[svn.haxx.se] · SVN Dev · SVN Users · SVN Org · TSVN Dev · TSVN Users · Subclipse Dev · Subclipse Users · this month's index

Re: BDB vs FSFS - OMG!

From: Lieven Govaerts <lgo_at_apache.org>
Date: Sun, 6 Jan 2013 11:04:15 +0100

Hi,

On Sun, Jan 6, 2013 at 10:43 AM, Bert Huijben <bert_at_qqmail.nl> wrote:
> The revprop and revision cache are in fsfs, not the repos layer...
>
> In what way are you then comparing the backends?
>
> You are now comparing a backend+caching to a backend without caching.
>
> Im not against dropping support, but if we do it we should do it for the
> right reasons, not by using skewed numbers.

This comparison shows how far FSFS + all the optimizations you guys
have added on top has improved over BDB. Seems like exactly the sort
of numbers we need for this particular discussion.
Unless of course if the cost of adding these caching mechanisms to the
BDB backend, or moving it to the repo layer, is very low, but I
seriously doubt that.

> Bert

Lieven

> Sent from Windows Mail
>
> From: Stefan Fuhrmann <stefan.fuhrmann_at_wandisco.com>
> Sent: January 6, 2013 3:01 AM
> To: Subversion Development <dev_at_subversion.apache.org>
> Subject: BDB vs FSFS - OMG!
>
> Hey there,
>
> So, I did some measurement based with a mirror of the boost repository.
> That is a 82,362 revs, 971,599 changes repository mainly containing
> source code. It is surprising how much BDB has fallen behind.
>
> Tests were run on 64 bit Ubuntu 12.10 with BDB 5.1. Repository configs,
> created with latest /trunk:
>
> (1): --fs-type bdb --bdb-txn-nosync; auto log removal implied
> (2): --fs-type fsfs
> (3): --fs-type fsfs + directory deltification + compressed revprops
>
> Repository sizes (BDB is ~4GB) don't look too bad but FSFS is still
> a clear winner:
>
> (1) : (2) : (3)
> 3.1 : 1.8 : 1 disk usage
>
> Now some runtime numbers. svnadmin tests were run with '-M 4000' cache
> size.
>
> (1): (2) : (3) performance (inverted runtime)
> 1 : 3.6 : 3.5 svnadmin load (user + sys)
> 1 : 26 : 20 svnadmin verify
>
> Client test were run with svn-bench against svnserve ('--cache-revprops yes
> --cache-fulltexts yes --cache-txdeltas yes -M 4000 -c 0 --client-speed 10').
> Two runs were made for each operation and svnserve was restarted after
> the second run. BDB did not show an improvement in the second runs.
>
> (1): (2) : (3) performance (inverted runtime)
> 1 : 3.6 : 3.5 svn-bench null-export (1st run)
> 1 : 52 : 47 (2nd run)
> 1 : 21 : 18 svn-bench null-log -v (1st run)
> 1 : 136 : 134 (2nd run)
> 1 : 14 : 13 svn-bench null-log -v -g (1st run)
> 1 : 41 : 40 (2nd run)
> 1 : 6.3 : 5.7 svn-bench null-list -v -R (1st run)
> 1 : 16 : 16 (2nd run)
>
> Given these numbers, merge operations are probably also much slower with
> BDB.
>
> -- Stefan^2.
>
> --
> Certified & Supported Apache Subversion Downloads:
>
> http://www.wandisco.com/subversion/download
Received on 2013-01-06 11:05:12 CET

This is an archived mail posted to the Subversion Dev mailing list.

This site is subject to the Apache Privacy Policy and the Apache Public Forum Archive Policy.