[Gluster-devel] a union to two stripes to fourteen mirrors...
Anand Avati
avati at zresearch.com
Mon Nov 19 17:13:33 UTC 2007
Nyberg,
I think that 32KB stripe size is too small for 7 stripe count, and stripe
count of 7 for all files is defenitely too high in any case. You would only
want to stripe large file (if you can classify them by a filename pattern
using unify/switch). A block-size of 512KB or 1MB with a stripe count of 2
or 3 should be ideal for boosting 'general purpose' performance from a
stripe.
Also your read-ahead seems a bit too aggressive for a gigabit ethernet
setup. It takes some amount of tuning to get stripe parameters right for
performance.
About the glusterfsd dieing, We're really interested to debug that. Can you
get the logs/core if any?
avati
2007/11/19, Jerker Nyberg <jerker at update.uu.se>:
>
>
> Hi,
>
> I'm trying out different configurations of GlusterFS. I have 7 nodes each
> with two 320 GB disks where 300 GB om each disk is for the distributed
> file system.
>
> Each node is called N. Every file system is on the server side mirrored to
> the other disk on the next node, wrapped around so that the last node
> mirrors its disk to the first. invented. The real config is included in
> the end of this mail.
>
> Pseudodefinitions:
>
> fs(1) = a file system on the first disk
> fs(2) = a file system on the second disk
> n(I, fs(J)) = the fs J on node I
> afr(N .. M) = mirror the volumes
> stripe(N .. M) = stripe the volumes
>
> Server:
>
> Forw(N) = afr(fs(1), node(N+1, fs(2))
> Back(N) = afr(fs(2), node(N-1, fs(1))
>
> Client:
>
> FStr(N .. M) = stripe(n(N, Forw(N)) .. n(N+i, Forw(N+1)) .. n(M, Forw(M)))
> BStr(N .. M) = stripe(n(N, Back(N)) .. n(N+i, Back(N+1)) .. n(M, Back(M)))
> mount /glusterfs = union(FStr(1 .. 7), BStr(1..7))
>
>
>
> The goal was to get good performance but also redundancy. But this setup
> will not will it? The stripes will not work when a part of is gone and the
> union will not not magically find the other part of a file on the other
> stripe? And where to put the union namespace for good performance?
>
> But my major question is this: I tried to stripe a single stripe (not
> using union on the client, just striping on the servers which in turn
> mirrored) When rsync'ing in data on it on a single server things worked
> fine, but when I put some load on it from the other nodes (dd'ing in and
> out some large files) the glusterfsd's on the first server died... Do you
> want me to check this up more and try to reproduce and narrow down the
> problem, or is this kind of setup in general not a good idea?
>
> Regards
> Jerker Nyberg.
>
> ### client config
>
> # remote slices
> volume brick2
> type protocol/client
> option transport-type tcp/client
> option remote-host 10.0.0.2
> option remote-subvolume brick
> end-volume
> volume brick3
> type protocol/client
> option transport-type tcp/client
> option remote-host 10.0.0.3
> option remote-subvolume brick
> end-volume
> volume brick4
> type protocol/client
> option transport-type tcp/client
> option remote-host 10.0.0.4
> option remote-subvolume brick
> end-volume
> volume brick5
> type protocol/client
> option transport-type tcp/client # for TCP/IP transport
> option remote-host 10.0.0.5
> option remote-subvolume brick
> end-volume
> volume brick6
> type protocol/client
> option transport-type tcp/client
> option remote-host 10.0.0.6
> option remote-subvolume brick
> end-volume
> volume brick7
> type protocol/client
> option transport-type tcp/client
> option remote-host 10.0.0.7
> option remote-subvolume brick
> end-volume
> volume brick8
> type protocol/client
> option transport-type tcp/client
> option remote-host 10.0.0.8
> option remote-subvolume brick
> end-volume
> volume stripe
> type cluster/stripe
> subvolumes brick2 brick3 brick4 brick5 brick6 brick7 brick8
> option block-size *:32KB
> end-volume
> ### Add iothreads
> volume iothreads
> type performance/io-threads
> option thread-count 32 # deault is 1
> option cache-size 64MB #64MB
> subvolumes stripe
> end-volume
> ### Add readahead feature
> volume readahead
> type performance/read-ahead
> option page-size 256kB # unit in bytes
> # option page-count 20 # cache per file = (page-count x page-size)
> option page-count 10 # cache per file = (page-count x page-size)
> subvolumes iothreads
> end-volume
> ### Add IO-Cache feature
> volume iocache
> type performance/io-cache
> option page-size 256KB
> # option page-size 100MB
> option page-count 10
> subvolumes readahead
> end-volume
> ### Add writeback feature
> volume writeback
> type performance/write-behind
> option aggregate-size 1MB
> option flush-behind off
> subvolumes iocache
> end-volume
>
> ### server config for the 10.0.0.2
>
> # posix
> volume ba
> type storage/posix
> option directory /hda/glusterfs-a
> end-volume
> volume bc
> type storage/posix
> option directory /hdc/glusterfs-c
> end-volume
> # remote mirror
> volume mc
> type protocol/client
> option transport-type tcp/client
> option remote-host 10.0.0.3 # the next node
> option remote-subvolume bc
> end-volume
> # join
> volume afr
> type cluster/afr
> subvolumes ba mc
> end-volume
> # lock
> volume pl
> type features/posix-locks
> subvolumes afr
> end-volume
> # threads
> volume brick
> type performance/io-threads
> option thread-count 16 # deault is 1
> option cache-size 128MB #64MB
> subvolumes pl
> end-volume
> # export
> volume server
> type protocol/server
> option transport-type tcp/server
> subvolumes brick, bc
> option auth.ip.brick.allow *
> option auth.ip.bc.allow *
> end-volume
>
>
>
> # glusterfs --version
> glusterfs 1.3.8 built on Nov 16 2007
> Copyright (c) 2006, 2007 Z RESEARCH Inc. <http://www.zresearch.com>
> GlusterFS comes with ABSOLUTELY NO WARRANTY.
> You may redistribute copies of GlusterFS under the terms of the GNU
> General Public License.
>
>
>
>
> _______________________________________________
> Gluster-devel mailing list
> Gluster-devel at nongnu.org
> http://lists.nongnu.org/mailman/listinfo/gluster-devel
>
--
It always takes longer than you expect, even when you take into account
Hofstadter's Law.
-- Hofstadter's Law
More information about the Gluster-devel
mailing list