Poor git write performance to NFS

Poor git write performance to NFS

[Daniel Steinborn]

Hi,

currently we are experiencing poor write performance when a repository 
is pushed to a nfs volume. Interestingly, this seems to be a problem in 
newer git versions:

v1.7.12.4: Very good performance

v2.1.4: Bad performance, up to 6 times slower

Are there any changed default settings or new features that can be the 
reason for that problem?

The tests are done on a Debian 8.2 VM.

Please ask for specific details if they are neccessary.

Thanks for your help! Best regards, Daniel Steinborn

-- 
Daniel Steinborn
Leibniz-Rechenzentrum
Boltzmannstraße 1
85748 Garching bei München

Re: Poor git write performance to NFS

[Junio C Hamano]

Daniel Steinborn <daniel.steinborn@lrz.de> writes:

> currently we are experiencing poor write performance when a repository
> is pushed to a nfs volume. Interestingly, this seems to be a problem
> in newer git versions:
>
> v1.7.12.4: Very good performance
>
> v2.1.4: Bad performance, up to 6 times slower
>
> Are there any changed default settings or new features that can be the
> reason for that problem?
>
> Please ask for specific details if they are neccessary.

Between 1.7.12.x series and v2.1.4, there are more than two years'
worth of changes, so it is unreasonable for anybody to expect that
such a question can be answered in a meaningful way.

Have you tried more recent versions yet?  2.1.x series is over a
year old, and I am reasonably sure there have been tons of "earlier
we did X for correctness, which unfortunately made things slower,
and this ensures the same correctness in a different way that is
much more performant" fixes since then.

Re: Poor git write performance to NFS

[Daniel Steinborn]

Hi,

thank you for your answer.
I removed the git Debian package and compiled v2.6.2 manually. The 
performance is much better now (on the same level als v1.7.12.4).
We opened a Debian bug report, to request a git version >= 2.6 to be 
backported to jessie: 
https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=802661

The issue is now solved for us, thanks again for your quick response.

Best regards,
Daniel Steinborn

Am 21.10.2015 um 20:20 schrieb Junio C Hamano:
> Daniel Steinborn <daniel.steinborn@lrz.de> writes:
>
>> currently we are experiencing poor write performance when a repository
>> is pushed to a nfs volume. Interestingly, this seems to be a problem
>> in newer git versions:
>>
>> v1.7.12.4: Very good performance
>>
>> v2.1.4: Bad performance, up to 6 times slower
>>
>> Are there any changed default settings or new features that can be the
>> reason for that problem?
>>
>> Please ask for specific details if they are neccessary.
> Between 1.7.12.x series and v2.1.4, there are more than two years'
> worth of changes, so it is unreasonable for anybody to expect that
> such a question can be answered in a meaningful way.
>
> Have you tried more recent versions yet?  2.1.x series is over a
> year old, and I am reasonably sure there have been tons of "earlier
> we did X for correctness, which unfortunately made things slower,
> and this ensures the same correctness in a different way that is
> much more performant" fixes since then.

-- 
Daniel Steinborn
Leibniz-Rechenzentrum
Boltzmannstraße 1
85748 Garching bei München

Re: Poor git write performance to NFS

[Junio C Hamano]

Daniel Steinborn <daniel.steinborn@lrz.de> writes:

> I removed the git Debian package and compiled v2.6.2 manually. The
> performance is much better now (on the same level als v1.7.12.4).

Thanks.  It is good that we must have done something good within the
past 12 months between v2.1 and v2.6, even though it feels somewhat
unsatisfactory not knowing exactly what that good thing is ;-)

Re: Poor git write performance to NFS

[Jeff King]

On Thu, Oct 22, 2015 at 09:07:42AM -0700, Junio C Hamano wrote:

> Daniel Steinborn <daniel.steinborn@lrz.de> writes:
> 
> > I removed the git Debian package and compiled v2.6.2 manually. The
> > performance is much better now (on the same level als v1.7.12.4).
> 
> Thanks.  It is good that we must have done something good within the
> past 12 months between v2.1 and v2.6, even though it feels somewhat
> unsatisfactory not knowing exactly what that good thing is ;-)

It was almost certainly the commits pulled in by e3b199a (Merge branch
'jk/prune-mtime', 2015-05-05).

And 0eeb077 (index-pack: avoid excessive re-reading of pack
directory, 2015-06-09) helps for certain commands, too.

-Peff