Re: improve inode allocation

[Ryusuke Konishi <konishi.ryusuke-Zyj7fXuS5i5L9jVzuh4AOg@public.gmane.org>]

On Wed, 24 Sep 2014 10:01:05 +0200, Andreas Rohner wrote:
> On 2014-09-23 18:35, Ryusuke Konishi wrote:
>> On Tue, 23 Sep 2014 16:21:33 +0200, Andreas Rohner wrote:
>>> On 2014-09-23 14:47, Ryusuke Konishi wrote:
>>>> By the way, if you are interested in improving this sort of bad
>>>> implemetation, please consider improving inode allocator that we can
>>>> see at nilfs_ifile_create_inode().
>>>>
>>>> It always searches free inode from ino=0.  It doesn't use the
>>>> knowledge of the last allocated inode number (inumber) nor any
>>>> locality of close-knit inodes such as a file and the directory that
>>>> contains it.
>>>>
>>>> A simple strategy is to start finding a free inode from (inumber of
>>>> the parent directory) + 1, but this may not work efficiently if the
>>>> namespace has multiple active directories, and requires that inumbers
>>>> of directories are suitably dispersed.  On the other hands, it
>>>> increases the number of disk read and also increases the number of
>>>> inode blocks to be written out if inodes are allocated too discretely.
>>>>
>>>> The optimal strategy may differ from that of other file systems
>>>> because inode blocks are not allocated to static places in nilfs.  For
>>>> example, it may be better if we gather inodes of frequently accessed
>>>> directories into the first valid inode block (on ifile) for nilfs.
>>>
>>> Sure I'll have a look at it, but this seems to be a hard problem.
>>>
>>> Since one inode has 128 bytes a typical block of 4096 contains 32
>>> inodes. We could just allocate every directory inode into an empty block
>>> with 31 free slots. Then any subsequent file inode allocation would
>>> first search the 31 slots of the parent directory and if they are full,
>>> fallback to a search starting with ino 0.
>> 
>> We can utilize several characteristics of metadata files for this
>> problem:
>> 
>> - It supports read ahead feature.  when ifile reads an inode block, we
>>   can expect that several subsequent blocks will be loaded to page
>>   cache in the background.
>> 
>> - B-tree of NILFS is efficient to hold sparse blocks.  This means that
>>   putting close-knit 32 * n inodes far from offset=0 is not so bad.
>> 
>> - ifile now can have private variables in nilfs_ifile_info (on-memory)
>>   struct.  They are available to store context information of
>>   allocator without compatibility issue.
>> 
>> - We can also use nilfs_inode_info struct of directories to store
>>   directory-based context of allocator without losing compatibility.
>> 
>> - Only caller of nilfs_ifile_create_inode() is nilfs_new_inode(), and
>>   this function knows the inode of the parent directory.
> 
> Then the only problem is how to efficiently allocate the directories. We
> could do something similar to the Orlov allocator used by the ext2/3/4
> file systems:
> 
> 1. We spread first level directories. Every one gets a full bitmap
>    block (or half a bitmap block)
> 2. For the other directories we will try to choose the bitmap block of
>    the parent unless the number of free inodes is below a certain
>    threshold. Within this bitmap block the directories should also
>    spread out.

In my understanding, the basic strategy of the Orlov allocator is to
physically spead out subtrees over cylinder groups.  This strategy is
effective for ext2/ext3/ext4 to mitigate overheads which come from
disk seeks.  The strategy increases the locality of data and metadata
and that of a parent directory and its childs nodes, but the same
thing isn't always true for nilfs because real block allocation of
ifile and other files including directories is virtualized and doesn't
reflect underlying phyiscs (e.g. relation between LBA and seek
time) as is.

I think the strategy 1 above doesn't make sense unlike ext2/3/4.

> File inodes will just start a linear search at the parents inode if
> there is enough space left in the bitmap.
> 
>>> This way if a directory has less than 32 files, all its inodes can be
>>> read in with one single block. If a directory has more than 32 files its
>>> inodes will spill over into the slots of other directories.
>>>
>>> But I am not sure if this strategy would pay off.
>> 
>> Yes, for small namespaces, the current implementation may be enough.
>> We should first decide how we evaluate the effect of the algorithm.
>> It may be the scalability of namespace.
> 
> It will be very difficult to measure the time accurately. I would
> suggest to simply count the number of reads and writes on the device.
> This can be easily done:
> 
> mkfs.nilfs2 /dev/sdb
> 
> cat /proc/diskstats > rw_before.txt
> 
> do_tests
> 
> extract_kernel_sources
> 
> ...
> 
> find /mnt
> 
> cat /proc/diskstats > rw_after.txt
> 
> The algorithm with fewer writes and reads wins.
> 
> I am still not convinced that all of this will pay off, but I will try a
> few things and see if it works.

How about measuring the following performance?

(1) Latency of inode allocation and deletion in a file system which
    holds millions of inodes.

    (Can you estimate the cost of bitmap scan per block and
     its relation with the number of inodes ?)

(2) Time to traverse a real model directory tree such as a full UNIX
    system tree or data storage of an actually-used samba server.

How do you think ?

Regards,
Ryusuke Konishi

> br,
> Andreas Rohner
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nilfs" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-nilfs" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html