vrv / linux-iopoll-deferral Goto Github PK

Deferred I/O polling for the Linux ATA driver
Author: Vijay Vasudevan ([email protected])

This work is aimed at improving the interrupt mitigation techniques
for modern flash devices. This patch is an experimental feature and is
not fully general (see note below).

It is targeted toward recent kernel versions (2.6.32+) because it
relies on the blk-iopoll infrastructure. This patch combines the ahci
blk-iopoll support patch (after its split into libahci) with the
deferred polling modifications.  It patches and merges properly with
Linux 2.6.36 to 2.6.39, and has been tested to run properly on 2.6.36.

Notes:

Our experience with this patch has shown significantly
increased random IOPS rates when random I/O performance is CPU-bound,
but this has slightly adverse impacts on sequential throughput.

This modification works best for single drive configurations: the
deferral of completion processing currently breaks out of the iopoll
loop.  If there are multiple iopoll structures (multiple drives), then
we should continue processing the other drives, especially if they
have a low I/O rate or are latency-sensitive.


Use:

The following sysctl has been added to play around with the parameters
of deferral. Requires kernel.blk_iopoll=1 to activate.

sysctl -w kernel.blk_iopoll_defer_by=X

  Defers the processing of an interrupt by X microseconds when the
  previous interrupt completed only 1 command.


TODOs:
1) Make defer_by value adaptive based on I/O rate.
2) Fix unfair polling by continuing instead of breaking.

Please contact [email protected] if you have any questions about the patch.
Pull requests encouraged!