Re: Make the git codebase thread-safe

From: Stefan Zager <szager@chromium.org>
Date: Feb 12, 2014, 20:27 UTC
Message-ID: <CAHOQ7J8Q1905pVwx9QVib1BM-Xxg8vTL=hDUjT7garX++VXm3g@mail.gmail.com>
In-Reply-To: <xmqqr478m6xx.fsf@gitster.dls.corp.google.com>

On Wed, Feb 12, 2014 at 12:06 PM, Junio C Hamano <gitster@pobox.com> wrote:

Show 30 quoted lines

> Stefan Zager <szager@chromium.org> writes:
>
>> ...  I used the Very Sleepy profiler
>> to see where all the time was spent on Windows: 55% of the time was
>> spent in OpenFile, and 25% in CloseFile (both in win32).
>
> This is somewhat interesting.
>
> When we check things out, checkout_paths() has a list of paths to be
> checked out, and iterates over them and call checkout_entry().
>
> I wonder if you can:
>
>  - introduce a version of checkout_entry() that takes file
>    descriptors to write to;
>
>  - have an asynchronous helper threads that pre-open the paths to be
>    written out and feed <ce, file descriptor to be written> to a
>    queue;
>
>  - restructure that loop so that it reads the <ce, file descriptor
>    to be written> from the queue, performs the actual writing out,
>    and then feeds <file descriptor to be closed> to another queue; and
>
>  - have another asynchronous helper threads that reads <file
>    descriptor to be closed> from the queue and close them.
>
> Calls to write (and preparation of data to be written) will then
> remain single-threaded, but it sounds like that codepath is not the
> bottleneck in your measurement, so....

Yes, I considered that as well. At a minimum, that would still require attr.c to implement thread locking, since attribute files must be parsed to look for stream filters. I have already done that work.

But I'm not sure it's the best long-term approach to add convoluted custom threading solutions to each git operation as it appears on the performance radar. I'm hoping to make the entire code base more thread-friendly, so that threading can be added in a more natural and idiomatic (and less painful) way.

For example, the most natural way to add threading to checkout would be in the loops over the index in check_updates() in unpack-trees.c. If attr.c and sha1_file.c were thread-safe, then it would be possible to thread checkout entirely in check_updates(), with a pretty compact code change. I have already done the work in attr.c; sha1_file.c is hairier, but do-able.

Stefan