On 6/16/2010 11:52 AM, Bill Davidsen wrote:

Not sure what's causing that, other than you just have your max set
pretty high for raid1.

200000 is just the default.

By leaving nothing in the way of capacity for
system operation you are filling all of memory with writes to the
resyncing drive. I would measure the speed of the disk read on the inner
tracks (dd with offset from sda to /dev/null) and not set max over 2/3
of that.

The outer tracks measure about 65000 KB/sec, and the inner tracks about
35000 KB/sec.

The problem I have with just setting sync_speed_max to a fixed, low
value like 30000 prior to starting the data-check is it needlessly slows
down the reading of the outer tracks, causing the check to take an extra
hour or so to complete.

I would prefer to see md use as much bandwidth as possible, but pause
whenever any I/O requests come in. This appears to be what the code is
designed to do -- and the log says "idle IO bandwidth" -- but given that
tasks are routinely hanging for 120+ seconds, it doesn't seem to be
working in my case.

Alternatively, you can try setting your io scheduler to
deadline,

Interesting idea. I'll give that a try and report back.

I think md would benefit from a limit on how much io can be outstanding
to any given device, but that would be a non-trivial change, I fear.

Perhaps an option to force msleep()s at regular intervals (e.g., every
50 MB) would help...

-- 
Jordan Russell