[OpenAFS] Speed of replication

Turbo Fredriksson turbo@bayour.com
Sun, 20 Feb 2005 13:12:50 +0100


I had a long talk with a friend of mine a few weeks ago that now
works at KTH (Stockholm, Sweden) about AFS, replication and long
running jobs (webservers etc).

One notable thing was that my replication takes AGES, whereas
theirs takes a matter of seconds (for roughly the same data).


I only have one AFS server (Sun Blade 1000, 2x750MHz Ultra III,
1Gb memory) with all the volumes on two software RAID5 (behind
a Symbios Logic 53c875 SCSI controller w/ two channels).

There are two AFS partitions (i.e. two RAID5 devices) - 'a' and
'b'. I've tried my best to split the disks evenly on both channels
to try to increase speed.


Doing a full replication of a volume that have roughly 1.3Gb data
takes me 76 minutes (all three steps below). This replication is
done from partition 'a' to partition 'b' on the same server.


I see that the replication is done in tree steps:
  1. Cloning			Takes about  8 minutes
  2. Forwarding                 Takes about 60 minutes
  3. Deleting release volume    Takes about  8 minutes

With this, I get 288Kb/second. Seems _VERY_ (!!) slow... ?
Any ideas? Where should I start looking?

----- s n i p -----
user.turbo.src.mine
    RWrite: 536871340     ROnly: 536871341
    number of sites -> 2
       server aurora.bayour.com partition /vicepa RW Site
       server aurora.bayour.com partition /vicepb RO Site
This is a complete release of the volume 536871340
----- s n i p -----