[Gluster-users] Gluster native mount is really slow compared to nfs

Jo Goossens jo.goossens at hosted-power.com
Tue Jul 11 15:14:51 UTC 2017


Hello Joe,

 
 
I really appreciate your feedback, but I already tried the opcache stuff (to not valildate at all). It improves of course then, but not completely somehow. Still quite slow.

 
I did not try the mount options yet, but I will now!

 
 
With nfs (doesnt matter much built-in version 3 or ganesha version 4) I can even host the site perfectly fast without these extreme opcache settings.

 
I still can't understand why the nfs mount is easily 80 times faster, actually no matter what options I set it seems. It's almost there is something really wrong somehow...

 
I tried the ceph mount now and out of the box it's comparable with gluster with nfs mount.

 
 
Regards

Jo

 
BE: +32 53 599 000

NL: +31 85 888 4 555

 
https://www.hosted-power.com/

 

 
-----Original message-----
From:Joe Julian <joe at julianfamily.org>
Sent:Tue 11-07-2017 17:04
Subject:Re: [Gluster-users] Gluster native mount is really slow compared to nfs
To:gluster-users at gluster.org; 
 

My standard response to someone needing filesystem performance for www traffic is generally, "you're doing it wrong". https://joejulian.name/blog/optimizing-web-performance-with-glusterfs/
 
 That said, you might also look at these mount options: attribute-timeout, entry-timeout, negative-timeout (set to some large amount of time), and fopen-keep-cache.

 
 
On 07/11/2017 07:48 AM, Jo Goossens wrote:
 

Hello,

 

 
 

 
 

Here is the volume info as requested by soumya:

 

 
 
#gluster volume info www
 
  
Volume Name: www
 
Type: Replicate
 
Volume ID: 5d64ee36-828a-41fa-adbf-75718b954aff
 
Status: Started
 
Snapshot Count: 0
 
Number of Bricks: 1 x 3 = 3
 
Transport-type: tcp
 
Bricks:
 
Brick1: 192.168.140.41:/gluster/www
 
Brick2: 192.168.140.42:/gluster/www
 
Brick3: 192.168.140.43:/gluster/www
 
Options Reconfigured:
 
cluster.read-hash-mode: 0
 
performance.quick-read: on
 
performance.write-behind-window-size: 4MB
 
server.allow-insecure: on
 
performance.read-ahead: disable
 
performance.readdir-ahead: on
 
performance.io-thread-count: 64
 
performance.io-cache: on
 
performance.client-io-threads: on
 
server.outstanding-rpc-limit: 128
 
server.event-threads: 3
 
client.event-threads: 3
 
performance.cache-size: 32MB
 
transport.address-family: inet
 
nfs.disable: on
 
nfs.addr-namelookup: off
 
nfs.export-volumes: on
 
nfs.rpc-auth-allow: 192.168.140.*
 
features.cache-invalidation: on
 
features.cache-invalidation-timeout: 600
 
performance.stat-prefetch: on
 
performance.cache-samba-metadata: on
 
performance.cache-invalidation: on
 
performance.md-cache-timeout: 600
 
network.inode-lru-limit: 100000
 
performance.parallel-readdir: on
 
performance.cache-refresh-timeout: 60
 
performance.rda-cache-limit: 50MB
 
cluster.nufa: on
 
network.ping-timeout: 5
 
cluster.lookup-optimize: on
 
cluster.quorum-type: auto
 
  
I started with none of them set and I added/changed while testing. But it was always slow, by tuning some kernel parameters it improved slightly (just a few percent, nothing reasonable)
 
  
I also tried ceph just to compare, I got this with default settings and no tweaks:
 
  
 ./smallfile_cli.py  --top /var/www/test --host-set 192.168.140.41 --threads 8 --files 5000 --file-size 64 --record-size 64
 
smallfile version 3.0
 
                           hosts in test : ['192.168.140.41']
 
                   top test directory(s) : ['/var/www/test']
 
                               operation : cleanup
 
                            files/thread : 5000
 
                                 threads : 8
 
           record size (KB, 0 = maximum) : 64
 
                          file size (KB) : 64
 
                  file size distribution : fixed
 
                           files per dir : 100
 
                            dirs per dir : 10
 
              threads share directories? : N
 
                         filename prefix :
 
