Challenges to use swap on fast NVMe disk

This proposal has been rejected.

*

One Line Summary

Discuss the possible solution for the challenges we found during our testing and optimizing to use swap on fast NVMe disk.

Abstract

Swap is an important Linux kernel mechanism to run more or bigger applications on a system that hasn’t enough RAM. With the improvement of the fast NVMe disk performance, swap is getting more attractive again. But the kernel swap subsystem becomes the bottleneck now. We have resolved some problems, for example, improving the scalability of the swap cache and the swap device management, etc. But there are still some challenges remaining, including but not limited to, scalability of swap cgroup, massive TLB shoot down for multi-thread workload, swap allocation/freeing speed, THP swap optimization, swap readahead optimization, etc. In this presentation, we will give a short introduction to Linux kernel swap subsystem and swap optimization done recently, show the challenges we found with the testing results, and discuss the potential resolution with the kernel developers.

Tags

memory management, swap, performance and scalability

Speaker

  • Biography

    Ying Huang is a senior software engineer in Intel’s Open Source Technology Center (OTC). At Intel over 10 years, he’s worked on Linux kernel, including x86 boot, kexec, RAS, etc. And now, he is working on 0day Linux kernel performance testing and memory management subsystem performance optimization. He got his doctor degree in Zhejiang University.