Research And Optimization Of Scalability In Distributed Shared Memory

The distributed shared memory(DSM)can provide a shared memory model for users on a cluster.It hides the communication details among machines and thus improves the portability and ease of programming.The main overhead of DSM is the cost of communication among different machines.With the recent evolvement of network interfaces,the network latency has been decreased to several microseconds and the throughput has been increased to several gigabits per second.This may probably mitigate the performance degradation of DSM systems introduced by the communication overhead.As a result,this paper will discuss the scalability issues of DSM.SNOP,as the DSM framework designed by this paper,has the characteristics of strong portability and modularized DSM protocol.As for the portability,it implemented the interfaces of memory allocation,thread management,synchronization and etc.The signatures of these interfaces are quite similar with the ones in the multiprocessor environment and this helps with the improvement of portability.SNOP leverages the features of memory layout and customized linker script to provide the transparency and convenience for the users.To support modularized DSM protocol,SNOP provides several easy-to-use network interfaces and a flexible message template.This paper will implement the protocol from IVY system firstly.This paper also gives three kinds of optimizations based on the SNOP framework.The first optimization designs a brand new DSM protocol,snooping-based DSM protocol,to improve the overall performance of the SNOP framework.The second kind of optimizations aims to improve the correctness and performance of the snooping-based DSM protocol.It discusses the correctness and liveness issues in this protocol.Then,this paper comes up with the sequence-based solution to solve them.To further improve the performance of this solution,it also leverages the features of current network interfaces.The last kind of optimizations aims to improve the performance of synchronization primitives.It first analyses the performance overhead of primitives based the DSM and figures out that this kind of implementations introduces multiple rounds of protocol executions.So this paper invents the synchronization manager to help improve the performance.Finally,this paper evaluates the SNOP framework from different aspects.This evaluation chooses two different memory access patterns and three kinds of applications.It analyses the synchronization performance,protocol performance and the scalability of the SNOP.For the synchronization primitives,it figures out that the spinlocks in the SNOP can improve the performance by 89 microseconds.For the comparisons of snooping-based protocol and IVY protocol,the results show that snooping-based protocol can improve the performance by 33%.At last,this evaluation also shows that snooping-based DSM protocol can improve the scalability of the SNOP framework.