                         filename suffix :
 
             hash file number into dir.? : N
 
                     fsync after modify? : N
 
          pause between files (microsec) : 0
 
                    finish all requests? : Y
 
                              stonewall? : Y
 
                 measure response times? : N
 
                            verify read? : Y
 
                                verbose? : False
 
                          log to stderr? : False
 
                           ext.attr.size : 0
 
                          ext.attr.count : 0
 
               permute host directories? : N
 
                remote program directory : /root/smallfile-master
 
               network thread sync. dir. : /var/www/test/network_shared
 
starting all threads by creating starting gate file /var/www/test/network_shared/starting_gate.tmp
 
host = 192.168.140.41,thr = 00,elapsed = 1.339621,files = 5000,records = 0,status = ok
 
host = 192.168.140.41,thr = 01,elapsed = 1.436776,files = 5000,records = 0,status = ok
 
host = 192.168.140.41,thr = 02,elapsed = 1.498681,files = 5000,records = 0,status = ok
 
host = 192.168.140.41,thr = 03,elapsed = 1.483886,files = 5000,records = 0,status = ok
 
host = 192.168.140.41,thr = 04,elapsed = 1.454833,files = 5000,records = 0,status = ok
 
host = 192.168.140.41,thr = 05,elapsed = 1.469340,files = 5000,records = 0,status = ok
 
host = 192.168.140.41,thr = 06,elapsed = 1.439060,files = 5000,records = 0,status = ok
 
host = 192.168.140.41,thr = 07,elapsed = 1.375074,files = 5000,records = 0,status = ok
 
total threads = 8
 
total files = 40000
 
100.00% of requested files processed, minimum is  70.00
 
1.498681 sec elapsed time
 
26690.134975 files/sec
 
  
  


 Regards

 

Jo

 

 
 -----Original message-----
 From: Jo Goossens <jo.goossens at hosted-power.com> <mailto:jo.goossens at hosted-power.com> 
 Sent: Tue 11-07-2017 12:15
 Subject: Re: [Gluster-users] Gluster native mount is really slow compared to nfs
 To: Soumya Koduri <skoduri at redhat.com> <mailto:skoduri at redhat.com> ; gluster-users at gluster.org <mailto:gluster-users at gluster.org> ; 
 CC: Ambarish Soman <asoman at redhat.com> <mailto:asoman at redhat.com> ; 
 

Hello,

 

 
 

 
 

Here is some speedtest with a new setup we just made with gluster 3.10, there are no other differences, except glusterfs versus nfs. The nfs is about 80 times faster:

 

 
 

 
 
root at app1:~/smallfile-master# mount -t glusterfs -o use-readdirp=no,log-level=WARNING,log-file=/var/log/glusterxxx.log 192.168.140.41:/www /var/www
 
root at app1:~/smallfile-master# ./smallfile_cli.py  --top /var/www/test --host-set 192.168.140.41 --threads 8 --files 500 --file-size 64 --record-size 64
 
smallfile version 3.0
 
                           hosts in test : ['192.168.140.41']
 
                   top test directory(s) : ['/var/www/test']
 
                               operation : cleanup
 
                            files/thread : 500
 
                                 threads : 8
 
           record size (KB, 0 = maximum) : 64
 
                          file size (KB) : 64
 
                  file size distribution : fixed
 
                           files per dir : 100
 
                            dirs per dir : 10
 
              threads share directories? : N
 
                         filename prefix :
 
                         filename suffix :
 
             hash file number into dir.? : N
 
                     fsync after modify? : N
 
          pause between files (microsec) : 0
 
                    finish all requests? : Y
 
                              stonewall? : Y
 
                 measure response times? : N
 
                            verify read? : Y
 
                                verbose? : False
 
                          log to stderr? : False
 
                           ext.attr.size : 0
 
                          ext.attr.count : 0
 
               permute host directories? : N
 
                remote program directory : /root/smallfile-master
 
               network thread sync. dir. : /var/www/test/network_shared
 
starting all threads by creating starting gate file /var/www/test/network_shared/starting_gate.tmp
 
host = 192.168.140.41,thr = 00,elapsed = 68.845450,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 01,elapsed = 67.601088,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 02,elapsed = 58.677994,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 03,elapsed = 65.901922,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 04,elapsed = 66.971720,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 05,elapsed = 71.245102,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 06,elapsed = 67.574845,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 07,elapsed = 54.263242,files = 500,records = 0,status = ok
 
total threads = 8
 
total files = 4000
 
100.00% of requested files processed, minimum is  70.00
 
71.245102 sec elapsed time
 
56.144211 files/sec
 
  
umount /var/www
 
  
root at app1:~/smallfile-master# mount -t nfs -o tcp 192.168.140.41:/www /var/www
 
root at app1:~/smallfile-master# ./smallfile_cli.py  --top /var/www/test --host-set 192.168.140.41 --threads 8 --files 500 --file-size 64 --record-size 64
 
smallfile version 3.0
 
                           hosts in test : ['192.168.140.41']
 
                   top test directory(s) : ['/var/www/test']
 
