[OT] block allocation algorithm [Was] Re: heavily fragmented file system.. How to defrag it on-line??

[Mike Fedyk]

Isaac Claymore wrote:
> On Tue, Mar 09, 2004 at 01:02:48AM -0700, Andreas Dilger wrote:
> 
>>On Mar 09, 2004  15:46 +0800, Isaac Claymore wrote:
>>
>>>I've got a workload that several clients tend to write to separate files
>>>under a same dir simultaneously, resulting in heavily fragmented files.
>>>And, even worse, those files are rarely read simultaneously, thus read
>>>performance degrades quite alot.
>>>
>>>I'm wondering whether there's any feature that helps alleviating
>>>fragmentation in such workloads. Does writing to different dirs(of a same
>>>filesystem) help?
>>
>>Very much yes.  Files allocated from different directories will get blocks
>>from different parts of the filesystem (if available), so they should be
>>less fragmented.  In 2.6 there is a heuristic that files opened by different
>>processes allocate from different parts of a group, even within the same
>>directory, but that only really helps if the files themselves aren't too
>>large (i.e. under 8MB or so).
>>
> 
> 
> Thanks, I did some test on this last weekend, and here're the results in
> case someone is interested:
> 

Yes, very interesting.  It would be nice if JFS was in this comparison.

Also, the next step is running each filesystem under your workload (Like 
test 5) and compare the fragmentation and performance over a longer 
period of time.

Read below for some more comments...

> 
> Test environment:
> 
> kernel: 2.6.3 with latest reiser4 patches applied.
> OS: Debian testing/unstable
> HW: Intel(R) Pentium(R) 4 CPU 1.80GHz, 256M RAM
> 
> For each FS configuration, my test went on as: dumping 3 files of 1G
> each simultaneously, and measure the fragmentation with 'filefrag'.
> 
> Each test iteration was done on a freshly formatted filesystem.
> 
> Here goes the figures & my evaluations:
> 
> 1. reiser3, 3 files under a same dir:
> 
> sandbox:/mnt/foo [1016]# dd if=/dev/zero of=f0 bs=16M count=64&;dd 
> if=/dev/zero of=f1 bs=16M count=64&;dd if=/dev/zero of=f2 bs=16M count=64&;wait
> 
> sandbox:/mnt/foo [1018]# filefrag f0 f1 f2
> f0: 470 extents found
> f1: 461 extents found
> f2: 470 extents found
> 
> 
> My Evaluation: badly fragmented!
> 
> 
> 2. reiser3, 3 files under 3 different dirs:
> 
> sandbox:/mnt [1028]# dd if=/dev/zero of=dir0/foo bs=16M count=64&;dd 
> if=/dev/zero of=dir1/foo bs=16M count=64&;dd if=/dev/zero of=dir2/foo 
> bs=16M count=64&;wait
> 
> sandbox:/mnt [1029]# filefrag dir0/foo dir1/foo dir2/foo
> dir0/foo: 448 extents found
> dir1/foo: 462 extents found
> dir2/foo: 443 extents found
> 
> 
> My Evaluation: still bad, spreading the files under different dirs did no 
> visible good.

How about with 10 dirs?

