[Gluster-users] Very slow rsync to gluster volume UNLESS `ls` or `find` scan dir on gluster volume first

[Tom Fite]

Hi all,

I have seen this issue as well, on Gluster 3.12.1. (3 bricks per box, 2
boxes, distributed-replicate) My testing shows the same thing -- running a
find on a directory dramatically increases lstat performance. To add
another clue, the performance degrades again after issuing a call to reset
the system's cache of dentries and inodes:

# sync; echo 2 > /proc/sys/vm/drop_caches

I think that this shows that it's the system cache that's actually doing
the heavy lifting here. There are a couple of sysctl tunables that I've
found helps out with this.

See here:

http://docs.gluster.org/en/latest/Administrator%20Guide/Linux%20Kernel%20Tuning/

Contrary to what that doc says, I've found that setting
vm.vfs_cache_pressure to a low value increases performance by allowing more
dentries and inodes to be retained in the cache.

# Set the swappiness to avoid swap when possible.
vm.swappiness = 10

# Set the cache pressure to prefer inode and dentry cache over file cache.
This is done to keep as many
# dentries and inodes in cache as possible, which dramatically improves
gluster small file performance.
vm.vfs_cache_pressure = 25

For comparison, my config is:

Volume Name: gv0
Type: Tier
Volume ID: d490a9ec-f9c8-4f10-a7f3-e1b6d3ced196
Status: Started
Snapshot Count: 13
Number of Bricks: 8
Transport-type: tcp
Hot Tier :
Hot Tier Type : Replicate
Number of Bricks: 1 x 2 = 2
Brick1: gluster2:/data/hot_tier/gv0
Brick2: gluster1:/data/hot_tier/gv0
Cold Tier:
Cold Tier Type : Distributed-Replicate
Number of Bricks: 3 x 2 = 6
Brick3: gluster1:/data/brick1/gv0
Brick4: gluster2:/data/brick1/gv0
Brick5: gluster1:/data/brick2/gv0
Brick6: gluster2:/data/brick2/gv0
Brick7: gluster1:/data/brick3/gv0
Brick8: gluster2:/data/brick3/gv0
Options Reconfigured:
performance.cache-max-file-size: 128MB
cluster.readdir-optimize: on
cluster.watermark-hi: 95
features.ctr-sql-db-cachesize: 262144
cluster.read-freq-threshold: 5
cluster.write-freq-threshold: 2
features.record-counters: on
cluster.tier-promote-frequency: 15000
cluster.tier-pause: off
cluster.tier-compact: on
cluster.tier-mode: cache
features.ctr-enabled: on
performance.cache-refresh-timeout: 60
performance.stat-prefetch: on
server.outstanding-rpc-limit: 2056
cluster.lookup-optimize: on
performance.client-io-threads: off
nfs.disable: on
transport.address-family: inet
features.barrier: disable
client.event-threads: 4
server.event-threads: 4
performance.cache-size: 1GB
network.inode-lru-limit: 90000
performance.md-cache-timeout: 600
performance.cache-invalidation: on
features.cache-invalidation-timeout: 600
features.cache-invalidation: on
performance.quick-read: on
performance.io-cache: on
performance.nfs.write-behind-window-size: 4MB
performance.write-behind-window-size: 4MB
performance.nfs.io-threads: off
network.tcp-window-size: 1048576
performance.rda-cache-limit: 64MB
performance.flush-behind: on
server.allow-insecure: on
cluster.tier-demote-frequency: 18000
cluster.tier-max-files: 1000000
cluster.tier-max-promote-file-size: 10485760
cluster.tier-max-mb: 64000
features.ctr-sql-db-wal-autocheckpoint: 2500
cluster.tier-hot-compact-frequency: 86400
cluster.tier-cold-compact-frequency: 86400
performance.readdir-ahead: off
cluster.watermark-low: 50
storage.build-pgfid: on
performance.rda-request-size: 128KB
performance.rda-low-wmark: 4KB
cluster.min-free-disk: 5%
auto-delete: enable


On Sun, Feb 4, 2018 at 9:44 PM, Amar Tumballi <atumball at redhat.com> wrote:

> Thanks for the report Artem,
>
> Looks like the issue is about cache warming up. Specially, I suspect rsync
> doing a 'readdir(), stat(), file operations' loop, where as when a find or
> ls is issued, we get 'readdirp()' request, which contains the stat
> information along with entries, which also makes sure cache is up-to-date
> (at md-cache layer).
>
> Note that this is just a off-the memory hypothesis, We surely need to
> analyse and debug more thoroughly for a proper explanation.  Some one in my
> team would look at it soon.
>
> Regards,
> Amar
>
> On Mon, Feb 5, 2018 at 7:25 AM, Vlad Kopylov <vladkopy at gmail.com> wrote:
>
>> You mounting it to the local bricks?
>>
>> struggling with same performance issues
>> try using this volume setting
>> http://lists.gluster.org/pipermail/gluster-users/2018-January/033397.html
>> performance.stat-prefetch: on might be it
>>
>> seems like when it gets to cache it is fast - those stat fetch which
>> seem to come from .gluster are slow
>>
>> On Sun, Feb 4, 2018 at 3:45 AM, Artem Russakovskii <archon810 at gmail.com>
>> wrote:
>> > An update, and a very interesting one!
>> >
>> > After I started stracing rsync, all I could see was lstat calls, quite
>> slow
>> > ones, over and over, which is expected.
>> >
>> > For example: lstat("uploads/2016/10/nexus2c
>> ee_DSC05339_thumb-161x107.jpg",
>> > {st_mode=S_IFREG|0664, st_size=4043, ...}) = 0
>> >
>> > I googled around and found
>> > https://gist.github.com/nh2/1836415489e2132cf85ed3832105fcc1, which is
>> > seeing this exact issue with gluster, rsync and xfs.
>> >
>> > Here's the craziest finding so far. If while rsync is running (or right
>> > before), I run /bin/ls or find on the same gluster dirs, it immediately
>> > speeds up rsync by a factor of 100 or maybe even 1000. It's absolutely
>> > insane.
>> >
>> > I'm stracing the rsync run, and the slow lstat calls flood in at an
>> > incredible speed as soon as ls or find run. Several hundred of files per
>> > minute (excruciatingly slow) becomes thousands or even tens of
>> thousands of
>> > files a second.
>> >
>> > What do you make of this?
>> >
>> >
>> > _______________________________________________
>> > Gluster-users mailing list
>> > Gluster-users at gluster.org
>> > http://lists.gluster.org/mailman/listinfo/gluster-users
>> _______________________________________________
>> Gluster-users mailing list
>> Gluster-users at gluster.org
>> http://lists.gluster.org/mailman/listinfo/gluster-users
>>
>
>
>
> --
> Amar Tumballi (amarts)
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://lists.gluster.org/mailman/listinfo/gluster-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.gluster.org/pipermail/gluster-users/attachments/20180205/ca4a388c/attachment.html>