Skip to contentSkip to conversation

btrfs: 21 minutes to read 1.2M file directory

DISCUSSION Andy Isaacson 3 participants 5 messages Dec 20, 2010 – Dec 22, 2010 Btrfs file system

lore
Share & tools
lore
#0 lore raw

I have a directory with 1.2M files in it, which makes readdir very slow on btrfs with cold caches (although it's reasonably fast with hot caches as in the first example below):

% time find /btr/foo > /btr/foo.list
find /btr/foo > /btr/foo.list 4.10s user 7.97s system 36% cpu 33.275 total % head /btr/foo.list
/btr/foo
/btr/foo/1281373625.777.fg.jpg
/btr/foo/1281373625.777.bg.jpg
/btr/foo/1281373625.948.fg.jpg
/btr/foo/1281373625.948.bg.jpg
/btr/foo/1281373626.096.fg.jpg
/btr/foo/1281373626.096.bg.jpg
/btr/foo/1281373626.218.fg.jpg
/btr/foo/1281373626.218.bg.jpg
/btr/foo/1281373626.350.fg.jpg
% wc !$
wc /btr/foo.list
12166666 12166666 401499940 /btr/foo.list
% wc -l /btr/foo.list
12166666 /btr/foo.list
% sudo sysctl -w vm.drop_caches=3 vm.drop_caches=0
vm.drop_caches = 3
vm.drop_caches = 0
% time find /btr/foo > /btr/foo.list.2
find /btr/foo > /btr/foo.list.2 5.62s user 24.54s system 2% cpu 21:40.90 total % uname -a
Linux pyron 2.6.36-rc7-00149-g29979aa #71 SMP Wed Oct 13 09:42:57 PDT 2010 x86_64 GNU/Linux

Interestingly, while readdir is busy I'm only seeing IO on sdb even though the btrfs is on 3 targets:

Label: btr uuid: 1271de53-b3d2-4d68-9d48-b19487e1c982

        Total devices 3 FS bytes used 555.13GB
        devid    1 size 18.65GB used 18.64GB path /dev/sda2
        devid    3 size 512.00GB used 44.13GB path /dev/sdc1
        devid    2 size 512.00GB used 511.76GB path /dev/sdb1

"iostat -k 1 | grep sdb" tells me:

Device: tps kB_read/s kB_wrtn/s kB_read kB_wrtn

sdb             173.00       692.00         0.00        692          0
sdb             185.00       740.00         0.00        740          0
sdb             198.00       792.00         0.00        792          0
sdb             177.00       712.00         0.00        712          0

I updated to a recent git and it's still slow (my test hasn't completed yet 19 minutes in):

Linux pyron 2.6.37-rc6-11882-g55ec86f #72 SMP Mon Dec 20 13:34:38 PST 2010 x86_64 GNU/Linux

The devices are:

[    1.834527] ata1.00: ATA-7: INTEL SSDSA2M040G2GC, 2CV102HD, max UDMA/133
[    1.834816] ata1.00: 78165360 sectors, multi 1: LBA48 NCQ (depth 31/32)
[    1.835369] ata1.00: configured for UDMA/133
[    1.835776] scsi 0:0:0:0: Direct-Access     ATA      INTEL SSDSA2M040 2CV1 PQ: 0 ANSI: 5
...
[    2.904919] ata3.00: ATA-8: ST31500341AS, CC1H, max UDMA/133
[    2.905206] ata3.00: 2930277168 sectors, multi 0: LBA48 NCQ (depth 31/32)
[    2.947393] ata3.00: configured for UDMA/133
[    2.947850] scsi 2:0:0:0: Direct-Access     ATA      ST31500341AS     CC1H PQ: 0 ANSI: 5
...
[    3.989664] ata5: SATA link up 3.0 Gbps (SStatus 123 SControl 300)
[    4.018524] ata5.00: ATA-8: ST31500341AS, CC1H, max UDMA/133
[    4.018811] ata5.00: 2930277168 sectors, multi 0: LBA48 NCQ (depth 31/32)
[    4.060838] ata5.00: configured for UDMA/133
[    4.061205] scsi 4:0:0:0: Direct-Access     ATA      ST31500341AS     CC1H PQ: 0 ANSI: 5

The host is a "Intel(R) Core(TM) i7 CPU 930 @2.80GHz" with 12GB RAM.

Thanks,
-andy

#1 lore raw

Sigh, wrong btrfs address on the original. Apologies for the double-post.

Quoted 75 lines · Andy Isaacson

I have a directory with 1.2M files in it, which makes readdir very slow on btrfs with cold caches (although it's reasonably fast with hot caches as in the first example below):

