Rethinking Design Metrics for Datacenter DRAM

Abstract

Over the years, the evolution of DRAM has provided a little improvement in access latencies, but has been optimized to deliver greater peak bandwidths from the devices. The combined bandwidth in a contemporary multi-socket server system runs into hundreds of GB/s. However datacenter scale applications running on server platforms care largely about having access to a large pool of low-latency main memory (DRAM), and in the best case, are unable to utilize even a small fraction of the total memory bandwidth. In this extended abstract, we use measured data from the state-of-the-art servers running memory intensive datacenter workloads like Memcached to argue for main memory design to steer away from optimizing traditional metrics for DRAM design like peak bandwidth so as to be able to cater the growing needs to the datacenter server industry for high density, low latency memory with moderate bandwidth requirements.