Tiramisu: Fast and Scalable Traffic Splitting on Commodity Switches

Abstract

Traffic splitting is a key task for load balancing over multiple servers or paths. Load balancers often rely on commodity switches to implement traffic splitting within the ternary content addressable memory (TCAM). It is critical to reduce the amount of TCAMs allocated for this task since TCAMs are power-hungry and have limited capacities. Previous traffic splitting schemes have concentrated on minimizing the number of TCAM entries for a single service. However, these schemes suffer a great issue of scalability since a commodity switch with small-sized TCAMs needs to handle hundreds or thousands of services simultaneously. In this paper, we propose Tiramisu, an overlay approach to fast and scalable traffic splitting for multiple services on commodity switches. Tiramisu exploits wide SRAM words available in TCAMs to overlay multiple groups of rules within one same TCAM table. With rule overlaying, Tiramisu reduces the total number of TCAM entries allocated for multiple services. Experimental results on software simulations and P4 switches show that Tiramisu achieves significant reductions in number of TCAM entries by up to one order of magnitude for the imbalance error of 1 % as well as achieves higher splitting throughput and faster rule updates compared to previous schemes.