Crazy memory over-commits without compromising response time.

This proposal has been rejected.


One Line Summary

A recipe to achieve crazy memory over-commits without degrading the response time.


The total memory used at any given point in a system is smaller than the total memory requested. Cloud providers leverage this fact to overcommit memory.
However there is a constant concern that a sudden spike in total usage can lead to huge swap latencies leading to broken SLAs. NVMe devices, if used as a swap device, significantly reduce swap latencies and drive higher overcommits. But on larger systems with more than 256GB memory, higher overcommits can easily saturate the NVMe bandwith leading to sluggish application response time.

We have developed a device mapper module that alleviates saturation of NVMe bandwidth by transparently compressing and decompressing the swap data. Its optimized to work with minimal memory allocation to avoid exacerbating the swap traffic in an already memory constrained environment. Further to avoid overloading the CPU with compression/decompression, it off-loads the work to offload engines if one is available.

We propose to talk about our setup, the design of the device mapper compressor and most importantly our exciting results.


nvme, compression, overcommit


  • Ram Pai



    Software Engineer at IBM’s Linux Technology Center.

Leave a private comment to organizers about this proposal