On 3/1/20 3:32 AM, Steven Davies wrote:
On 02/01/2020 10:12, Anand Jain wrote:
As of now we use %pid method to read stripped mirrored data. So
application's process id determines the stripe id to be read. This type
of read IO routing typically helps in a system with many small
independent applications tying to read random data. On the other hand
the %pid based read IO distribution policy is inefficient if there is a
single application trying to read large data and the overall disk
bandwidth remains under utilized.
So this patch introduces a framework where we could add more readmirror
policies, such as routing the IO based on device's wait-queue or manual
when we have a read-preferred device or a policy based on the target
storage caching.
I think the idea is good but that it would be cleaner if the tunable was
named read_policy rather than readmirror as it's more obvious that it
contains a policy tunable.
Um. 'read_policy' sounds good, but I hope it is clear enough to
indicate that we are talking about read for only mirrored-chunks.
Will rename to read_policy.
Do you envisage allowing more than one policy to be active for a
filesystem? If not, what about using the same structure as the CPU
frequency and block IO schedulers with the format
#cat /sys/block/sda/queue/scheduler
noop [deadline] cfq
Such that btrfs would (eventually) have something like
#cat /sys/fs/btrfs/UUID/read_policy
by_pid [user_defined_device] by_shortest_queue
And in case of user_defined_device, the device for the read shall be
specified in
cat /sys/fs/btrfs/UUID/devinfo/<devid>/read_preferred
0 = unset, 1 = set.
(devinfo patches are in the ML [1] open for comment)
[1]
[PATCH v3 4/4] btrfs: sysfs, add devid/dev_state kobject and attribute
And the policy would be changed by echo'ing the new policy name to the
read_policy kobject.
I like this approach, will change it to use this format.
Thanks, Anand
Steve
