Coda File System

Re: Extreme slow performance

From: Jan Harkes <>
Date: Wed, 19 Jun 2002 09:51:19 -0400
On Wed, Jun 19, 2002 at 10:37:44AM +0200, wrote:
> we at EDS are evaluating CODA for our web server farm. We use SuSE Linux
> 8.0 (Kernel 2.4.18) and set up one SCM and two Non-SCMs (release 5.3.19-1).
> All machines are conntected with 100Mbit links. Each of the machine has
> ~30GB disk space. We set up a replicated volume called "wwwroot" over all
> three machines. Log and data partitions are raw devices: log is 30M and
> data 1G.
> We now experience an extreme performance problem. Creating and/or copying a
> file with 1MB needs aprox. 1 Minute (!). The log files doesn't show any
> problems as far as I can see.

Strange, here is my result on a triply replicated volume. I'm connected
over a 100Mbit-HD link to the servers, which are connected to a switch
and have 100Mbit-FD between them. The servers are PentiumII-200's with
128MB of memory (maybe 256MB).

Just to show that I'm not write-disconnected.

    jaharkes_at_ravel:/coda/usr/jaharkes/test$ cfs lv .
    Status of volume 0x7f000492 (2130707602) named "vmm:u.jaharkes"
    Volume type is ReadWrite
    Connection State is Connected
    Minimum quota is 0, maximum quota is unlimited
    Current blocks used are 8592
    The partition has 160118 blocks available out of 2350051
    Write-back is disabled

And here is the time it takes me to store 1MB.

    jaharkes_at_ravel:/coda/usr/jaharkes/test$ time dd if=/dev/zero of=testfile bs=1024 count=1024
    1024+0 records in
    1024+0 records out

    real    0m0.549s
    user    0m0.000s
    sys     0m0.010s

So this is about 100 times faster than what you're seeing.

Did some more timings,

    first store (create + store)
    real    0m0.492s

    second store (store only)
    real    0m0.014s

    first store (create + store)
    real    0m0.483s

    second store (store only)
    real    0m0.430s

Running it a couple of times shows that connected mode stores are all
typically around 1/2 second, whether we create or overwrite the file.
When disconnected, only the create takes approximately half a second,
but all subsequent stores are in the order of 2/100ths of a second.

Connected store-only should be faster than the create+store because we
only need one RPC instead of two. Not sure why these operations are the
same speed, it is probably the data transfer that is taking most of the
time here.

Write-disconnected create should be a lot faster, we are not sending any
data or RPCs to the servers. Not sure where the time is lost here.

Received on 2002-06-19 09:52:52