UW-Madison Logo

The ADvanced Systems Laboratory (ADSL)
Overview

:: Home
:: Projects
:: People
:: Pictures
:: Publications
:: Software
:: Sponsors
:: Internal

NVM Storage Technology

Unwritten Contracts: Characterizing Optane Performance

New NVM technologies provide unprecedented performance for persistent storage. Such devices offer significantly lower latency than Flash-based SSD and can be a costeffective alternative to DRAM. One excellent example is Intel's 3D XPoint memory from Intel and Micron, available on the open market under the brand name Optane. It is available in various form factors, including Optane memory (a caching layer between DRAM and block device), Optane SSD (a block device), and Optane DC Persistent Memory. Using new technology effectively requires understanding its performance and reliability characteristics; these rules, or unwritten contract, are already known for hard disks and flash-based SSDs.

We have begun forming the unwritten contract for the Optane SSD in terms of six rules.

  1. To obtain low latency, users should issue small requests ( > 4 KB) and keep a small number of outstanding IOs (Access with Low Request Scale rule).
  2. Dfferent from HDD/Flash SSDs, Optane SSD clients should not consider sequential workloads special (Random Access is OK rule)
  3. To avoid contention among requests, clients should not issue parallel accesses to a single chunk (4KB) (Avoid Crowded Accesses rule).
  4. To achieve optimal latency, the user needs to control the overall load of both reads and writes (Control Overall Load rule).
  5. To exploit the bandwidth of Optane SSD, clients should never issue requests less than 4KB (Avoid Tiny Accesses rule).
  6. To get the best latency, requests issued to Optane SSD should align to eight sectors (Issue 4KB Aligned Requests rule).
  7. When serving sustained workloads, there is no cost of garbage collection in Optane SSD (Forget Garbage Collection rule).
A few of the on-going research questions include:
  • How general is this unwritten contract for other emerging devices?
  • Previously, memory and storage were treated as a hierarchy, with the memory layer closer to the processor being smaller, faster (both higher throughput and lower latency), and more expensive than the layer beneath it. These hierarchical properties no longer hold across these layers with rich performance characteristcs. How can all of these different storage technologies be best used together? How should caching and partitioning of data be performed given non-hierarchical performance?

Towards an Unwritten Contract of Intel Optane SSD
Kan Wu, Andrea C. Arpaci-Dusseau, Remzi H. Arpaci-Dusseau
HotStorage '19
Renton, WA, July 2019.
Available as: PDF