On Sat, Jun 27, 2015 at 6:20 PM, Stefan Priebe <s.priebe@xxxxxxxxxxxx> wrote:
> Dear Ilya,
>
> Am 25.06.2015 um 14:07 schrieb Ilya Dryomov:
>>
>> On Wed, Jun 24, 2015 at 10:29 PM, Stefan Priebe <s.priebe@xxxxxxxxxxxx>
>> wrote:
>>>
>>>
>>> Am 24.06.2015 um 19:53 schrieb Ilya Dryomov:
>>>>
>>>>
>>>> On Wed, Jun 24, 2015 at 8:38 PM, Stefan Priebe <s.priebe@xxxxxxxxxxxx>
>>>> wrote:
>>>>>
>>>>>
>>>>>
>>>>> Am 24.06.2015 um 16:55 schrieb Nick Fisk:
>>>>>>
>>>>>>
>>>>>>
>>>>>> That kernel probably has the bug where tcp_nodelay is not enabled.
>>>>>> That
>>>>>> is fixed in Kernel 4.0+, however also in 4.0 blk-mq was introduced
>>>>>> which
>>>>>> brings two other limitations:-
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> blk-mq is terrible slow. That's correct.
>>>>
>>>>
>>>>
>>>> Is that a general sentiment or your experience with rbd? If the
>>>> latter, can you describe your workload and provide some before and
>>>> after blk-mq numbers? We'd be very interested in identifying and
>>>> fixing any performance regressions you might have on blk-mq rbd.
>>>
>>>
>>>
>>> oh i'm sorry. I accidently compiled blk-mq into the kernel when 3.18.1
>>> came
>>> out and i was wondering why my I/O waits on my ceph osds where doubled or
>>> even tripled. After reverting back to cfq everything was fine again. I
>>> didn't digged deeper into it as i thought blk-mq is experimental in 3.18.
>>
>>
>> That doesn't make sense - rbd was switched to blk-mq in 4.0. Or did
>> you try to apply the patch from the mailing list to 3.18?
>
>
> I'm talking about the ceph-osd process / side not about rbd client side.
Ah, sorry - Nick was clearly talking about the kernel client and
I replied to his mail. The kernel you run your OSDs on shouldn't
matter much, as long as it's not something ancient (except when you
need to work around a particular filesystem bug), so I just assumed
you and German were talking about the kernel client.
Thanks,
Ilya
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
