(A) low-cost dynamic DRAM partitioning policy in tiered-memory systems

Abstract

Persistent memory (PM) is rising as an excellent candidate for the next-generation memory resource. A tiered-memory system refers to a system where persistent memory is used as the second-level memory resource after DRAM. In this work, we investigate the memory usage characteristics in an environment where DRAM and PM are used simultaneously by multiple processes. We find that, in a tiered memory system, appropriate DRAM partitioning can yield higher system performance. Based on key observations of applications' memory usage patterns and the unique characteristics of tiered-memory systems, we design a policy for the operating system to dynamically allocate DRAM resource among processes. Our policy is reactive and efficient

it does not require any prior knowledge of applications and uses negligible system resources. We implement and evaluate our policy in the Linux kernel using its native features. The experimental results show that the new partitioning policy well understands, captures, and exploits how applications use memory in the target system to achieve higher system performance.