xfs
[Top] [All Lists]

Re: specify agsize?

To: aurfalien <aurfalien@xxxxxxxxx>
Subject: Re: specify agsize?
From: Eric Sandeen <sandeen@xxxxxxxxxxx>
Date: Sun, 14 Jul 2013 11:14:59 -0500
Cc: xfs@xxxxxxxxxxx
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <9AB8D1D3-29D7-4C43-A624-37024CA4EFD9@xxxxxxxxx>
References: <6A14EB72-A699-47AF-937D-D6DA1CF12ACB@xxxxxxxxx> <51E2092D.7090409@xxxxxxxxxxx> <9AB8D1D3-29D7-4C43-A624-37024CA4EFD9@xxxxxxxxx>
User-agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.8; rv:17.0) Gecko/20130620 Thunderbird/17.0.7
On 7/13/13 11:20 PM, aurfalien wrote:
> 
> On Jul 13, 2013, at 7:13 PM, Eric Sandeen wrote:
> 
>> On 7/13/13 7:11 PM, aurfalien wrote:
>>> Hello again,
>>>
>>> I have a Raid 6 x16 disk array with 128k stripe size and a 512 byte block 
>>> size.
>>>
>>> So I do;
>>>
>>> mkfs.xfs -f -l size=512m -d su=128k,sw=14 /dev/mapper/vg_doofus_data-lv_data
>>>
>>> And I get;
>>>
>>> meta-data=/dev/mapper/vg_doofus_data-lv_data isize=256    agcount=32, 
>>> agsize=209428640 blks
>>>         =                       sectsz=512   attr=2, projid32bit=0
>>> data     =                       bsize=4096   blocks=6701716480, imaxpct=5
>>>         =                       sunit=32     swidth=448 blks
>>> naming   =version 2              bsize=4096   ascii-ci=0
>>> log      =internal log           bsize=4096   blocks=131072, version=2
>>>         =                       sectsz=512   sunit=32 blks, lazy-count=1
>>> realtime =none                   extsz=4096   blocks=0, rtextents=0
>>>
>>>
>>> All is fine but I was recently made aware of tweaking agsize.
>>
>> Made aware by what?  For what reason?
> 
> Autodesk has this software called Flame which requires very very fast
> local storage using XFS. They have an entire write up on how to calc
> proper agsize for optimal performance.

http://wikihelp.autodesk.com/Creative_Finishing/enu/2012/Help/05_Installation_Guides/Installation_and_Configuration_Guide_for_Linux_Workstations/0118-Advanced118/0194-Manually194/0199-Creating199

I guess?

That's quite a procedure!  And I have to say, a slightly strange one at first 
glance.

It'd be nice if they said what they were trying to accomplish rather than just 
giving you a long recipe.

In the end, I think they are trying to create 128AGs and maybe work around some 
mkfs corner case or other.

> I never mess with agsize but it is require  when creating the XFS
> file system for use with Flame.  I realize its tailored for there
> apps particular IO characteristics, so I'm curious about it.

In general more AGs allow more concurrency for some operations;
it also will generally change how/where files in multiple directories get
allocated.

>>> So I would like to mess around and iozone any diffs between the above
>>> agcount of 32 and whatever agcount changes I may do.
>>
>> Unless iozone is your machine's normal workload, that will probably prove to 
>> be uninteresting.
> 
> Well, it will give me a base line comparison of non tweaked agsize vs tweaked 
> agsize.

Not necessarily, see above; I'm not sure what iozone invocation would
show any effects from more or fewer AGs.  Anyway, iozone != flame, not
by a long shot! :)

>>> I didn't see any mention of agsize/agcount on the XFS FAQ and would
>>> like to know, based on the above, why does XFS think I have 32
>>> allocation groups with the corresponding size?
>>
>> It doesn't think so, it _knows_ so, because it made them itself.  ;)
> 
> Yea but based on what?
> 
> Why 32 at there current size?

see calc_default_ag_geometry()

Since you are in multidisk mode (you have stripe geometry) it uses more AGs for 
more AGs since it knows you have more spindles:

        } else if (dblocks > GIGABYTES(512, blocklog))
                shift = 5;

2^5 = 32

If you hadn't been in multidisk mode you would have gotten 25 AGs due to the 
max AG size of 1T.

>>> And are these optimal
>>> numbers?
>>
>> How high is up?
>>
>> Here's the appropriate faq entry:
>>
>> http://xfs.org/index.php/XFS_FAQ#Q:_I_want_to_tune_my_XFS_filesystems_for_.3Csomething.3E
> 
> Problem is I run Centos so the line;
> 
> "As of kernel 3.2.12, the default i/o scheduler, CFQ, will defeat much of the 
> parallelization in XFS. "
> 
> ... doesn't really apply.

Well, my point was that your original question, "are these optimal numbers?" 
included absolutely no context of your workload, so the best answer is yes - 
the default mkfs behavior is optimal for a generic, unspecified workload.

I don't have access to Autodesk Flame so I really don't know how it behaves or 
what an optimal tuning might be.

Anyway, I think the calc_default_ag_geometry() info above answered your 
original question of "why does XFS think I have 32 allocation groups with the 
corresponding size?" - that's simply the default mkfs algorithm when in 
multidisk mode, for a disk of this size.

-Eric

<Prev in Thread] Current Thread [Next in Thread>