Optimal Polling for Latency-Throughput Tradeoffs in Queue-Based Network Interfaces for Clusters

Abstract

Abstract. We consider a networking subsystem for message-passing clusters that uses two unidirectional queues for data transfers between the network interface card (NIC) and the lower protocol layers, with polling as the primary mechanism for reading data off these queues. We suggest that for accurate mathematical analysis of such an organization, the values of the system's states probabilities have to be taken into consideration, in addition to the well-known mean-value estimates. A single server single queue polling system with server's "vacations" is then articulated via an M/G/1 queueing model. We present a method for obtaining the values of system's states probabilities in such a scheme that can be used to identify "sweet spots" of the polling period that result in a reasonable compromise between the message passing latency and the overall system's throughput.

Analytical Model and Performance Analysis of a Network Interface Card

“…The Myrinet NIC and software system in [24] is modeled at a meso-granularity level (i.e., in between the macro-granularity level in [23] and the micro-granularity level we have considered here) using a system of 2 queues such that the processor polls between queues at the software and the hardware layers, thereby incurring time during the context switch. However, since the entire NIC is modeled as a single-server queue in [24], it is not clear how various designs inside the NIC can be evaluated. The motivation for considering a micro-granularity level for modeling the Myrinet NIC in this paper is to test various NIC designs and evaluate the performance improvement across the NIC.…”