Re: [PATCH 2/2] archive: avoid spawning `gzip`

From: René Scharfe <l.s.r@web.de>
Date: Apr 27, 2019, 09:59 UTC
Message-ID: <f6f32bc0-109c-e0eb-f7d2-9e46647f260c@web.de>
In-Reply-To: <nycvar.QRO.7.76.6.1904261047560.45@tvgsbejvaqbjf.bet>

Am 26.04.19 um 16:51 schrieb Johannes Schindelin:> Hi Peff,

Show 19 quoted lines

>
> On Mon, 15 Apr 2019, Jeff King wrote:
>
>> On Sun, Apr 14, 2019 at 12:01:10AM +0200, René Scharfe wrote:
>>
>>>>> As we already link to the zlib library, we can perform the compression
>>>>> without even requiring gzip on the host machine.
>>>>
>>>> Very cool. It's nice to drop a dependency, and this should be a bit more
>>>> efficient, too.
>>>
>>> Getting rid of dependencies is good, and using zlib is the obvious way to
>>> generate .tgz files. Last time I tried something like that, a separate gzip
>>> process was faster, though -- at least on Linux [1].  How does this one
>>> fare?
>>
>> I'd expect a separate gzip to be faster in wall-clock time for a
>> multi-core machine, but overall consume more CPU. I'm slightly surprised
>> that your timings show that it actually wins on total CPU, too.

My initial expectation back then was that moving data between processes is costly and that compressing in-process would improve the overall performance. Your expectation is more in line with what I then actually saw. The difference in total CPU time wasn't that big, perhaps just noise.

> If performance is really a concern, you'll be much better off using `pigz`
> than `gzip`.

Performance is always a concern, but on the other hand I didn't see any complaints about slow archiving so far.

Show 14 quoted lines

>> Here are best-of-five times for "git archive --format=tar.gz HEAD" on
>> linux.git (the machine is a quad-core):
>>
>>    [before, separate gzip]
>>    real	0m21.501s
>>    user	0m26.148s
>>    sys	0m0.619s
>>
>>    [after, internal gzwrite]
>>    real	0m25.156s
>>    user	0m25.059s
>>    sys	0m0.096s
>>
>> which does show what I expect (longer overall, but less total CPU).

I get similar numbers with hyperfine:

Benchmark #1: git archive --format=tar.gz HEAD >/dev/null
  Time (mean ± σ):     16.683 s ±  0.451 s    [User: 20.230 s, System: 0.375 s]
  Range (min … max):   16.308 s … 17.852 s    10 runs

Benchmark #2: ~/src/git/git-archive --format=tar.gz HEAD >/dev/null
  Time (mean ± σ):     19.898 s ±  0.228 s    [User: 19.825 s, System: 0.073 s]
  Range (min … max):   19.627 s … 20.355 s    10 runs

Benchmark #3: git archive --format=zip HEAD >/dev/null
  Time (mean ± σ):     16.449 s ±  0.075 s    [User: 16.340 s, System: 0.109 s]
  Range (min … max):   16.326 s … 16.611 s    10 runs

#1 is git v2.21.0, #2 is with the two patches applied, #3 is v2.21.0 again, but with zip output, just to put things into perspective.

Show 11 quoted lines

>> Which one you prefer depends on your situation, of course. A user on a
>> workstation with multiple cores probably cares most about end-to-end
>> latency and using all of their available horsepower. A server hosting
>> repositories and receiving many unrelated requests probably cares more
>> about total CPU (though the differences there are small enough that it
>> may not even be worth having a config knob to un-parallelize it).
>
> I am a bit sad that this is so noticeable. Nevertheless, I think that
> dropping the dependency is worth it, in particular given that `gzip` is
> not exactly fast to begin with (you really should switch to `pigz` or to a
> faster compression if you are interested in speed).

We could import pigz verbatim, it's just 11K LOCs total. :)

Show 7 quoted lines

>>> Doing compression in its own thread may be a good idea.
>>
>> Yeah. It might even make the patch simpler, since I'd expect it to be
>> implemented with start_async() and a descriptor, making it look just
>> like a gzip pipe to the caller. :)
>
> Sadly, it does not really look like it is simpler.

