[OpenAFS] Speed of release or incremental release (replication)?

[Turbo Fredriksson]

>>>>> "Sergio" == Sergio Gelato <Sergio.Gelato@astro.su.se> writes:

    Sergio> Most probably a configuration error, may be a performance
    Sergio> problem.  "vos release" works fine for me; I just did an
    Sergio> initial release of a 300MB volume (three RO replicas on
    Sergio> different servers, as usual) and it took about one minute,
    Sergio> two at most. Not too bad for 100Mb/s Ethernet and ATA/100
    Sergio> disks.

I posted a mail about two hours ago, stating that I might have found
the problem (the release is going through an ancient SPARC).

    Sergio> That's one thing we're doing differently: I normally have
    Sergio> a RO replica colocated with the RW, as the manuals
    Sergio> recommend.

Basically that's what I'm doing, just not on the same partition (it
shouldn't matter because the disks are fast enough)...

    Sergio> Question: do you have a lot of writes to user.turbo.mail
    Sergio> during the release?

Absolutely 0. The volume locks during cloning, and after the clone is
done (and the replication is done), there ARE writes. Quite a lot,
but not HUGE... I get lots of mails, but not horrendous (1000/day
roughly :).

    Sergio> Shouldn't matter once the local RO clone is ready, other
    Sergio> than competing for kernel time. Was the initial cloning fast?

No, it was almost WORSE!

    >> It say 'This is a complete release of the volume' so this leads
    >> me to think that it should be possible to do a 'non-complete'
    >> (ie, incremental) release... ?

    Sergio> I'd have to check, but I rather suspect it refers to the
    Sergio> possibility of some RO copies being outdated while others
    Sergio> are up to date.

Since there is only one RW and one RO (same machine, different disks)
and the replication only takes place twice every twelve hours (it was
just not possible to do it more often), then yes, the RO _is_ outdated...

    Sergio> What OS kernel is this, and which version of OpenAFS?

Kernel: 2.4.21-xfs1.3.0pre2
OpenAFS kernel module: 1.2.6-1.TF.1+10.00.Custom
OpenAFS user land: 1.2.7-1

It was recommended in another thread that I should test without XFS, and
I'm currently compiling a 2.6.0-test8 kernel (with OpenAFS module 1.2.7-1).

    Sergio> One wonders about SMP locking bugs... Can you repeat the
    Sergio> experiment with a non-SMP kernel (using only one of your
    Sergio> CPUs)? If it's faster, you've got your culprit.

Oki, I've cancelled the kernel build and de-selecting SMP. I'll get back
when I know more...

    >> I MUST have configured something wrong, because I can't believe
    >> that AFS is this slow!!
    >> 
    >> The machine is a dual PII 333MHz with Ultra160 disks 'in full
    >> working order'... The machine IS loaded, but not that much...

    Sergio> By the way, wouldn't it make more sense to simply place
    Sergio> your /vicepc on a RAID-1 metadevice?

Absolutely! I'll send you the bill! :)

    Sergio> AFS replication is nice when you have multiple servers,
    Sergio> and then only for volumes with nearly-static contents.  I
    Sergio> wouldn't use it on home volumes, where nearly all users
    Sergio> will want to access the RW version most of the time.

Currently I only have one file server, but the intention is to build
more when I can afford it and separate the 'interactive' (such as shell,
IMAP/pop) services from the 'non-interactive' (such as web and SMTP).

    Sergio> For backups, there are better tools such as "vos dump" and
    Sergio> "backup".

Yes, that is done once every night to tape...


-- 
spy AK-47 South Africa Kennedy 747 toluene Nazi $400 million in gold
bullion Waco, Texas kibo pits plutonium Delta Force president class
struggle
[See http://www.aclu.org/echelonwatch/index.html for more about this]