[ATTEND] Want to attend LSF 2010

[ATTEND] Want to attend LSF 2010

[Vivek Goyal]

Resending this mail. This time also CCing linux-scsi mailing list. Sorry
for the duplicate copy.

Hi,

I wanted to attend LSF 2010. Sepcifically I was interested in discussing
couple of things.

- CFQ performance issues on higher end storage
        - Deadline outperforms CFQ on higher end storage (storage arrays
          and also on host based hardware RAID). I was wondering if there
          is a way to fix it or it is just design limitation. In the past
          Jens had mentioned that he wants to get rid of deadline also and
          be replaced with CFQ. Are we still targeting that and if yes,
          how to achieve that (some kind of auto tuning).

- Max bandwidth IO controller
        - A basic proportional weight controller (blkio) based off CFQ is
          now in. Now there is also a need to implement throttling/max
          bandwidth controller. Wanted to get some ideas on how to go
          go about it and what is the best place to impement it. Implement
          it in CFQ or a new device mapper target or something else.

Thanks
Vivek

Re: [ATTEND] Want to attend LSF 2010

[Ric Wheeler]

On 07/01/2010 02:04 PM, Vivek Goyal wrote:
> Resending this mail. This time also CCing linux-scsi mailing list. Sorry
> for the duplicate copy.
>
> Hi,
>
> I wanted to attend LSF 2010. Sepcifically I was interested in discussing
> couple of things.
>
> - CFQ performance issues on higher end storage
>          - Deadline outperforms CFQ on higher end storage (storage arrays
>            and also on host based hardware RAID). I was wondering if there
>            is a way to fix it or it is just design limitation. In the past
>            Jens had mentioned that he wants to get rid of deadline also and
>            be replaced with CFQ. Are we still targeting that and if yes,
>            how to achieve that (some kind of auto tuning).
>
> - Max bandwidth IO controller
>          - A basic proportional weight controller (blkio) based off CFQ is
>            now in. Now there is also a need to implement throttling/max
>            bandwidth controller. Wanted to get some ideas on how to go
>            go about it and what is the best place to impement it. Implement
>            it in CFQ or a new device mapper target or something else.
>
> Thanks
> Vivek
>

Thanks Vivek,

I know that this IO scheduling issue has been a hot issue for us in our 
performance testing. We would love to figure out how to get CFQ to displace 
deadline totally and definitely it would be very interesting to have a 
conversation around what needs done/what can be done.

Vivek is a local, so no travel needed....

Thanks!

Ric

Re: [Lsf10-pc] [ATTEND] Want to attend LSF 2010

[Chris Mason]

On Thu, Jul 01, 2010 at 02:10:06PM -0400, Ric Wheeler wrote:
> On 07/01/2010 02:04 PM, Vivek Goyal wrote:
> > Resending this mail. This time also CCing linux-scsi mailing list. Sorry
> > for the duplicate copy.
> >
> > Hi,
> >
> > I wanted to attend LSF 2010. Sepcifically I was interested in discussing
> > couple of things.
> >
> > - CFQ performance issues on higher end storage
> >          - Deadline outperforms CFQ on higher end storage (storage arrays
> >            and also on host based hardware RAID). I was wondering if there
> >            is a way to fix it or it is just design limitation. In the past
> >            Jens had mentioned that he wants to get rid of deadline also and
> >            be replaced with CFQ. Are we still targeting that and if yes,
> >            how to achieve that (some kind of auto tuning).
> >
> > - Max bandwidth IO controller
> >          - A basic proportional weight controller (blkio) based off CFQ is
> >            now in. Now there is also a need to implement throttling/max
> >            bandwidth controller. Wanted to get some ideas on how to go
> >            go about it and what is the best place to impement it. Implement
> >            it in CFQ or a new device mapper target or something else.
> >
> > Thanks
> > Vivek
> >
> 
> Thanks Vivek,
> 
> I know that this IO scheduling issue has been a hot issue for us in our 
> performance testing. We would love to figure out how to get CFQ to displace 
> deadline totally and definitely it would be very interesting to have a 
> conversation around what needs done/what can be done.
> 

I'll second that...one scheduler to rule them all.

-chris

Re: [Lsf10-pc] [ATTEND] Want to attend LSF 2010

[Jens Axboe]

On 01/07/10 20.40, Chris Mason wrote:
> On Thu, Jul 01, 2010 at 02:10:06PM -0400, Ric Wheeler wrote:
>> On 07/01/2010 02:04 PM, Vivek Goyal wrote:
>>> Resending this mail. This time also CCing linux-scsi mailing list. Sorry
>>> for the duplicate copy.
>>>
>>> Hi,
>>>
>>> I wanted to attend LSF 2010. Sepcifically I was interested in discussing
>>> couple of things.
>>>
>>> - CFQ performance issues on higher end storage
>>>          - Deadline outperforms CFQ on higher end storage (storage arrays
>>>            and also on host based hardware RAID). I was wondering if there
>>>            is a way to fix it or it is just design limitation. In the past
>>>            Jens had mentioned that he wants to get rid of deadline also and
>>>            be replaced with CFQ. Are we still targeting that and if yes,
>>>            how to achieve that (some kind of auto tuning).
>>>
>>> - Max bandwidth IO controller
>>>          - A basic proportional weight controller (blkio) based off CFQ is
>>>            now in. Now there is also a need to implement throttling/max
>>>            bandwidth controller. Wanted to get some ideas on how to go
>>>            go about it and what is the best place to impement it. Implement
>>>            it in CFQ or a new device mapper target or something else.
>>>
>>> Thanks
>>> Vivek
>>>
>>
>> Thanks Vivek,
>>
>> I know that this IO scheduling issue has been a hot issue for us in our 
>> performance testing. We would love to figure out how to get CFQ to displace 
>> deadline totally and definitely it would be very interesting to have a 
>> conversation around what needs done/what can be done.
>>
> 
> I'll second that...one scheduler to rule them all.

Agreed :-)

-- 
Jens Axboe