% time find /btr/foo > /btr/foo.list
find /btr/foo > /btr/foo.list 4.10s user 7.97s system 36% cpu 33.275 total % head /btr/foo.list
/btr/foo
/btr/foo/1281373625.777.fg.jpg
/btr/foo/1281373625.777.bg.jpg
/btr/foo/1281373625.948.fg.jpg
/btr/foo/1281373625.948.bg.jpg
/btr/foo/1281373626.096.fg.jpg
/btr/foo/1281373626.096.bg.jpg
/btr/foo/1281373626.218.fg.jpg
/btr/foo/1281373626.218.bg.jpg
/btr/foo/1281373626.350.fg.jpg
% wc !$
wc /btr/foo.list
12166666 12166666 401499940 /btr/foo.list
% wc -l /btr/foo.list
12166666 /btr/foo.list
% sudo sysctl -w vm.drop_caches=3 vm.drop_caches=0
vm.drop_caches = 3
vm.drop_caches = 0
% time find /btr/foo > /btr/foo.list.2
find /btr/foo > /btr/foo.list.2 5.62s user 24.54s system 2% cpu 21:40.90 total % uname -a
Linux pyron 2.6.36-rc7-00149-g29979aa #71 SMP Wed Oct 13 09:42:57 PDT 2010 x86_64 GNU/Linux

Interestingly, while readdir is busy I'm only seeing IO on sdb even though the btrfs is on 3 targets:

Label: btr uuid: 1271de53-b3d2-4d68-9d48-b19487e1c982

        Total devices 3 FS bytes used 555.13GB
        devid    1 size 18.65GB used 18.64GB path /dev/sda2
        devid    3 size 512.00GB used 44.13GB path /dev/sdc1
        devid    2 size 512.00GB used 511.76GB path /dev/sdb1

"iostat -k 1 | grep sdb" tells me:

Device: tps kB_read/s kB_wrtn/s kB_read kB_wrtn

sdb             173.00       692.00         0.00        692          0
sdb             185.00       740.00         0.00        740          0
sdb             198.00       792.00         0.00        792          0
sdb             177.00       712.00         0.00        712          0

I updated to a recent git and it's still slow (my test hasn't completed yet 19 minutes in):

Linux pyron 2.6.37-rc6-11882-g55ec86f #72 SMP Mon Dec 20 13:34:38 PST 2010 x86_64 GNU/Linux

The devices are:

[    1.834527] ata1.00: ATA-7: INTEL SSDSA2M040G2GC, 2CV102HD, max UDMA/133
[    1.834816] ata1.00: 78165360 sectors, multi 1: LBA48 NCQ (depth 31/32)
[    1.835369] ata1.00: configured for UDMA/133
[    1.835776] scsi 0:0:0:0: Direct-Access     ATA      INTEL SSDSA2M040 2CV1 PQ: 0 ANSI: 5
...
[    2.904919] ata3.00: ATA-8: ST31500341AS, CC1H, max UDMA/133
[    2.905206] ata3.00: 2930277168 sectors, multi 0: LBA48 NCQ (depth 31/32)
[    2.947393] ata3.00: configured for UDMA/133
[    2.947850] scsi 2:0:0:0: Direct-Access     ATA      ST31500341AS     CC1H PQ: 0 ANSI: 5
...
[    3.989664] ata5: SATA link up 3.0 Gbps (SStatus 123 SControl 300)
[    4.018524] ata5.00: ATA-8: ST31500341AS, CC1H, max UDMA/133
[    4.018811] ata5.00: 2930277168 sectors, multi 0: LBA48 NCQ (depth 31/32)
[    4.060838] ata5.00: configured for UDMA/133
[    4.061205] scsi 4:0:0:0: Direct-Access     ATA      ST31500341AS     CC1H PQ: 0 ANSI: 5

The host is a "Intel(R) Core(TM) i7 CPU 930 @2.80GHz" with 12GB RAM.

Thanks,
-andy

#3 lore raw
Quoted 9 lines · Felipe Contreras
Quoted 3 lines · Andy Isaacson

I have a directory with 1.2M files in it, which makes readdir very slow on btrfs with cold caches (although it's reasonably fast with hot caches as in the first example below):

Sounds like:

Bug 21562 - btrfs is dead slow due to fragmentation
https://bugzilla.kernel.org/show_bug.cgi?id=21562

Hmmm, how do I look at the btree layout for a given inode?

btrfs-image for this filesystem is 1.7GiB .bz2, so I'm afraid it's not reasonable to publish it.

-andy

#4 lore raw
Quoted 6 lines · Andy Isaacson
Quoted 5 lines · Felipe Contreras
Quoted 3 lines · Andy Isaacson

I have a directory with 1.2M files in it, which makes readdir very slow on btrfs with cold caches (although it's reasonably fast with hot caches as in the first example below):

Quoted 6 lines
Quoted 4 lines

Sounds like:

Bug 21562 - btrfs is dead slow due to fragmentation
https://bugzilla.kernel.org/show_bug.cgi?id=21562

Hmmm, how do I look at the btree layout for a given inode?

There's documentation on the tree structures at [1] and [2]. If you
know the inode number of the object you're interested in, you need to look in the FS tree for the subvolume it's in and find the (inode_number, EXTENT_DATA, ...) keys for the file. Each of those records will reference an individual disk extent -- and you can get the disk start position and length of the extent from the data stored under the key.

Hugo.

[1] https://btrfs.wiki.kernel.org/index.php/Btree_Items
[2] https://btrfs.wiki.kernel.org/index.php/Data_Structures

Signature · 4 lines
=== Hugo Mills: hugo@... carfax.org.uk | darksatanic.net | lug.org.uk ===
  PGP key: 515C238D from wwwkeys.eu.pgp.net or http://www.carfax.org.uk
   --- Hail and greetings.  We are a flat-pack invasion force from ---   
                     Planet Ikea. We come in pieces.