Power-aware Computing with Optane Persistent Memory Modules

Abstract

The increasing popularity and interest in heterogeneous memory solutions has led to the emergence of new high-performance computing (HPC) memory technologies requiring in-depth analysis for efficient performance. On the other hand, energy consumption in modern HPC clusters has reached prohibitive constraints posing a major challenge to power-efficient computation. This paper explores the performance characteristics of Intel’s Optane Data Center Persistent Memory modules (PMM) and investigates power consumption optimization strategies. We evaluate the bandwidth and latency tradeoffs between DRAM and PMM memories and assess the power-saving potential with two approaches: power capping and frequency scaling using Intel’s SST-CP technology. Two different generation systems equipped with Optane PMMs were used: Intel’s Cascade Lake and Ice Lake CPUs. Our results indicate that up to 22% of power consumption can be saved at 4% increased time with memory-bound kernels accessing PMM memory.

Authors 4

1. Anara Kozhokanova Aachen IT centre

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,IT Center,Aachen,Germany

   IT Center, RWTH Aachen University, Aachen, Germany

2. Bo Wang Aachen IT centre

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,IT Center,Aachen,Germany

   IT Center, RWTH Aachen University, Aachen, Germany

3. Christian Terboven Aachen IT centre

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,IT Center,Aachen,Germany

   IT Center, RWTH Aachen University, Aachen, Germany

4. Matthias Mueller Aachen IT centre

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,IT Center,Aachen,Germany

   IT Center, RWTH Aachen University, Aachen, Germany