Re: iostat show constants write to osd disk with writeahead journal, normal behaviour ?

[Mark Nelson <mark.nelson@inktank.com>]

On 6/18/12 9:04 AM, Alexandre DERUMIER wrote:
> Hi Mark,
>
>>> Sorry I got behind at looking at your output last week. I've created a
>>> seekwatcher movie of your blktrace results here:
>>>
>>> http://nhm.ceph.com/movies/mailinglist-tests/alex-test-3.4.mpg
>
> how do you create seekwatcher movie from blktrace ? (I'd like to create them myself, seem good to debug)

You'll need to download seekwatcher from Chris Mason's website.  Get the 
newest unstable version.  To make movies you'll need mencoder.  (It also 
needs numpy and matplotlib).  There is a small bug in the code where "&> 
/dev/null" should be changed to "> /dev/null 2>&1".  If you have trouble 
let me know and I can send you a fixed version of the script.

>
>
>>> The results match up well with your iostat output. Peaks and valleys in
>>> the writes every couple of seconds. Low numbers of seeks, so probably
>>> not limited by the filestore (a quick "osd tell X bench" might confirm
>>> that).
>
> yet, i'm pretty sure that the limitation if not hardware. (each osd are 15k drive, handling around 10MB/S during the test, so I think it should be ok ^_^ )
> how do you use "osd tell X bench" ?

Yeah, I just wanted to make sure that the constant writes weren't 
because the filestore was falling behind.  You may want to take a look 
at some of the information that is provided by the admin socket for the 
OSD while the test is running. dump_ops_in_flight, perf schema, and perf 
dump are all useful.

Try:

ceph --admin-daemon <socket> help

The osd admin sockets should be available in /var/run/ceph.

>
>>> I'm wondering if you increase "filestore max sync interval" to something
>>> bigger (default is 5s) if you'd see somewhat different behavior. Maybe
>>> try something like 30s and see what happens?
>
> I have done test with 30s, that doesn't change nothing.
> I have try with filestore min sync interval = 29  + filestore max sync interval = 30
>

Nuts.  Do you still see the little peaks/valleys every couple seconds?

>
>
>
> ----- Mail original -----
>
> De: "Mark Nelson"<mark.nelson@inktank.com>
> À: "Alexandre DERUMIER"<aderumier@odiso.com>
> Cc: ceph-devel@vger.kernel.org
> Envoyé: Lundi 18 Juin 2012 15:29:58
> Objet: Re: iostat show constants write to osd disk with writeahead journal, normal behaviour ?
>
> On 6/18/12 7:34 AM, Alexandre DERUMIER wrote:
>> Hi,
>>
>> I'm doing test with rados bench, and I see constant writes to osd disks.
>> Is it the normal behaviour ? with write-ahead should write occur each 20-30 seconde ?
>>
>>
>> Cluster is
>> 3 nodes (ubuntu precise - glibc 2.14 - ceph 0.47.2) with each node 1 journal on tmpfs 8GB - 1 osd (xfs) on sas disk - 1 gigabit link
>>
>>
>> 8GB journal can handle easily 20s of write (1 gigabit link)
>>
>> [osd]
>> osd data = /srv/osd.$id
>> osd journal = /tmpfs/osd.$id.journal
>> osd journal size = 8000
>> journal dio = false
>> filestore journal parallel = false
>> filestore journal writeahead = true
>> filestore fiemap = false
>>
>>
>>
>>
>> I have done tests with differents kernel (3.0,3.2,3.4) , differents filesystem (xfs,btrfs,ext4), forced journal mode to writeahead.
>> Bench were done write rados bench and fio.
>>
>> I always have constant write since the first second of bench start.
>>
>> Any idea ?
>
> Hi Alex,
>
> Sorry I got behind at looking at your output last week. I've created a
> seekwatcher movie of your blktrace results here:
>
> http://nhm.ceph.com/movies/mailinglist-tests/alex-test-3.4.mpg
>
> The results match up well with your iostat output. Peaks and valleys in
> the writes every couple of seconds. Low numbers of seeks, so probably
> not limited by the filestore (a quick "osd tell X bench" might confirm
> that).
>
> I'm wondering if you increase "filestore max sync interval" to something
> bigger (default is 5s) if you'd see somewhat different behavior. Maybe
> try something like 30s and see what happens?
>
> Mark
>
>
>
>

--
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