Intel® Optane™ Persistent Memory: Comprehensive Performance Evaluation and Configuration Guide

Intel® Optane™ Persistent Memory (PMem) is highlighted as a disruptive technology that combines high capacity, low latency, and data persistence, promising to reshape traditional memory and storage architectures.

Basic performance test: Comparing a configuration of 8 DRAM + 8 PMem with 8 DRAM + 1 PMem shows that the 8 + 8 setup delivers roughly eight times the throughput of the 8 + 1 setup, while latency remains in the 170‑320 ns range—comparable to DRAM and far faster than the microsecond‑level latency of NVMe SSDs.

Redis workload: In a test of 120 million requests, PMem in KEME DAX mode achieves about 96 % of DRAM read performance, 90 % of DRAM write performance, and latency nearly identical to DRAM, demonstrating that a modest DRAM + PMem server can match the performance of a DRAM‑only server at lower cost.

RocksDB workload: PMem delivers random‑write performance close to twice that of an NVMe SSD and random‑read performance up to ten times faster than NVMe, confirming its advantage for high‑throughput database scenarios.

Configuration modes: PMem supports three primary modes—Memory Mode, App Direct, and KMEM DAX—each suited to different application requirements such as non‑persistent memory, persistent high‑bandwidth storage, or flexible DRAM/PMem placement.

Configuration steps:

Memory Mode:

ipmctl create -f -goal memorymode=100

# Reboot the system

App Direct (AD) Mode:

ipmctl create -f -goal persistentmemorytype=appdirect

# Reboot the system

ndctl create-namespace

mkfs.ext4 /dev/pmem0

mount -o dax /dev/pmem0 /mnt/pmem0

KMEM DAX Mode:

# Set AD mode first and reboot

# Create DEVDAX namespace

# Add PMem as a new NUMA node

In conclusion, the extensive testing and straightforward configuration guide demonstrate that PMem can provide DRAM‑like performance with SSD‑level persistence, offering a cost‑effective solution for memory‑intensive workloads, large‑scale databases, and private‑cloud virtualization.