ceph-osd cpu usage

ceph-osd cpu usage

[Stefan Priebe - Profihost AG]

[-- Attachment #1: Type: text/plain, Size: 368 bytes --]

Hello list,

my main problem right now is that ceph does not scale for me (more vms 
using rbd). It does not scale as the ceph-osd is using all my CPU core 
all the time (8 cores) with just 4 SSDs. The SSDs are far away from 
being loaded.

What is the best way to find out what the ceph-osd process is doing all 
the time?

A gperf graph is attached.

Greets,
Stefan

[-- Attachment #2: out.pdf --]
[-- Type: application/pdf, Size: 19302 bytes --]

Re: ceph-osd cpu usage

[Alexandre DERUMIER]

cpu usage is same for read and write  ?


----- Mail original ----- 

De: "Stefan Priebe - Profihost AG" <s.priebe@profihost.ag> 
À: ceph-devel@vger.kernel.org 
Envoyé: Jeudi 15 Novembre 2012 11:56:37 
Objet: ceph-osd cpu usage 

Hello list, 

my main problem right now is that ceph does not scale for me (more vms 
using rbd). It does not scale as the ceph-osd is using all my CPU core 
all the time (8 cores) with just 4 SSDs. The SSDs are far away from 
being loaded. 

What is the best way to find out what the ceph-osd process is doing all 
the time? 

A gperf graph is attached. 

Greets, 
Stefan 
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: ceph-osd cpu usage

[Stefan Priebe - Profihost AG]

Am 15.11.2012 12:18, schrieb Alexandre DERUMIER:
> cpu usage is same for read and write  ?

no for read it is just around 25%. And i get "full" (limited by rbd / 
librbd) 23.000 iops per vm.


> ----- Mail original -----
>
> De: "Stefan Priebe - Profihost AG" <s.priebe@profihost.ag>
> À: ceph-devel@vger.kernel.org
> Envoyé: Jeudi 15 Novembre 2012 11:56:37
> Objet: ceph-osd cpu usage
>
> Hello list,
>
> my main problem right now is that ceph does not scale for me (more vms
> using rbd). It does not scale as the ceph-osd is using all my CPU core
> all the time (8 cores) with just 4 SSDs. The SSDs are far away from
> being loaded.
>
> What is the best way to find out what the ceph-osd process is doing all
> the time?
>
> A gperf graph is attached.
>
> Greets,
> Stefan
>
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: ceph-osd cpu usage

[Mark Nelson]

Out of curiosity, does it help much if you disable crc32c calculations? 
Use the "nocrc" option in your ceph.conf file.  I've had my eye on 
crcutil as an alternative to how we do crc32c now.

http://code.google.com/p/crcutil/

Mark

On 11/15/2012 06:19 AM, Stefan Priebe - Profihost AG wrote:
> Am 15.11.2012 12:18, schrieb Alexandre DERUMIER:
>> cpu usage is same for read and write  ?
>
> no for read it is just around 25%. And i get "full" (limited by rbd /
> librbd) 23.000 iops per vm.
>
>
>> ----- Mail original -----
>>
>> De: "Stefan Priebe - Profihost AG" <s.priebe@profihost.ag>
>> À: ceph-devel@vger.kernel.org
>> Envoyé: Jeudi 15 Novembre 2012 11:56:37
>> Objet: ceph-osd cpu usage
>>
>> Hello list,
>>
>> my main problem right now is that ceph does not scale for me (more vms
>> using rbd). It does not scale as the ceph-osd is using all my CPU core
>> all the time (8 cores) with just 4 SSDs. The SSDs are far away from
>> being loaded.
>>
>> What is the best way to find out what the ceph-osd process is doing all
>> the time?
>>
>> A gperf graph is attached.
>>
>> Greets,
>> Stefan
>>
> --
> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: ceph-osd cpu usage

[Sage Weil]

On Thu, 15 Nov 2012, Stefan Priebe - Profihost AG wrote:
> Hello list,
> 
> my main problem right now is that ceph does not scale for me (more vms using
> rbd). It does not scale as the ceph-osd is using all my CPU core all the time
> (8 cores) with just 4 SSDs. The SSDs are far away from being loaded.
> 
> What is the best way to find out what the ceph-osd process is doing all the
> time?
> 
> A gperf graph is attached.

Hmm, most significant time seems to be in the allocator and doing 
fsetxattr(2) (10%!).  Also some path traversal stuff.

Can you try the wip-fd-simple-cache branch, which tries to spend less time 
closing and reopening files?  I'm curious how much of a different it will 
make for you for both IOPS and CPU utilization.

It is also possible to use leveldb for most attrs.  If you set 
'filestore xattr use omap = true' it should put most attrs in leveldb.

sage

Re: ceph-osd cpu usage

[Stefan Priebe - Profihost AG]

Am 15.11.2012 16:14, schrieb Sage Weil:
> On Thu, 15 Nov 2012, Stefan Priebe - Profihost AG wrote:
>> Hello list,
>>
>> my main problem right now is that ceph does not scale for me (more vms using
>> rbd). It does not scale as the ceph-osd is using all my CPU core all the time
>> (8 cores) with just 4 SSDs. The SSDs are far away from being loaded.
>>
>> What is the best way to find out what the ceph-osd process is doing all the
>> time?
>>
>> A gperf graph is attached.
>
> Hmm, most significant time seems to be in the allocator and doing
> fsetxattr(2) (10%!).  Also some path traversal stuff.
>
> Can you try the wip-fd-simple-cache branch, which tries to spend less time
> closing and reopening files?  I'm curious how much of a different it will
> make for you for both IOPS and CPU utilization.

