Hi,
I noticed that several random IO-heavy Firefox files got fragmented
easily. Our cache suffers most.
The cache works by creating a flat file and storing fixed-size entries
in it. I though if I fallocate() the file first, then all of the writes
within the allocated area would not cause additional fragmentation.
This doesn't seem to completely cure fragmentation with ext4 in 2.6.33.
If I allocate a 4mb file, it gets more and more fragmented over time.
fallocate() does reduce fragmentation, but not as much as I expected.
I assumed that writing to an fallocate()ed area will not cause
additional fragmentation. Is my assumption incorrect?
Thanks,
Taras
ps. I'm using filefrag for measuring fragmentation.
ps2. Does running filefrag on a directory mean anything in ext4?
On 2010-09-24, at 18:05, Taras Glek wrote:
> I noticed that several random IO-heavy Firefox files got fragmented easily. Our cache suffers most.
> The cache works by creating a flat file and storing fixed-size entries in it. I though if I fallocate() the file first, then all of the writes within the allocated area would not cause additional fragmentation.
>
> This doesn't seem to completely cure fragmentation with ext4 in 2.6.33. If I allocate a 4mb file, it gets more and more fragmented over time. fallocate() does reduce fragmentation, but not as much as I expected.
Have you checked filefrag immediately after fallocating the file? Is it OK?
It may be that the issue is that an fallocate()'d file is using "unwritten extents" and converting these extents to "normal" extents may cause apparent fragmentation. However, depending on which version of e2fsprogs/filefrag you are using, it may well be that these extents only appear to be fragmented due to the different extent types.
> ps2. Does running filefrag on a directory mean anything in ext4?
With newer e2fsprogs using FIEMAP and kernels it should be possible to get useful filefrag data from a directory. Older kernels/e2fsprogs using FIBMAP will just fail outright.
Andreas Dilger wrote:
> On 2010-09-24, at 18:05, Taras Glek wrote:
>> I noticed that several random IO-heavy Firefox files got fragmented
>> easily. Our cache suffers most. The cache works by creating a flat
>> file and storing fixed-size entries in it. I though if I
>> fallocate() the file first, then all of the writes within the
>> allocated area would not cause additional fragmentation.
>>
>> This doesn't seem to completely cure fragmentation with ext4 in
>> 2.6.33. If I allocate a 4mb file, it gets more and more fragmented
>> over time. fallocate() does reduce fragmentation, but not as much
>> as I expected.
>
> Have you checked filefrag immediately after fallocating the file? Is
> it OK?
>
> It may be that the issue is that an fallocate()'d file is using
> "unwritten extents" and converting these extents to "normal" extents
> may cause apparent fragmentation. However, depending on which
> version of e2fsprogs/filefrag you are using, it may well be that
> these extents only appear to be fragmented due to the different
> extent types.
Agreed, please include filefrag (-v) output right after it's fallocated,
and also when you see this fragmentation, and then we'll have a better idea
about what you're seeing. And, the newer the filefrag the better. :)
-Eric
>> ps2. Does running filefrag on a directory mean anything in ext4?
>
> With newer e2fsprogs using FIEMAP and kernels it should be possible
> to get useful filefrag data from a directory. Older
> kernels/e2fsprogs using FIBMAP will just fail outright.
On 09/25/2010 10:26 AM, Eric Sandeen wrote:
> Andreas Dilger wrote:
>> On 2010-09-24, at 18:05, Taras Glek wrote:
>>> I noticed that several random IO-heavy Firefox files got fragmented
>>> easily. Our cache suffers most. The cache works by creating a flat
>>> file and storing fixed-size entries in it. I though if I
>>> fallocate() the file first, then all of the writes within the
>>> allocated area would not cause additional fragmentation.
>>>
>>> This doesn't seem to completely cure fragmentation with ext4 in
>>> 2.6.33. If I allocate a 4mb file, it gets more and more fragmented
>>> over time. fallocate() does reduce fragmentation, but not as much
>>> as I expected.
>> Have you checked filefrag immediately after fallocating the file? Is
>> it OK?
>>
>> It may be that the issue is that an fallocate()'d file is using
>> "unwritten extents" and converting these extents to "normal" extents
>> may cause apparent fragmentation. However, depending on which
>> version of e2fsprogs/filefrag you are using, it may well be that
>> these extents only appear to be fragmented due to the different
>> extent types.
