Re: [PATCH -mm 0/6] cgroup: block device i/o controller (v11)

[Gui Jianfeng <guijianfeng-BthXqXjhjHXQFUHtdCDX3A@public.gmane.org>]

Andrea Righi wrote:
> On 2008-11-03 08:01, Gui Jianfeng wrote:
>> Hi Andrea,
>>
>> I'v tested io-throttle(v11), it seems io-throttle doesn't work well 
>> when a process writes(delay-write) a small file(<50M may be). it can't 
>> throttle the request process accurately.
>> I guess the following reason causes this problem.
>> The write-request process terminates too early to see all accounting
>> calculated in pdflush. So io-throttle can't throttle
>> enough time for that process. Am i right?
> 
> Hi Gui,
> 
> sorry for my late first of all. You're right in part. The fact is that
> with asynchronous IO (no O_DIRECT more exactly) the accounting is
> performed when the blocks are submitted to the IO subsystem (submit_bio)
> and the throttling is performed post-facto when the same process or the
> other processes in the same cgroup write additional pages mapped to the
> limited device (throttling is performed in
> balance_dirty_pages_ratelimited_nr).
> 
> Anyway, if the process did a single small write in memory the operation
> is not accounted until pdflush writes the data back to the block device.
> When pdflush (or the process itself) performs the writeback the
> operation is accounted and next reads/writes that come from the same
> cgroup can now be throttled according to the defined limits.
> 
> So, the accounting is not lost, this information is only used later to
> throttle the next operations that will cause IO requests. IO spikes in
> this way are not prevented. With an early version of io-throttle I did
> both throttling and accounting in balance_dirty_pages_ratelimited_nr
> (when pages are written to memory), this can prevent IO spikes, but only
> if dirty_ratio is set to small values.

  Hi, Andrea,

  Thanks for your detailed explanation, I'v got your point.
  I think it makes more sense if throttling can be fair among other tasks. ;-)

> 
> Hope this clarify,
> -Andrea
> 
> 
> 

-- 
Regards
Gui Jianfeng