> 
> 
> 
> 3. ext3, 3 files under a same dir:
> 
> sandbox:/mnt/foo [1041]# dd if=/dev/zero of=f0 bs=16M count=64&;dd 
> if=/dev/zero of=f1 bs=16M count=64&;dd if=/dev/zero of=f2 bs=16M 
> count=64&;wait
> 
> sandbox:/mnt/foo [1044]# filefrag f0 f1 f2
> f0: 202 extents found, perfection would be 9 extents
> f1: 207 extents found, perfection would be 9 extents
> f2: 208 extents found, perfection would be 9 extents
> 
> 
> My Evaluation: much better than reiser3, yet far from perfection.
> 
> 
> 
> 4. ext3, 3 files under 3 different dirs:
> 
> sandbox:/mnt [1054]# dd if=/dev/zero of=dir0/foo bs=16M count=64&;dd
> if=/dev/zero of=dir1/foo bs=16M count=64&;dd if=/dev/zero of=dir2/foo 
> bs=16M count=64&;wait
> 
> sandbox:/mnt [1056]# sandbox:/mnt [1056]# filefrag dir0/foo dir1/foo dir2/foo
> dir0/foo: 91 extents found, perfection would be 9 extents
> dir1/foo: 9 extents found
> dir2/foo: 95 extents found, perfection would be 9 extents
> 
> 
> My Evaluation: spreading the files under different dirs DID help quite
> alot! but can we get even better result by spread the files more sparsely?
> (see next test)
> 
> 
> 5. still ext3, mkdir 10 dirs first, then dumping the files under the
> 1st, 5th, and 9th dirs:
> 
> sandbox:/mnt [1085]# dd if=/dev/zero of=dir0/foo bs=16M count=64&;dd 
> if=/dev/zero of=dir4/foo bs=16M count=64&;dd if=/dev/zero of=dir9/foo 
> bs=16M count=64&;wait
> 
> sandbox:/mnt [1086]# filefrag dir{0,4,9}/foo
> dir0/foo: 11 extents found, perfection would be 9 extents
> dir4/foo: 11 extents found, perfection would be 9 extents
> dir9/foo: 10 extents found, perfection would be 9 extents
> 
> 
> My Evaluation: almost perfect!
> 
> 
> 
> 6. XFS, 3 files under a same dir:
> 
> sandbox:/mnt/foo [1112]# dd if=/dev/zero of=f0 bs=16M count=64&;dd if=/dev/zero 
> of=f1 bs=16M count=64&;dd if=/dev/zero of=f2 bs=16M count=64&;wait
> 
> sandbox:/mnt/foo [1114]# filefrag f0 f1 f2
> f0: 25 extents found
> f1: 11 extents found
> f2: 20 extents found
> 
> 
> My Evaluation: this'd be the BEST result I got, when dumping into a same dir.
> 
> 
> 
> 
> 7. XFS, dumping into 3 dirs among ten, similar to test 5:
> 
> sandbox:/mnt [1127]# dd if=/dev/zero of=dir0/foo bs=16M count=64&;dd if=/dev/zero 
> of=dir4/foo bs=16M count=64&;dd if=/dev/zero of=dir9/foo bs=16M count=64&;wait
> 
> sandbox:/mnt [1128]# filefrag dir0/foo dir4/foo dir9/foo
> dir0/foo: 1 extent found
> dir4/foo: 1 extent found
> dir9/foo: 1 extent found
> 
> 
> My Evaluation: impressed! cant be any better now.
> 

How about with 3 dirs?

> 
> 8. Reiser4, 1 dir
> 
> sandbox:/mnt/foo [1155]# dd if=/dev/zero of=f0 bs=16M count=64&;dd if=/dev/zero 
> of=f1 bs=16M count=64&;dd if=/dev/zero of=f2 bs=16M count=64&;wait
> 
> sandbox:/mnt/foo [1156]# filefrag f0 f1 f2
> f0: 45 extents found
> f1: 6011 extents found
> f2: 45 extents found
> 
> 
> My Evaluation: far better than it's brother reiser3. the 6011 extents of
> f1 was weird, i'd have done more iterations to get an average, just blame
> lazy me ;)
> 
> 
> 
> 9. Reiser4, 3 dirs among 10:
> 
> sandbox:/mnt [1165]# dd if=/dev/zero of=dir0/foo bs=16M count=64&;dd 
> if=/dev/zero of=dir4/foo bs=16M count=64&;dd if=/dev/zero of=dir9/foo 
> bs=16M count=64&;wait
> 
> sandbox:/mnt [1167]# filefrag dir{0,4,9}/foo
> dir0/foo: 42 extents found
> dir4/foo: 50 extents found
> dir9/foo: 46 extents found
> 
> 
> My Evaluation: nice figures, really. and unlike its elder brother, using
> more dirs DID help.

How about with 3 dirs?

Mike