The Implementation And Optimization Of NHD System For Service On Demand

With the development of information technology, the computers have been extensivelyused in many different fields, such as enterprise computing, multimedia teaching, etc.Currently, most of these applications adopt the dispersed computing model in which eachcomputer has its own local attached storage. To some extent, this computing model meets therequirements of scalability and openness, but it has obvious inherent limitations, such as lackof dynamics, low availability and complicated management.In an attempt to solve these problems, the National Research Center for High PerformanceComputer (NRCHPC) has developed a specialized system called SOND(Service On Demand),which is based on the theory model of physical separation between storing and computingresources. This system utilizes a special hardware equipment called NHD (Network HardDisk) and installed on the client side to dynamically establish the network connection betweenthe network-attached storage and the local computing system, which functions as a localphysical disk. This newly reconstructed computing environment has advantages of dynamics,individuation and easy management.As a key component of the SOND, the NHD system is implemented as a specialembedded device running customized software, which is able to provide a data access channelbetween local computing resource and remote storing resource. This paper first makes anin-depth description of the underlying design principles of NHD system, as well as itsfunctions and the involved data channel. Based on the careful analysis of the transportationlatency of the data channel, we propose I/O performance optimization solutions for the NHDsystem including various optimizing technologies such as DMA streaming, zero-copy, TCPchecksum offloading etc. The experiment results show that after the optimization, the I/Operformance of NHD system has enhanced over three times than before.