SC17 Denver, CO

P97: Profile Guided Kernel Optimization for Individual Container Execution on Bare-Metal Container


Authors: Kuniyasu Suzaki (National Institute of Advanced Industrial Science and Technology), Hidetaka Koie (National Institute of Advanced Industrial Science and Technology), Ryousei Takano (National Institute of Advanced Industrial Science and Technology)

Abstract: Container technologies become popular on supercomputers as well as in data centers. They use a container image as a package of an application, which makes easy to customize the computing environment. Unfortunately, they are not allowed to change the kernel. It means that an application cannot get the benefit of kernel optimization. Especially, Profile Guided Kernel Optimization (PGKO) is not applied.

Bare-Metal Container (BMC) tries to solve this problem. BMC utilizes remote machine management technologies (IPMI, Intel AMT, and WakeupOnLAN) to run a container image on a remote machine with a suitable Linux kernel. It enables to use PGKO easily, because the trial execution to get a profile and the optimized execution are executed automatically. Furthermore, BMC easily changes the target machine, and the user can compare the effects. We measured the performance of PGKO on big data workloads (Apache and Redis) on Xeon and i7 and found the difference.

Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf


Poster Index