[Gluster-devel] performance seems extremely bad

[Anand Avati]

Dale,
 it is tough to relate the performance numbers provided by dbench to actual
real-life application's "performance" over a filesystem. dbench usually
reports very low numbers on filesystem implementations via fuse, due to the
marginally increased latency for every meta data operation. But for real
life applications the perfrmance is almost as good as a local disk and
sometimes better for very heavy I/O. Can you compare the dbench numbers (on
the latest TLA patchset) with, say, NFS or other similar network
filesystems?

thanks,
avati

2007/7/5, Dale Dude <dale at oc3networks.com>:
>
> Kernel 2.6.15. mainline-2.5 patch 275. fuse 2.6.5
>
> Tested with: dbench -t 10 10. Is performance supposed to be this bad?
>
> Glusterfs /volumes: Throughput 15.8983 MB/sec 10 procs
>
> Bypass glusterfs direct to /volume1: Throughput 65.0482 MB/sec 10 procs
>
> Bypass glusterfs direct to /volume2: Throughput 66.5139 MB/sec 10 procs
>
>
>
> =============
> client.vol:
>
> volume server1
>          type protocol/client
>          option transport-type tcp/client     # for TCP/IP transport
>          option remote-host 127.0.0.1     # IP address of the remote brick
>          option remote-subvolume volumenamespace
> end-volume
>
> volume server1vol1
>          type protocol/client
>          option transport-type tcp/client     # for TCP/IP transport
>          option remote-host 127.0.0.1     # IP address of the remote brick
>          option remote-subvolume clusterfs1
> end-volume
>
>
> volume server1vol2
>          type protocol/client
>          option transport-type tcp/client     # for TCP/IP transport
>          option remote-host 127.0.0.1     # IP address of the remote brick
>          option remote-subvolume clusterfs2
> end-volume
>
> volume bricks
>   type cluster/unify
>   option namespace server1
>   option readdir-force-success on  # ignore failed mounts
>   subvolumes server1vol1 server1vol2
>
>   option scheduler rr
>   option rr.limits.min-free-disk 5 #%
> end-volume
>
> volume writebehind   #writebehind improves write performance a lot
>   type performance/write-behind
>   option aggregate-size 131072 # in bytes
>   subvolumes bricks
> end-volume
>
> volume readahead
>   type performance/read-ahead
>   option page-size 65536     # unit in bytes
>   option page-count 16       # cache per file  = (page-count x page-size)
>   subvolumes writebehind
> end-volume
>
> volume iothreads
>    type performance/io-threads
>    option thread-count 32
>    subvolumes readahead
> end-volume
>
> ==============================
> server.vol:
>
> volume volume1
>   type storage/posix
>   option directory /volume1
> end-volume
>
> #volume posixlocks1
>   #type features/posix-locks
>   #option mandatory on          # enables mandatory locking on all files
>   #subvolumes volume1
> #end-volume
>
> volume clusterfs1
>    type performance/io-threads
>    option thread-count 16
>    subvolumes volume1
> end-volume
>
> #######
>
> volume volume2
>   type storage/posix
>   option directory /volume2
> end-volume
>
> #volume posixlocks2
>   #type features/posix-locks
>   #option mandatory on          # enables mandatory locking on all files
>   #subvolumes volume2
> #end-volume
>
> volume clusterfs2
>    type performance/io-threads
>    option thread-count 16
>    subvolumes volume2
> end-volume
>
> #######
>
> volume volumenamespace
>   type storage/posix
>   option directory /volume.namespace
> end-volume
>
> ###
>
> volume clusterfs
>   type protocol/server
>   option transport-type tcp/server
>   subvolumes clusterfs1 clusterfs2 volumenamespace
>   option auth.ip.clusterfs1.allow *
>   option auth.ip.clusterfs2.allow *
>   option auth.ip.volumenamespace.allow *
> end-volume
>
>
> _______________________________________________
> Gluster-devel mailing list
> Gluster-devel at nongnu.org
> http://lists.nongnu.org/mailman/listinfo/gluster-devel
>



-- 
Anand V. Avati