On 02/18/2013 04:24 PM, Mark Tinguely wrote:
> On 02/18/13 15:08, Brian Foster wrote:
>> Hi guys,
>>
>> I was running a sanity check of my quota throttling stuff rebased
>> against the updated speculative prealloc algorithm:
>>
>> a1e16c26 xfs: limit speculative prealloc size on sparse files
>>
>> ... and ran into an interesting behavior on my baseline test (quota
>> disabled).
>>
>> The test I'm running is a concurrent write of 32 files (10GB each) via
>> iozone (I'm not testing performance, just using it as a concurrent
>> writer):
>>
>> iozone -w -c -e -i 0 -+n -r 4k -s 10g -t 32 -F /mnt/data/file{0..31}
>>
>> ... what I noticed is that from monitoring du during the test,
>> speculative preallocation seemed to be ineffective. From further
>> tracing, I observed that imap[0].br_blockcount in
>> xfs_iomap_eof_prealloc_initial_size() was fairly consistently maxed out
>> at around 32768 blocks (128MB).
>>
>> Without the aforementioned commit, preallocation occurs as expected and
>> the files result in 7-9 extents after the test. With the commit, I'm in
>> the 70s to 80s range of number of extents with a max extent size of
>> 128MB. A couple examples of xfs_bmap output are appended to this
>> message. It seems like initial fragmentation might be throwing the
>> algorithm out of whack..?
>>
>> Brian
>
> ... the patched version increases in doubles
>
> + if (imap[0].br_startblock == HOLESTARTBLOCK)
> + return 0;
>
> vvvvvv
> + if (imap[0].br_blockcount <= (MAXEXTLEN >> 1))
> + return imap[0].br_blockcount;
> ^^^^^^
>
> + return XFS_B_TO_FSB(mp, XFS_ISIZE(ip));
> +}
>
> have you experimented without the middle if statement.
> If I remember correctly when I reviewed the code, that should be moving
> code closer to the original code; namely use the file size as the
> preallocation value.
>
Just a quick update...
I've tested the change above and a suggestion Dave made on IRC to return
(imap[0].br_blockcount << 1) and both resolve the immediate issue. I
need to verify the original test case still works and I'll post a patch.
Thanks...
Brian
> --Mark.
_______________________________________________
xfs mailing list
xfs@xxxxxxxxxxx
http://oss.sgi.com/mailman/listinfo/xfs
