Accelerated Socket Communications in System Area Networks

In recent years, a great need for large scale servers has arisen. One practical solution is to construct such servers as cluster systems. For applications reside on them, inter process communications(IPCs) among them must be high performance. The current generation of System Area Networks(SANs) have very wide bandwidths, and very low communication latencies, and they usually consume relatively little CPU time. Although such hardwares have quite good performance, we cannot utilize them fully with socket APIs that are used as a de facto standard IPC. We can easily see the performance problem of socket APIs by comparing the results of benchmark test of socket APIs and new APIs such as VIPL[1, 2] over the same SAN hardware. For example, the result of our measurment of bandwidth says that we can get twice as good performance as socket APIs. In case of communication latencies and CPU utilization, we can get same kind of results. This means that these performance problems come from the overheads of underlying network protocol stacks. Indeed there are new APIs that are well optimized for SAN environment, none of them could not have replaced socket APIs. So, if we want to widely deploy an improved IPC over SANs, it is essential to retain compatibility with socket APIs. Socket APIs are used to abstract the endpoints of communications via underlying network protocols as files. To be compatible with socket APIs, it is important to retain compatibility with the UNIX semantics for file descriptors without mentioning the semantics of the underlying network protocols.