> Agreed, please include filefrag (-v) output right after it's fallocated,
> and also when you see this fragmentation, and then we'll have a better idea
> about what you're seeing. And, the newer the filefrag the better. :)
Thanks for clarification. Turns out ext4 is performing as expected,
nevermind my previous message.
I was confused by discrepancy in number of extents reported by filefrag
1.41.10 with/without -v flag.
filefrag _CACHE_003_
_CACHE_003_: 17 extents found
filefrag -v _CACHE_003_
Filesystem type is: ef53
File size of _CACHE_003_ is 4194304 (1024 blocks, blocksize 4096)
ext logical physical expected length flags
0 0 232448 128
1 128 232576 1 unwritten
2 129 232577 95
3 224 232672 1 unwritten
4 225 232673 31
5 256 232704 1 unwritten
6 257 232705 63
7 320 232768 1 unwritten
8 321 232769 95
9 416 232864 1 unwritten
10 417 232865 255
11 672 233120 1 unwritten
12 673 233121 191
13 864 233312 1 unwritten
14 865 233313 127
15 992 233440 3
16 995 233443 29 unwritten,eof
_CACHE_003_: 1 extent found
Thanks,
Taras
Taras Glek wrote:
> On 09/25/2010 10:26 AM, Eric Sandeen wrote:
>> Andreas Dilger wrote:
>>> On 2010-09-24, at 18:05, Taras Glek wrote:
>>>> I noticed that several random IO-heavy Firefox files got fragmented
>>>> easily. Our cache suffers most. The cache works by creating a flat
>>>> file and storing fixed-size entries in it. I though if I
>>>> fallocate() the file first, then all of the writes within the
>>>> allocated area would not cause additional fragmentation.
>>>>
>>>> This doesn't seem to completely cure fragmentation with ext4 in
>>>> 2.6.33. If I allocate a 4mb file, it gets more and more fragmented
>>>> over time. fallocate() does reduce fragmentation, but not as much
>>>> as I expected.
>>> Have you checked filefrag immediately after fallocating the file? Is
>>> it OK?
>>>
>>> It may be that the issue is that an fallocate()'d file is using
>>> "unwritten extents" and converting these extents to "normal" extents
>>> may cause apparent fragmentation. However, depending on which
>>> version of e2fsprogs/filefrag you are using, it may well be that
>>> these extents only appear to be fragmented due to the different
>>> extent types.
>> Agreed, please include filefrag (-v) output right after it's fallocated,
>> and also when you see this fragmentation, and then we'll have a better
>> idea
>> about what you're seeing. And, the newer the filefrag the better. :)
> Thanks for clarification. Turns out ext4 is performing as expected,
> nevermind my previous message.
>
> I was confused by discrepancy in number of extents reported by filefrag
> 1.41.10 with/without -v flag.
:) yeah, that's an odd discrepancy, I guess...
I think filefrag needs to decide what it means by "number of extents"
found, and stick to it. For ext3 it merges in contiguous metadata,
for ext4 it doesn't count adjacent extents as separate, etc...
it does get confusing.
Anyway, glad that what you're actually seeing on disk is pretty
much what's expected. :) I am a little surprised that we are
leaving those little interspersed unwritten extents though, bits
of the code try to extend the conversion to avoid that, I thought.
-Eric
> filefrag _CACHE_003_
> _CACHE_003_: 17 extents found
> filefrag -v _CACHE_003_
> Filesystem type is: ef53
> File size of _CACHE_003_ is 4194304 (1024 blocks, blocksize 4096)
> ext logical physical expected length flags
> 0 0 232448 128
> 1 128 232576 1 unwritten
> 2 129 232577 95
> 3 224 232672 1 unwritten
> 4 225 232673 31
> 5 256 232704 1 unwritten
> 6 257 232705 63
> 7 320 232768 1 unwritten
> 8 321 232769 95
> 9 416 232864 1 unwritten
> 10 417 232865 255
> 11 672 233120 1 unwritten
> 12 673 233121 191
> 13 864 233312 1 unwritten
> 14 865 233313 127
> 15 992 233440 3
> 16 995 233443 29 unwritten,eof
> _CACHE_003_: 1 extent found
>
> Thanks,
> Taras