On 12/11/07, Linus Torvalds <[EMAIL PROTECTED]> wrote:
>
>
> On Tue, 11 Dec 2007, Jon Smirl wrote:
> >
> > So why does our threaded code take 20 CPU minutes longer (12%) to run
> > than the same code with a single thread?
>
> Threaded code *always* takes more CPU time. The only thing you can hope
> for is a wall-clock reduction. You're seeing probably a combination of
> (a) more cache misses
> (b) bigger dataset active at a time
> and a probably fairly miniscule
> (c) threading itself tends to have some overheads.
>
> > Q6600 is just two E6600s in the same package, the caches are not shared.
>
> Sure they are shared. They're just not *entirely* shared. But they are
> shared between each two cores, so each thread essentially has only half
> the cache they had with the non-threaded version.
>
> Threading is *not* a magic solution to all problems. It gives you
> potentially twice the CPU power, but there are real downsides that you
> should keep in mind.
>
> > Why does the threaded code need 2.24GB (google allocator, 2.85GB gcc)
> > with 4 threads? But only need 950MB with one thread? Where's the extra
> > gigabyte going?
>
> I suspect that it's really simple: you have a few rather big files in the
> gcc history, with deep delta chains. And what happens when you have four
> threads running at the same time is that they all need to keep all those
> objects that they are working on - and their hash state - in memory at the
> same time!
>
> So if you want to use more threads, that _forces_ you to have a bigger
> memory footprint, simply because you have more "live" objects that you
> work on. Normally, that isn't much of a problem, since most source files
> are small, but if you have a few deep delta chains on big files, both the
> delta chain itself is going to use memory (you may have limited the size
> of the cache, but it's still needed for the actual delta generation, so
> it's not like the memory usage went away).
This makes sense. Those runs that blew up to 4.5GB were a combination
of this effect and fragmentation in the gcc allocator. Google
allocator appears to be much better at controlling fragmentation.
Is there a reasonable scheme to force the chains to only be loaded
once and then shared between worker threads? The memory blow up
appears to be directly correlated with chain length.
>
> That said, I suspect there are a few things fighting you:
>
> - threading is hard. I haven't looked a lot at the changes Nico did to do
> a threaded object packer, but what I've seen does not convince me it is
> correct. The "trg_entry" accesses are *mostly* protected with
> "cache_lock", but nothing else really seems to be, so quite frankly, I
> wouldn't trust the threaded version very much. It's off by default, and
> for a good reason, I think.
>
> For example: the packing code does this:
>
> if (!src->data) {
> read_lock();
> src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
> read_unlock();
> ...
>
> and that's racy. If two threads come in at roughly the same time and
> see a NULL src->data, theÿ́'ll both get the lock, and they'll both
> (serially) try to fill it in. It will all *work*, but one of them will
> have done unnecessary work, and one of them will have their result
> thrown away and leaked.
That may account for the threaded version needing an extra 20 minutes
CPU time. An extra 12% of CPU seems like too much overhead for
threading. Just letting a couple of those long chain compressions be
done twice
>
> Are you hitting issues like this? I dunno. The object sorting means
> that different threads normally shouldn't look at the same objects (not
> even the sources), so probably not, but basically, I wouldn't trust the
> threading 100%. It needs work, and it needs to stay off by default.
>
> - you're working on a problem that isn't really even worth optimizing
> that much. The *normal* case is to re-use old deltas, which makes all
> of the issues you are fighting basically go away (because you only have
> a few _incremental_ objects that need deltaing).
I agree, this problem only occurs when people import giant
repositories. But every time someone hits these problems they declare
git to be screwed up and proceed to thrash it in their blogs.
> In other words: the _real_ optimizations have already been done, and
> are done elsewhere, and are much smarter (the best way to optimize X is
> not to make X run fast, but to avoid doing X in the first place!). The
> thing you are trying to work with is the one-time-only case where you
> explicitly disable that big and important optimization, and then you
> complain about the end result being slow!
>
> It's like saying that you're compiling with extreme debugging and no
> optimizations, and then complaining that the end result doesn't run as
> fast as if you used -O2. Except this is a hundred times worse, because
> you literally asked git to do the really expensive thing that it really
> really doesn't want to do ;)
>
> > Is there another allocator to try? One that combines Google's
> > efficiency with gcc's speed?
>
> See above: I'd look around at threading-related bugs and check the way we
> lock (or don't) accesses.
>
> Linus
>
--
Jon Smirl
[EMAIL PROTECTED]
