Date:	Wed, 12 Jan 2011 12:41:28 -0700
From:	Andreas Dilger <adilger.kernel@...ger.ca>
To:	Amir Goldstein <amir73il@...il.com>
Cc:	Lukas Czerner <lczerner@...hat.com>, linux-ext4@...r.kernel.org,
	tytso@....edu, sandeen@...hat.com, kalpak@...geny.com
Subject: Re: [RFC PATCH 0/2] Add rbtree backend for storing bitmaps

On 2011-01-12, at 05:44, Amir Goldstein wrote:
> On Mon, Jan 10, 2011 at 5:18 PM, Lukas Czerner <lczerner@...hat.com> wrote:
>> To address this I have created rbtree backend for storing bitmaps. It stores
>> just used extents of bitmaps, it means, that it can be more memory efficient
>> in most cases and also with careful use it might be much faster as well.
> 
> 2. Yet another bitmap backend.
> 
> It seems to me, that if half of the block groups are not in use and
> the other half is dense and fragmented, then neither rb_trees, nor
> arrays are the optimal backend.  A data structure similar to the page
> table, could be quite efficient in this use case, both from the memory
> usage aspect and the test/set_bit speed aspect.
> 
> While it is rare to see a block group occupied with few used block, it
> could certainly happen for inodes, so I would choose a "page" size equal
> to block size for "in-use block bitmap" and a much smaller "page" size
> for "in-use inode" bitmap.

This is essentially the "ebitmap" proposal that I had made some years ago.
It is a simple tree of data structures, and the leaf nodes are either
bitmaps (if a lot of bits are set) or a simple list (if very few bits are
set).  We might consider a third type of leaf with extents/RLE for when
there are a lot of bits set but mostly in contiguous ranges.

Initial selection of the leaf type for a given bitmap could be hinted by
the caller based on expected usage, but conversion between leaf types is
done transparently based on usage (e.g. list->bitmap when the list memory
usage exceeds a bitmap block).

> 3. More ways to reduce memory usage of e2fsck.
> 
> The most recent case of e2fsck OOM I remember showing on this list
> was cause by a file system with many hard links that were created by
> rsnapshot (http://rsnapshot.org/)
> 
> If I recall correctly, the overblown memory usage was caused by the
> icount cache, which creates an entry for every inode with nlink > 1.

I posted a patch to at least partly avoid this problem, by allowing
the icount cache to be allocated in chunks, instead of a single huge
allocation.  That allows e2fsck/kernel to swap out parts of the icount
array, and avoids the need to use 2x the memory when it is doing the
realloc() of the array to increase its size.

> If the hard linked inodes are in fact sparsely distributed
> and if the larger the refcounts, the fewer the inodes
> (let's neglect directory hard links, because we've given up on them
> for #subdirs > 32000 anyway),

Actually, for directories I think we could optimize the icount usage
somewhat as well.  We could use the inode_dir_map/S_ISDIR as a +1 link
count for the directory "." reference, so that the icount can only
count other links.  That means that directories will not need any entries
in the icount array, and I think this would reduce memory usage for
icount considerably in normal usage.

> then it is possible to replace the icount cache with 16 inode
> bitmaps, each one representing a single bit in the u16 i_links_count.
> 
> Assuming that in a normal file system the maximum links count is bounded,
> then most of these bitmaps will be empty and the rest very sparse.

I had thought of something similar, with a limited number of bitmaps
(say 2 or 3) to handle the common hard-link cases, and then fall back
to the old icount list for inodes with more links than that.

With rbtree the cost of a very sparse bitmap for the high bits would be
fairly low, so the memory usage of having 16 bitmaps is not so bad, but
there would be 16x bitmap lookups for each inode.  That might still be a
good trade-off (if memory is tight) because CPU speed is improving relative
to memory size over time.

> Even in a highly linked file system generated by 256 rsnapshots,
> the memory usage is still only about 8bits per inode, while icount
> cache stores 64bit per inode.

That isn't quite correct.  With rbtree the memory usage is in the order
of several bytes per bit in the worst case, IIRC from Lukas' email.  I
think this idea would need a lot of benchmarking before it is accepted.


Cheers, Andreas





--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists