Near-Memory Acceleration of Software-Defined Far Memory
Presented by Neel Patel
The stagnation of DRAM capacity scaling has led to the deployment of a new memory tier between DRAM and storage, which is known as far memory. Far memory tends to have a larger capacity than local memory with higher access latency and lower bandwidth. In this presentation, we first perform a cost-benefit analysis for two implementations of far memory: Disaggregated Far Memory (DFM), which adds extra DRAM modules over a system bus to increase the memory capacity, and Software-defined Far Memory (SFM) that compresses part of the existing DRAM capacity to increase the effective memory capacity. Our results show that it takes several years for a DFM to break even with the cost of an SFM for workloads with a moderate compression ratio. Then we discuss the opportunities in accelerating SFM and present an early-stage, near-memory accelerated SFM solution. Our preliminary evaluations show that performing SFM-related compression and decompression operations inside the memory provides performance isolation for co-running applications, significantly reduces the server’s energy consumption, and enables the memory management unit to better utilize the SFM space.
Bio: Neel Patel is a first-year graduate student in the Architecture Research Lab at the University of Kansas, advised by Professor Mohammad Alian. His research interest is to address data movement bottlenecks of scale-out systems by exploiting the underlying characteristics of memory and system interface technologies. Neel’s current research is on the near-memory acceleration of the datacenter tax.