[PATCH v2 0/3]CFQ: add think time check for service tree and group

[PATCH v2 0/3]CFQ: add think time check for service tree and group

[Shaohua Li]

Hi,
  Currently when the last queue of a service tree/group is empty, CFQ
doesn't expire the queue. This is to allow requests from the tree/group
come soon, so tree/group doesn't miss its share. But if the think time
is big, the assumption isn't correct. idling the queue is just wasting
bandwidth.

Originally I was hoping this can resolve Vivek's fsync issue, but it
doesn't. The fsync issue is caused by queue idling. But since think time
check only helps for think time above default queue idle time (8ms),
think time check doesn't help for the fsync issue.

On the other hand, think time check is still helpful for queues with
think time. I had test case in follow patches show throughput
improvement without sacrifice tree/group shares.

v1->v2:
1. addressed some good comments from Vivek
2. add more test to make sure the patches don't change behavior of
queues without think time.

Thanks,
Shaohua

Re: [PATCH v2 0/3]CFQ: add think time check for service tree and group

[Vivek Goyal]

On Tue, Jul 12, 2011 at 09:37:52AM +0800, Shaohua Li wrote:
> Hi,
>   Currently when the last queue of a service tree/group is empty, CFQ
> doesn't expire the queue. This is to allow requests from the tree/group
> come soon, so tree/group doesn't miss its share. But if the think time
> is big, the assumption isn't correct. idling the queue is just wasting
> bandwidth.
> 
> Originally I was hoping this can resolve Vivek's fsync issue, but it
> doesn't. The fsync issue is caused by queue idling. But since think time
> check only helps for think time above default queue idle time (8ms),
> think time check doesn't help for the fsync issue.
> 
> On the other hand, think time check is still helpful for queues with
> think time. I had test case in follow patches show throughput
> improvement without sacrifice tree/group shares.
> 
> v1->v2:
> 1. addressed some good comments from Vivek
> 2. add more test to make sure the patches don't change behavior of
> queues without think time.

This patch series looks good to me. Thanks Shaohua. We still shall have
to do something about fsync issue though in groups.

Acked-by: Vivek Goyal <vgoyal@redhat.com>

Thanks
Vivek

Re: [PATCH v2 0/3]CFQ: add think time check for service tree and group

[Jens Axboe]

On 2011-07-12 03:37, Shaohua Li wrote:
> Hi,
>   Currently when the last queue of a service tree/group is empty, CFQ
> doesn't expire the queue. This is to allow requests from the tree/group
> come soon, so tree/group doesn't miss its share. But if the think time
> is big, the assumption isn't correct. idling the queue is just wasting
> bandwidth.
> 
> Originally I was hoping this can resolve Vivek's fsync issue, but it
> doesn't. The fsync issue is caused by queue idling. But since think time
> check only helps for think time above default queue idle time (8ms),
> think time check doesn't help for the fsync issue.
> 
> On the other hand, think time check is still helpful for queues with
> think time. I had test case in follow patches show throughput
> improvement without sacrifice tree/group shares.
> 
> v1->v2:
> 1. addressed some good comments from Vivek
> 2. add more test to make sure the patches don't change behavior of
> queues without think time.

I applied 1-3, thanks. 1 was hand applied, since it didn't apply
directly to for-3.1/core after I merged Justins patch to get rid of the
meta data request special treatment.

-- 
Jens Axboe