A novel hybrid SRAM/DRAM memory architecture for fast packet buffers

Abstract

This paper addresses the design of fast packet buffers for high speed Internet routers and switches. These buffers usually use a memory hierarchy that consist of expensive but fast SRAM and cheap but slow DRAM to meet both speed and capacity requirements. One challenge building these packet buffers is to provide worst-case bandwidth guarantees and fixed latencies, not to stall pipelines or to reduce throughput. My colleagues and I propose a novel packet buffer architecture along with a new memory management algorithm which reduces the amount of required SRAM compared to other architectures, e. g. by 73% for a 100 Gbps system using DDR3-DRAM. Furthermore, our architecture scales well with line rate.