Re: [PATCH 3/5] submodule: helper to run foreach in parallel

From: Junio C Hamano <gitster@pobox.com>
Date: Aug 25, 2015, 22:23 UTC
Message-ID: <xmqqy4gzvwh5.fsf@gitster.dls.corp.google.com>
In-Reply-To: <CAGZ79kb2N_5_tJv-GURL9_ESFs=pHp=L-Mujn3Df_+-T74_9Dg@mail.gmail.com>

Stefan Beller <sbeller@google.com> writes:

Show 13 quoted lines

>>> +     while (1) {
>>> +             ssize_t len = xread(cp->err, buf, sizeof(buf));
>>> +             if (len < 0)
>>> +                     die("Read from child failed");
>>> +             else if (len == 0)
>>> +                     break;
>>> +             else {
>>> +                     strbuf_add(&out, buf, len);
>>> +             }
>>
>> ... and the whole thing is accumulated in core???
>
> The pipes have a limit, so we need to empty them to prevent back-pressure?

Of course. But that does not lead to "we hold everything in core". This side could choose to emit (under protection of args->mutex) early, e.g. after reading a line, emit it to our standard output (or our standard error).

> And because we want to have the output of one task at a time, we need to
> save it up until we can put out the whole output, no?

I do not necessarily agree, and I think I said that already:

  http://thread.gmane.org/gmane.comp.version-control.git/276273/focus=276321

Show 29 quoted lines

>>> +     }
>>> +     if (finish_command(cp))
>>> +             die("command died with error");
>>> +
>>> +     sem_wait(args->mutex);
>>> +     fputs(out.buf, stderr);
>>> +     sem_post(args->mutex);
>>
>> ... and emitted to standard error?
>>
>> I would have expected that the standard error would be left alone
>
> `git fetch` which may be a good candidate for such an operation
> provides progress on stderr, and we don't want to intermingle
> 2 different submodule fetch progress displays
> ("I need to work offline for a bit, so let me get all of the latest stuff,
> so I'll run `git submodule foreach -j 16 -- git fetch --all" though ideally
> we want to have `git fetch --recurse-submodules -j16` instead )
>
>> (i.e. letting warnings from multiple jobs to be mixed together
>> simply because everybody writes to the same file descriptor), while
>> the standard output would be line-buffered, perhaps captured by the
>> above loop and then emitted under mutex, or something.
>
>>
>> I think I said this earlier, but latency to the first output counts
>
> "to the first stderr"
> in this case?

I didn't mean "output==the standard output stream". As I said in $gmane/276321, an early output, as an indication that we are doing something, is important.

> Why would we want to unplug the task queue from somewhere else?

When you have a dispatcher more intelligent than a stupid FIFO, I would imagine that you would want to be able to do this pattern, especially when coming up with a task (not performing a task) takes non-trivial amount of work:

	prepare task queue and have N threads waiting on it;

	plug the queue, i.e. tell threads that do not start picking
	tasks out of it yet;

	large enough loop to fill the queue to a reasonable size
	while keeping the threads waiting;

	unplug the queue.  Now the threads can pick tasks from the
	queue, but they have many to choose from, and a dispatcher
	can do better than simple FIFO can take advantage of it;

	keep filling the queue with more tasks, if necessary;

        and finally, wait for everything to finish.

Without "plug/unplug" interface, you _could_ do the above by doing something stupid like

	prepare a task queue and have N threads waiting on it;

	loop to find enough number of tasks but do not put them to
	task queue, as FIFO will eat them one-by-one; instead hold
	onto them in a custom data structure that is outside the
	task queue system;

	tight and quick loop to move them to the task queue;

	keep finding more tasks and feed them to the task queue;

        and finally, wait for everything to finish.