K 10
svn:author
V 3
imp
K 8
svn:date
V 27
2015-03-06T00:24:21.210068Z
K 7
svn:log
V 1407
Implement the limiting code for the scheduler. reads and writes go to
different queues (phyiscally now, logically eventually). The limit on
the number of outstanding reads / writes can be controlled
independently on a per drive basis.

Implement proper 32.32 * 32.32 multiplication and use it when we're
multiplying two 32.32 numbers.

Ensure SD calculation is as robust as possible, and return 0 when
we're unable to properly estimate it (such as early in startup when
the long term values haven't converged).

It is looking like work loads very too much to make the fast decaying
average a useful metric. Implement an alternate method of estimating a
steering point for unacceptable latency. We now train for the first
few hundred I/Os to see what the drive will normally do unloaded, then
use a multiple of the 95% latency to set a threshold for badness (to
reflect the number of channels in the disk). We retain the old slow
decaying average as well.

Create a read_bias parameter for breaking ties between read and writes
when both are present. This ensures that we make some progress on
writes, while still balancing them against reads that are competing
against them. Set it to 100 to 1.

Allow for a tunable to control whether we do the netflix schedule
stuff or not. Due to the large amount of state we keep, and due to
the difficulty in 'draining' pending I/Os we only allow it to be set
at boot.

END