Re: rgw has serious GET perfomance problem with objects < 1448 bytes long

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Casey Bodley <cbodley@xxxxxxxxxx> writes:

> This feature is already present in upstream civetweb (so mimic or 
> later). You can turn it on with your rgw_frontends configuration, ie. 
> rgw_frontends = civetweb port=80 tcp_nodelay=1

It is a part of Luminous release as well; as we pulled up the version of
civetweb for the first L release.  
>
> There's an rfe to support this in the beast frontend at 
> http://tracker.ceph.com/issues/34308
>
>
> On 08/28/2018 12:44 PM, Mark Nelson wrote:
>> Excellent news Matt!  Is it worth creating a tracker ticket or would 
>> that be redundant at this point?
>>
>> Mark
>>
>>
>> On 08/28/2018 11:37 AM, Matt Benjamin wrote:
>>> That has come up before.  We have a downstream change to disable that,
>>> and it's being upstreamed.
>>>
>>> Matt
>>>
>>> On Tue, Aug 28, 2018 at 12:29 PM, Mark Nelson 
>>> <mark.a.nelson@xxxxxxxxx> wrote:
>>>> Hi Mark,
>>>>
>>>> Thanks for testing this!  Is there any chance I could get you to 
>>>> create a
>>>> tracker ticket and upload your data here?
>>>>
>>>> https://tracker.ceph.com/projects/rgw
>>>>
>>>> I doubt very many people have run performance tests on sub-4k byte 
>>>> objects.
>>>> In bluestore at least, you'll also get hit by the min-alloc size 
>>>> which means
>>>> pretty significant write and space amplification.  Out of curiosity, 
>>>> do you
>>>> know of many folks or applications that write out swift/S3 objects this
>>>> small?
>>>>
>>>> Mark
>>>>
>>>> On 08/28/2018 08:16 AM, Mark Seger wrote:
>>>>> I tried to include some data and the mail was rejected because of a
>>>>> suspected virus ;(
>>>>>
>>>>> Anyhow, running getput to talk to RGW via the swift API with varying
>>>>> object sizes I discovered GETs of objects with sizes less than 1448 
>>>>> bytes
>>>>> were taking 40msec longer which exactly corresponds to the delay 
>>>>> the nagel
>>>>> alorgithm inserts under certain buffer alignment conditiions. I 
>>>>> found this
>>>>> exact same problem with swift a number of years ago.
>>>>>
>>>>> The curious thing is I see in
>>>>> http://docs.ceph.com/docs/master/rados/configuration/network-config-ref/?highlight=public%20addr#tcp 
>>>>>
>>>>> it says nagel IS disabled in rados but maybe not RGW?
>>>>>
>>>>> -mark
>>>>>
>>>
>>>
>>
>
>




[Index of Archives]     [CEPH Users]     [Ceph Large]     [Information on CEPH]     [Linux BTRFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux