Docker Performance With KPTI Page Table Isolation Patches

Written by Michael Larabel in Operating Systems on 5 January 2018. Page 2 of 3. 18 Comments

First up was the Himeno pressure solver scientific test... No measurable change in performance with these Linux distributions on Docker. Arch and Clear Linux were slightly faster than Debian and Ubuntu (Fedora's compiler had issues building this test).

While with the Stockfish chess benchmark, having KPTI enabled led to only very slight decreases in performance. Arch Linux, Clear Linux, and Fedora also were shown to be faintly slower with KPTI enabled, but nothing too meaningful and the results for Debian and Ubuntu were indifferent.

In the case of OpenSSL signing performance, no change in performance.

Or when running Caffe on the supported distributions, no measurable change in performance. This is one of the tests where Clear Linux was measurably faster than the others.

Redis is one of the benchmarks where KPTI has shown to have an impact on the performance. When testing within Docker, the losses weren't too bad -- at least nothing jaw-dropping. But for those concerned about maximum performance and not wanting to regress, when using Intel's own Clear Linux distribution compared to the other containers there was the fastest performance. Clear Linux even when the host was using KPTI was still faster than the other tested environments when KPTI was disabled on the host.

Related Articles