Increasing number of PGs

Increasing number of PGs

[Vladimir Bashkirtsev]

Hello,

I am working on optimization of ceph performance: CPU load vs OSD data 
load. Right now I have 576 PGs in total. Three pools: metadata, data, 
rbd. Each pool has 192 PGs. data is not used heavily, rbd is in heavy 
use. In total 6 OSDs in cluster. I have read recommendation about 100 
PGs per OSD. Roughly I do have 100 PGs per OSD. But now it seems that 
PGs in data pool are mostly empty while PGs in rbd pool are quite busy. 
Would it make sense to increase number of PGs to be 100 per rbd pool per 
OSD? Technically it should take some memory and CPU but because other 
two pools are virtually stand still it should not make big difference 
while improving data placement on OSDs (which is currently somewhat skewed).

So here two questions:

1. Should I increase number of PGs in rbd pool or better leave it where 
it is?
2. Wiki says that increase in PGs is not tested and only should be 
attempted on empty pool. Date on Wiki is quite old. Is this still an 
issue? Is it safe to increase PGs number on 0.49 on pool in use?

Regards,
Vladimir

Re: Increasing number of PGs

[Josh Durgin]

On 07/28/2012 02:59 AM, Vladimir Bashkirtsev wrote:
> Hello,
>
> I am working on optimization of ceph performance: CPU load vs OSD data
> load. Right now I have 576 PGs in total. Three pools: metadata, data,
> rbd. Each pool has 192 PGs. data is not used heavily, rbd is in heavy
> use. In total 6 OSDs in cluster. I have read recommendation about 100
> PGs per OSD. Roughly I do have 100 PGs per OSD. But now it seems that
> PGs in data pool are mostly empty while PGs in rbd pool are quite busy.
> Would it make sense to increase number of PGs to be 100 per rbd pool per
> OSD? Technically it should take some memory and CPU but because other
> two pools are virtually stand still it should not make big difference
> while improving data placement on OSDs (which is currently somewhat
> skewed).

If you're not using them, you can delete the data and metadata pools.
You can always recreate them later, or use different pools for cephfs.

> So here two questions:
>
> 1. Should I increase number of PGs in rbd pool or better leave it where
> it is?

It would be better to increase it from a data balancing point of view,
but it's not clear that it would help performance.

> 2. Wiki says that increase in PGs is not tested and only should be
> attempted on empty pool. Date on Wiki is quite old. Is this still an
> issue? Is it safe to increase PGs number on 0.49 on pool in use?

Right now the number of pgs in a pool can't be increased. This should
be possible in a couple months. If you can stop I/O to the current pool,
you can create a new one with more pgs, copy all the data to it, delete
the original and rename the new one.

Josh

> Regards,
> Vladimir

Re: Increasing number of PGs

[Vladimir Bashkirtsev]

On 31/07/12 01:46, Josh Durgin wrote:
>
>> So here two questions:
>>
>> 1. Should I increase number of PGs in rbd pool or better leave it where
>> it is?
>
> It would be better to increase it from a data balancing point of view,
> but it's not clear that it would help performance.
I would care about data balancing first - as I said it is currently 
skewed a bit. Performance wise perhaps it will take a bit more 
CPU/memory to process but not to the extent where it will affect 
performance.
>
>> 2. Wiki says that increase in PGs is not tested and only should be
>> attempted on empty pool. Date on Wiki is quite old. Is this still an
>> issue? Is it safe to increase PGs number on 0.49 on pool in use?
>
> Right now the number of pgs in a pool can't be increased. This should
> be possible in a couple months. If you can stop I/O to the current pool,
> you can create a new one with more pgs, copy all the data to it, delete
> the original and rename the new one.
Then I am stuck with current number of PGs: pool is in heavy use and 
there virtually no chance to increase PGs number offline. Will wait for 
a version which will make it possible.

Increasing number of PGs

[Vladimir Bashkirtsev]

Hello,

I am working on optimization of ceph performance: CPU load vs OSD data 
load. Right now I have 576 PGs in total. Three pools: metadata, data, 
rbd. Each pool has 192 PGs. data is not used heavily, rbd is in heavy 
use. In total 6 OSDs in cluster. I have read recommendation about 100 
PGs per OSD. Roughly I do have 100 PGs per OSD. But now it seems that 
PGs in data pool are mostly empty while PGs in rbd pool are quite busy. 
Would it make sense to increase number of PGs to be 100 per rbd pool per 
OSD? Technically it should take some memory and CPU but because other 
two pools are virtually stand still it should not make big difference 
while improving data placement on OSDs (which is currently somewhat skewed).

So here two questions:

1. Should I increase number of PGs in rbd pool or better leave it where 
it is?
2. Wiki says that increase in PGs is not tested and only should be 
attempted on empty pool. Date on Wiki is quite old. Is this still an 
issue? Is it safe to increase PGs number on 0.49 on pool in use?

Regards,
Vladimir