The genetic library patch uses a genetic algorithm to tweak the tunables of various schedulers.
As workloads change, the algorithm adjusts the tunables to optimize for performance. The patch is very useful for specific workloads that are difficult to tune.

Goal

Current Plugins

Future Work

Performance

Issues

Why in kernel

Config Locations

Patches

Goal

Tuning for workloads is tough, espcially when there are varying workloads. The goal of this patch is that sys-admins and users never have to explicitly tune their system, but have it done autonomically.

Currently I have plugins for:

Zaphod CPU scheduler

Anticipatory IO scheduler

Deadline IO scheduler (currently broken)

Possible Future work:

Incorporate pre-defined "sets of tunables". The child's genes would be a set known good of tunable configurations, and the algorithm would just use which ever set performed the best.

Save optimal tunings for workloads. Recognize changing workloads and switch to optimal settings.

Make sure genes can be negative.

Investigate using on packet scheduling.

Investigate using on SMT scheduler.

Investigate using Lamarckian Evolution on discrete parameters.

Performance

On FFSB there was an average improvement of 9%. On some workloads, I saw a 23% improvement. The only degradation was in sequential read, I saw a less than a 1%. Which was expected since AS is tuned for sequential read. I'll put my formal paper up soon w/ pretty graphics.

On SpecJBB I've seen about a 1-5% gain on the Zaphod Scheduler plugin.

On UnixBench I saw about a 2% gain for the file read/write/copy on the Anticipatory Scheduler.

I've had no luck getting a performance improvement w/ the Deadline scheduler plugin.

Issues

Many times there is no IO while a child is running. The child that does not get IO is inadvertenly penalized.

Why have it in the kernel?

Easier to hook

Low latency and not have to go inbetween user/kernel (very important for short life cycle)

Flexibility for the components to use the library

Might have inconsistent results if was application, because it would be dependent on being scheduled, but itself was in charge of modifying the scheduler. It may set the sched very wrong and never be scheduled.

Config locations

genetic-library -  "Library Routines" -> "Genetic Library"

genetic-as-sched - "Device Drivers" -> "Block IO" -> "Genetic Anticipatory IO Scheduler"

genetic-zaphod-cpu-sched - "Library Routines" -> "Genetic Library - Zaphod CPU Scheduler"

Patches in development

2.6.12 Patches



2.6.11 Patches

genetic-lib - Base genetic library patch

genetic-zaphod-cpu-sched - Zaphod CPU scheduler plugin. Depends on zaphod 6.2 patch

genetic-io-sched - Base plugin for IO schedulers

genetic-as-sched - Anticipatory scheduler plugin. Depends on genetic-io-sched.

2.6.10 Patches

genetic-lib - Base genetic library patch

genetic-zaphod-cpu-sched - Zaphod CPU scheduler plugin. Depends on zaphod 6.2 patch

genetic-io-sched - Base plugin for IO schedulers

genetic-as-sched - Anticipatory scheduler plugin. Depends on genetic-io-sched.

2.6.9 Patches

genetic-lib - Base genetic library patch

genetic-zaphod-cpu-sched - Zaphod CPU scheduler plugin. Depends on zaphod patch

genetic-io-sched - Base plugin for IO schedulers

genetic-as-sched - Anticipatory scheduler plugin. Depends on genetic-io-sched.

genetic-deadline-sched - Deadline scheduler plugin. Depends on genetic-io-sched. (currently broken)