Abstract not found

For many cluster network technologies, a considerable part of the communication latency experienced by applications stems from interrupt handling. To reduce network communication latency, many scienti c parallel applications and communication libraries for such applications continuously poll the network adapters when waiting for incoming communication. This approach is hard to use in a more general cluster computing setting, as such busy waiting wastes processor resources and is vulnerable to long delays caused by the execution of compute bound processes. We propose integrating polling with the operating system scheduler, as the scheduler has extensive knowledge of the states of the processes, and thus is in a good position to decide when to poll. We have implemented such a polling scheduler in the Linux operating system. Our results show a latency reduction by 56% for small messages when using a stream protocol for a Scalable Coherent Interface (SCI) cluster network. 1 Introduction ...

We propose a SAN architecture called Queue Pair IP (QPIP) that combines the interface from industry proposals for low overhead, high bandwidth networks , e.g. Infiniband, with the well established inter-network protocol suite. We evaluate how effectively the queue pair abstraction enables inter-network protocol offload. We develop a prototype QPIP system that implements basic queue pair operations over a subset of TCP, UDP and IPv6 protocols using a programmable network adapter,. We assess this prototype in terms of basic application performance, underlying processing costs, and a network storage application. With modest hardware support, QPIP can perform as well as traditional inter-network protocol implementations at a fraction of the host CPU overhead. With hardware support equivalent to Infiniband, QPIP would achieve similar performance targets. 1.