Will try that this evening.

> It is also possible to use leveldb for most attrs.  If you set
> 'filestore xattr use omap = true' it should put most attrs in leveldb.

Do i have to recreate the cephfs?

Stefan

Re: ceph-osd cpu usage

[Stefan Priebe]

Am 15.11.2012 16:12, schrieb Mark Nelson:
> Out of curiosity, does it help much if you disable crc32c calculations?
> Use the "nocrc" option in your ceph.conf file.  I've had my eye on
> crcutil as an alternative to how we do crc32c now.
>
> http://code.google.com/p/crcutil/

Will try that how and where to i set the nocrc option?

Is it
[global]
nocrc = true

Stefan

Re: ceph-osd cpu usage

[Sage Weil]

On Thu, 15 Nov 2012, Stefan Priebe wrote:
> Am 15.11.2012 16:12, schrieb Mark Nelson:
> > Out of curiosity, does it help much if you disable crc32c calculations?
> > Use the "nocrc" option in your ceph.conf file.  I've had my eye on
> > crcutil as an alternative to how we do crc32c now.
> > 
> > http://code.google.com/p/crcutil/
> 
> Will try that how and where to i set the nocrc option?
> 
> Is it
> [global]
> nocrc = true

ms nocrc = true

sage

Re: ceph-osd cpu usage

[Stefan Priebe]

Hi Mark,

Am 15.11.2012 16:12, schrieb Mark Nelson:
> Out of curiosity, does it help much if you disable crc32c calculations?
> Use the "nocrc" option in your ceph.conf file.  I've had my eye on
> crcutil as an alternative to how we do crc32c now.
>
> http://code.google.com/p/crcutil/
>
> Mark
This changes nothing. CPU Load doesn't change on osds. This was a bit 
tricky ;-) i had to set nocrc on kvm host as well otherwise kvm wasn't 
starting due to bad crc.

Greets,
Stefan

Re: ceph-osd cpu usage

[Stefan Priebe]

Am 15.11.2012 16:14, schrieb Sage Weil:
> On Thu, 15 Nov 2012, Stefan Priebe - Profihost AG wrote:
> Hmm, most significant time seems to be in the allocator and doing
> fsetxattr(2) (10%!).  Also some path traversal stuff.
Yes fsetxattr seems to be CPU hungry.

> Can you try the wip-fd-simple-cache branch, which tries to spend less time
> closing and reopening files?  I'm curious how much of a different it will
> make for you for both IOPS and CPU utilization.
It seems to give me around 1000 iops across 3 VMs.

> It is also possible to use leveldb for most attrs.  If you set
> 'filestore xattr use omap = true' it should put most attrs in leveldb.
Tried this but this raises CPU by 20%.

Any other ideas how to reduce ceph-osd while doing randwrite?

Randread gives me with 3 VMs: 60.000 iops
Randwrite gives me with 3 VMs: 25.000 iops

Stefan

Re: ceph-osd cpu usage

[Alexandre DERUMIER]

>>Any other ideas how to reduce ceph-osd while doing randwrite? 
>>
>>Randread gives me with 3 VMs: 60.000 iops 
>>Randwrite gives me with 3 VMs: 25.000 iops 

Great to see that read scale !

For randwrite, what is the bottleneck now with filestore xattr use omap = true ?

Always cpu ?




----- Mail original ----- 

De: "Stefan Priebe" <s.priebe@profihost.ag> 
À: "Sage Weil" <sage@inktank.com> 
Cc: ceph-devel@vger.kernel.org 
Envoyé: Jeudi 15 Novembre 2012 21:26:06 
Objet: Re: ceph-osd cpu usage 

Am 15.11.2012 16:14, schrieb Sage Weil: 
> On Thu, 15 Nov 2012, Stefan Priebe - Profihost AG wrote: 
> Hmm, most significant time seems to be in the allocator and doing 
> fsetxattr(2) (10%!). Also some path traversal stuff. 
Yes fsetxattr seems to be CPU hungry. 

> Can you try the wip-fd-simple-cache branch, which tries to spend less time 
> closing and reopening files? I'm curious how much of a different it will 
> make for you for both IOPS and CPU utilization. 
It seems to give me around 1000 iops across 3 VMs. 

> It is also possible to use leveldb for most attrs. If you set 
> 'filestore xattr use omap = true' it should put most attrs in leveldb. 
Tried this but this raises CPU by 20%. 

Any other ideas how to reduce ceph-osd while doing randwrite? 

Randread gives me with 3 VMs: 60.000 iops 
Randwrite gives me with 3 VMs: 25.000 iops 

Stefan 
-- 
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in 
the body of a message to majordomo@vger.kernel.org 
More majordomo info at http://vger.kernel.org/majordomo-info.html 
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: ceph-osd cpu usage

[Stefan Priebe - Profihost AG]

Am 16.11.2012 09:41, schrieb Alexandre DERUMIER:
>>> Any other ideas how to reduce ceph-osd while doing randwrite?
>>>
>>> Randread gives me with 3 VMs: 60.000 iops
>>> Randwrite gives me with 3 VMs: 25.000 iops
>
> Great to see that read scale !
Yes that works fine.


> For randwrite, what is the bottleneck now with filestore xattr use omap = true ?
Oh no the result was still with default fsetxattr. Filestore xattr use 
omap added another 20% load an resultet in much lower iops. Seems i 
wasn't clear enough about that.

Stefan