LLM: A low latency messaging infrastructure for linux clusters

Abstract

In this paper, we develop a messaging infrastructure, called LLM, to arrive at a robust and efficient low latency message passing infrastructure for kernel-to-kernel communication. The main focus is to overcome the high latencies associated with the conventional communication protocol stack management of TCP/IP. The LLM provides a transport protocol that offers high reliability at the fragment level keeping the acknowledgment overhead low given the high reliability levels of the LAN. The system utilizes some of the architectural facilities provided by the Linux kernel specially designed for optimization in the respective areas. Reliability against fragment losses is ensured by using a low overhead negative acknowledgment scheme. The implementation is in the form of loadable modules extending the Linux OS. In a typical implementation on a cluster of two nodes, each of uniprocessor Intel Pentium 400 MHz on a 10/100 Mbps LAN achieved an average round trip latency of 169ms as compared to the.531ms obtained by ICMP (Ping) protocol. A relative comparison of LLM with others is also provided.