On 2/21/16 4:56 AM, David Casier wrote:
> I made a simple test with XFS
>
> dm-sdf6-sdg1 :
> -------------------------------------------------------------------------------------------
> || sdf6 : SSD part || sdg1 : HDD (4TB) ||
> -------------------------------------------------------------------------------------------
If this is in response to my concern about not working on small
filesystems, the above is sufficiently large that inode32
won't be ignored.
> [root@aotest ~]# mkfs.xfs -f -i maxpct=0.2 /dev/mapper/dm-sdf6-sdg1
Hm, why set maxpct? This does affect how the inode32 allocator
works, but I'm wondering if that's why you set it. How did you arrive
at 0.2%? Just want to be sure you understand what you're tuning.
Thanks,
-Eric
> [root@aotest ~]# mount -o inode32 /dev/mapper/dm-sdf6-sdg1 /mnt
>
> 8 directory with 16, 32, ..., 128 sub-directory and 16, 32, ..., 128
> files (82 bytes)
> 1 xattr per dir and 3 xattr per file (user.cephosd...)
>
> 3 800 000 files and directory
> 16 GiB was written on SSD
>
> ------------------------------------------------------
> || find | wc -l ||
> ------------------------------------------------------
> || Objects per dir || % IOPS on SSD ||
> ------------------------------------------------------
> || 16 || 99 ||
> || 32 || 100 ||
> || 48 || 93 ||
> || 64 || 88 ||
> || 80 || 88 ||
> || 96 || 86 ||
> || 112 || 87 ||
> || 128 || 88 ||
> -----------------------------------------------------
>
> ------------------------------------------------------
> || find -exec getfattr '{}' \; ||
> ------------------------------------------------------
> || Objects per dir || % IOPS on SSD ||
> ------------------------------------------------------
> || 16 || 96 ||
> || 32 || 97 ||
> || 48 || 96 ||
> || 64 || 95 ||
> || 80 || 94 ||
> || 96 || 93 ||
> || 112 || 94 ||
> || 128 || 95 ||
> -----------------------------------------------------
>
> It is true that filestore is not designed to make Big Data and the
> cache must work inode / xattr
>
> I hope to see quiclky Bluestore in production :)
>
> 2016-02-19 18:06 GMT+01:00 Eric Sandeen <esandeen@xxxxxxxxxx>:
>>
>>
>> On 2/15/16 9:35 PM, Dave Chinner wrote:
>>> On Mon, Feb 15, 2016 at 04:18:28PM +0100, David Casier wrote:
>>>> Hi Dave,
>>>> 1TB is very wide for SSD.
>>>
>>> It fills from the bottom, so you don't need 1TB to make it work
>>> in a similar manner to the ext4 hack being described.
>>
>> I'm not sure it will work for smaller filesystems, though - we essentially
>> ignore the inode32 mount option for sufficiently small filesystems.
>>
>> i.e. if inode numbers > 32 bits can't exist, we don't change the allocator,
>> at least not until the filesystem (possibly) gets grown later.
>>
>> So for inode32 to impact behavior, it needs to be on a filesystem
>> of sufficient size (at least 1 or 2T, depending on block size, inode
>> size, etc). Otherwise it will have no effect today.
>>
>> Dave, I wonder if we need another mount option to essentially mean
>> "invoke the inode32 allocator regardless of filesystem size?"
>>
>> -Eric
>>
>>>> Exemple with only 10GiB :
>>>> https://www.aevoo.fr/2016/02/14/ceph-ext4-optimisation-for-filestore/
>>>
>>> It's a nice toy, but it's not something that is going scale reliably
>>> for production. That caveat at the end:
>>>
>>> "With this model, filestore rearrange the tree very
>>> frequently : + 40 I/O every 32 objects link/unlink."
>>>
>>> Indicates how bad the IO patterns will be when modifying the
>>> directory structure, and says to me that it's not a useful
>>> optimisation at all when you might be creating several thousand
>>> files/s on a filesystem. That will end up IO bound, SSD or not.
>>>
>>> Cheers,
>>>
>>> Dave.
>>>
>
>
>
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