                               operation : cleanup
 
                            files/thread : 500
 
                                 threads : 8
 
           record size (KB, 0 = maximum) : 64
 
                          file size (KB) : 64
 
                  file size distribution : fixed
 
                           files per dir : 100
 
                            dirs per dir : 10
 
              threads share directories? : N
 
                         filename prefix :
 
                         filename suffix :
 
             hash file number into dir.? : N
 
                     fsync after modify? : N
 
          pause between files (microsec) : 0
 
                    finish all requests? : Y
 
                              stonewall? : Y
 
                 measure response times? : N
 
                            verify read? : Y
 
                                verbose? : False
 
                          log to stderr? : False
 
                           ext.attr.size : 0
 
                          ext.attr.count : 0
 
               permute host directories? : N
 
                remote program directory : /root/smallfile-master
 
               network thread sync. dir. : /var/www/test/network_shared
 
starting all threads by creating starting gate file /var/www/test/network_shared/starting_gate.tmp
 
host = 192.168.140.41,thr = 00,elapsed = 0.962424,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 01,elapsed = 0.942673,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 02,elapsed = 0.940622,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 03,elapsed = 0.915218,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 04,elapsed = 0.934349,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 05,elapsed = 0.922466,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 06,elapsed = 0.954381,files = 500,records = 0,status = ok
 
host = 192.168.140.41,thr = 07,elapsed = 0.946127,files = 500,records = 0,status = ok
 
total threads = 8
 
total files = 4000
 
100.00% of requested files processed, minimum is  70.00
 
0.962424 sec elapsed time
 
4156.173189 files/sec
 
  

 
 

 
 -----Original message-----
 From: Jo Goossens <jo.goossens at hosted-power.com> <mailto:jo.goossens at hosted-power.com> 
 Sent: Tue 11-07-2017 11:26
 Subject: Re: [Gluster-users] Gluster native mount is really slow compared to nfs
 To: gluster-users at gluster.org <mailto:gluster-users at gluster.org> ; Soumya Koduri <skoduri at redhat.com> <mailto:skoduri at redhat.com> ; 
 CC: Ambarish Soman <asoman at redhat.com> <mailto:asoman at redhat.com> ; 
 

Hi all,

 

 
 

 
 

One more thing, we have 3 apps servers with the gluster on it, replicated on 3 different gluster nodes. (So the gluster nodes are app servers at the same time). We could actually almost work locally if we wouldn't need to have the same files on the 3 nodes and redundancy :)

 

 
 

Initial cluster was created like this:

 

 
 
gluster volume create www replica 3 transport tcp 192.168.140.41:/gluster/www 192.168.140.42:/gluster/www 192.168.140.43:/gluster/www force
 
gluster volume set www network.ping-timeout 5
 
gluster volume set www performance.cache-size 1024MB
 
gluster volume set www nfs.disable on # No need for NFS currently
 
gluster volume start www
 
  
To my understanding it still wouldn't explain why nfs has such great performance compared to native ...
 
  

 
 

Regards

 

Jo

 

 
 


  

 -----Original message-----
 From: Soumya Koduri <skoduri at redhat.com> <mailto:skoduri at redhat.com> 
 Sent: Tue 11-07-2017 11:16
 Subject: Re: [Gluster-users] Gluster native mount is really slow compared to nfs
 To: Jo Goossens <jo.goossens at hosted-power.com> <mailto:jo.goossens at hosted-power.com> ; gluster-users at gluster.org <mailto:gluster-users at gluster.org> ; 
 CC: Ambarish Soman <asoman at redhat.com> <mailto:asoman at redhat.com> ; Karan Sandha <ksandha at redhat.com> <mailto:ksandha at redhat.com> ; 
 + Ambarish
 
