RE: Re: FSFS: Please test r9880 on Windows

On Tue, 2004-05-25 at 10:58, Leeuw van der, Tim wrote:
> Any indications of the performance of FSFS vs. BDB storage?

That's a complicated question.

For reading:

  * The server workload will generally be higher, because of the way we
store deltas. For the Subversion project repository, the server has to
do something like twice the amount of work for a checkout of the head
revision. (But with transport and working copy time added in, the
increase in observed time is much less than a factor of 2, of course.)

  * Despite the above, scalability may be better, because there are no
read-locks and no potential interference between readers and other
readers, or between readers and writers.

  * For directories with many files, FSFS will perform much better than
BDB, due to directory caching. But there's no reason this caching
couldn't be ported to BDB, and I will probably do that at some point.

For writing:

  * FSFS commits seem to be faster than BDB commits, generally speaking.

  * However, more of the work is done during the finalization stage of
the commit (after the client has submitted all of the changes), which
happens with no user feedback.

  * For directories with many many files, FSFS commits will be much
better than BDB commits, due to caching and better transaction directory
storage. BDB cannot match this performance without a schema change
because of the way it stores directories.

For storage:

  * FSFS repositories are generally 10-15% smaller than BDB
repositories, in my experience.

  * If you have many branches, savings may be much better.

  * If you have many small repositories (e.g. you're a massive hosting
site and many of the projects you host never get off the ground),
savings may be better because FSFS has less overhead.

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Tue May 25 17:13:01 2004