation with compllting so as to hide communication latency. As a result, thismodel is especially suitable for multithreaded-kernel operating systems and muItithreadedCPUs. Furthermore, through the relaxation of conditions of Thread Consistency Model, wepropose some lazy implementations, trying to reduce the overhead in maintaining consistencyso as to imProve system perfOrmance.Aiming at optimization technologies of shared-memory management, we investigate ondata locality optimianion policies. Since page migraion is an efficiellt way to improvelocality we stress on it. Current rePresentaive policies are in need of special hardware torecord history page access infonnation, resu1ted in their poor generality. To overcome thisdeficiency, we propose an Instantaneous InfOrmation-based Dynarnic Page Migration Policy,which doesn't need any support of hardware. In this policy, we use some instantaneous$Iv Ninformation directly obtained from caches to substitute traditional history information. The results of our experiment show that this policy could approach the effect of traditional policies.