High Performance Networking For Home-based Software DSMs

It is widely accepted that communication overhead is the main obstacle that prevents thesoftware DSMs becoming the desirable vehicle for parallel computing. Several techniqueswere proposed to reduce the communication overhead in the past, such as using multiple writeprotocol to reduce the frequency of communication, using write vector technique to reducecommunication traffic, using data prefectching scheme to tolerate communication latency, etc..However, as consitency protocols have matured, there is little space for reducing thefrequency and traffic of communications dramatically. The goal of this dissertation is toinvestigate efficient communication system to enhance the performance of home-basedsoftware DSMs further.In this dissertation, three communication characteristics of home-based software DSMsare discussed using JIAJIA as a representative example of this type of system. Thecharacteristics include: (1). Both communication latency and bandwidth are important tosoftware DSMs, (2). The communication model of"request-reply"not only introducesasynchronous processing overhead, but also isn't suitable for remote page fetching which iswidely existed in software DSMs, (3). Concurrent communication and collision are two majorreasons for low efficiency in application based on software DSMs. According to thosecharacteristics, a new communication system called FRAMP is designed and implemented.FRAMP is composed of a full customized switch, four network interface cards, and acommunication protocol. The switch adopts crossbar-based structure and circuit switchingtechnology. The two communication models, message passing and remote direct memoryaccess, are supported by FRAMP network interface card. Those features make FRAMPbecome a high bandwidth and low latency communication system on cluster.Raw performance evaluation shows that 5.7μs one way latency and 101MB/s asymptoticbandwidth are delivered in FRAMP. The performance improvement of applications built onJIAJIA is also evaluated and a dramatic enhancement is obtained by the advantages ofFRAMP.The work in this dissertation shows that, comparing with the traditionalcommunication system, specific cluster networking customized for software DSMs canmostly improve end-user performance. The shared memory platform with high performanceand easy programming for parallel computing is promising on cluster architecture.