 On 07/11/2017 02:31 PM, Jo Goossens wrote:
 > Hello,
 >
 >
 >
 >
 >
 > We tried tons of settings to get a php app running on a native gluster
 > mount:
 >
 >
 >
 > e.g.: 192.168.140.41:/www /var/www glusterfs
 > defaults,_netdev,backup-volfile-servers=192.168.140.42:192.168.140.43,direct-io-mode=disable
 > 0 0
 >
 >
 >
 > I tried some mount variants in order to speed up things without luck.
 >
 >
 >
 >
 >
 > After that I tried nfs (native gluster nfs 3 and ganesha nfs 4), it was
 > a crazy performance difference.
 >
 >
 >
 > e.g.: 192.168.140.41:/www /var/www nfs4 defaults,_netdev 0 0
 >
 >
 >
 > I tried a test like this to confirm the slowness:
 >
 >
 >
 > ./smallfile_cli.py  --top /var/www/test --host-set 192.168.140.41
 > --threads 8 --files 5000 --file-size 64 --record-size 64
 >
 > This test finished in around 1.5 seconds with NFS and in more than 250
 > seconds without nfs (can't remember exact numbers, but I reproduced it
 > several times for both).
 >
 > With the native gluster mount the php app had loading times of over 10
 > seconds, with the nfs mount the php app loaded around 1 second maximum
 > and even less. (reproduced several times)
 >
 >
 >
 > I tried all kind of performance settings and variants of this but not
 > helped , the difference stayed huge, here are some of the settings
 > played with in random order:
 >
 
 Request Ambarish & Karan (cc'ed who have been working on evaluating 
 performance of various access protocols gluster supports) to look at the 
 below settings and provide inputs.
 
 Thanks,
 Soumya
 
 >
 >
 > gluster volume set www features.cache-invalidation on
 > gluster volume set www features.cache-invalidation-timeout 600
 > gluster volume set www performance.stat-prefetch on
 > gluster volume set www performance.cache-samba-metadata on
 > gluster volume set www performance.cache-invalidation on
 > gluster volume set www performance.md-cache-timeout 600
 > gluster volume set www network.inode-lru-limit 250000
 >
 > gluster volume set www performance.cache-refresh-timeout 60
 > gluster volume set www performance.read-ahead disable
 > gluster volume set www performance.readdir-ahead on
 > gluster volume set www performance.parallel-readdir on
 > gluster volume set www performance.write-behind-window-size 4MB
 > gluster volume set www performance.io-thread-count 64
 >
 > gluster volume set www performance.client-io-threads on
 >
 > gluster volume set www performance.cache-size 1GB
 > gluster volume set www performance.quick-read on
 > gluster volume set www performance.flush-behind on
 > gluster volume set www performance.write-behind on
 > gluster volume set www nfs.disable on
 >
 > gluster volume set www client.event-threads 3
 > gluster volume set www server.event-threads 3
 >
 >
 >
 >
 >
 >
 > The NFS ha adds a lot of complexity which we wouldn't need at all in our
 > setup, could you please explain what is going on here? Is NFS the only
 > solution to get acceptable performance? Did I miss one crucial settting
 > perhaps?
 >
 >
 >
 > We're really desperate, thanks a lot for your help!
 >
 >
 >
 >
 >
 > PS: We tried with gluster 3.11 and 3.8 on Debian, both had terrible
 > performance when not used with nfs.
 >
 >
 >
 >
 >
 >
 >
 > Kind regards
 >
 > Jo Goossens
 >
 >
 >
 >
 >
 >
 >
 >
 >
 > _______________________________________________
 > Gluster-users mailing list
 > Gluster-users at gluster.org <mailto:Gluster-users at gluster.org> 
 > http://lists.gluster.org/mailman/listinfo/gluster-users <http://lists.gluster.org/mailman/listinfo/gluster-users> 
 >
 
 


 _______________________________________________  Gluster-users mailing list  Gluster-users at gluster.org <mailto:Gluster-users at gluster.org>   http://lists.gluster.org/mailman/listinfo/gluster-users <http://lists.gluster.org/mailman/listinfo/gluster-users> 

 
 


_______________________________________________  Gluster-users mailing list  Gluster-users at gluster.org <mailto:Gluster-users at gluster.org>   http://lists.gluster.org/mailman/listinfo/gluster-users <http://lists.gluster.org/mailman/listinfo/gluster-users> 

 
 
 


_______________________________________________ Gluster-users mailing list Gluster-users at gluster.org <mailto:Gluster-users at gluster.org>  http://lists.gluster.org/mailman/listinfo/gluster-users <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

 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.gluster.org/pipermail/gluster-users/attachments/20170711/c443aad4/attachment.html>


More information about the Gluster-users mailing list