I have to agree -- at least I was unable to pull off the stdout plumbing trick. Is there a way? But it doesn't look too bad, and the performance is closer to using the real gzip:

Benchmark #1: ~/src/git/git-archive --format=tar.gz HEAD >/dev/null
  Time (mean ± σ):     17.300 s ±  0.198 s    [User: 20.825 s, System: 0.356 s]
  Range (min … max):   17.042 s … 17.638 s    10 runs

This is with the following patch:

---
 archive-tar.c | 63 +++++++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 59 insertions(+), 4 deletions(-)

diff --git a/archive-tar.c b/archive-tar.c
index 3e53aac1e6..c889b84c2c 100644
--- a/archive-tar.c
+++ b/archive-tar.c
@@ -38,11 +38,13 @@ static int write_tar_filter_archive(const struct archiver *ar,
 #define USTAR_MAX_MTIME 077777777777ULL
 #endif

+static int out_fd = 1;
+
 /* writes out the whole block, but only if it is full */
 static void write_if_needed(void)
 {
 	if (offset == BLOCKSIZE) {
-		write_or_die(1, block, BLOCKSIZE);
+		write_or_die(out_fd, block, BLOCKSIZE);
 		offset = 0;
 	}
 }
@@ -66,7 +68,7 @@ static void do_write_blocked(const void *data, unsigned long size)
 		write_if_needed();
 	}
 	while (size >= BLOCKSIZE) {
-		write_or_die(1, buf, BLOCKSIZE);
+		write_or_die(out_fd, buf, BLOCKSIZE);
 		size -= BLOCKSIZE;
 		buf += BLOCKSIZE;
 	}
@@ -101,10 +103,10 @@ static void write_trailer(void)
 {
 	int tail = BLOCKSIZE - offset;
 	memset(block + offset, 0, tail);
-	write_or_die(1, block, BLOCKSIZE);
+	write_or_die(out_fd, block, BLOCKSIZE);
 	if (tail < 2 * RECORDSIZE) {
 		memset(block, 0, offset);
-		write_or_die(1, block, BLOCKSIZE);
+		write_or_die(out_fd, block, BLOCKSIZE);
 	}
 }

@@ -434,6 +436,56 @@ static int write_tar_archive(const struct archiver *ar,
 	return err;
 }

+static int internal_gzip(int in, int out, void *data)
+{
+	int *levelp = data;
+	gzFile gzip = gzdopen(1, "wb");
+	if (!gzip)
+		die(_("gzdopen failed"));
+	if (gzsetparams(gzip, *levelp, Z_DEFAULT_STRATEGY) != Z_OK)
+		die(_("unable to set compression level"));
+
+	for (;;) {
+		char buf[BLOCKSIZE];
+		ssize_t read = xread(in, buf, sizeof(buf));
+		if (read < 0)
+			die_errno(_("read failed"));
+		if (read == 0)
+			break;
+		if (gzwrite(gzip, buf, read) != read)
+			die(_("gzwrite failed"));
+	}
+
+	if (gzclose(gzip) != Z_OK)
+		die(_("gzclose failed"));
+	close(in);
+	return 0;
+}
+
+static int write_tar_gzip_archive(const struct archiver *ar,
+				  struct archiver_args *args)
+{
+	struct async filter;
+	int r;
+
+	memset(&filter, 0, sizeof(filter));
+	filter.proc = internal_gzip;
+	filter.data = &args->compression_level;
+	filter.in = -1;
+
+	if (start_async(&filter))
+		die(_("unable to fork off internal gzip"));
+	out_fd = filter.in;
+
+	r = write_tar_archive(ar, args);
+
+	close(out_fd);
+	if (finish_async(&filter))
+		die(_("error in internal gzip"));
+
+	return r;
+}
+
 static int write_tar_filter_archive(const struct archiver *ar,
 				    struct archiver_args *args)
 {
@@ -445,6 +497,9 @@ static int write_tar_filter_archive(const struct archiver *ar,
 	if (!ar->data)
 		BUG("tar-filter archiver called with no filter defined");

+	if (!strcmp(ar->data, "gzip -cn"))
+		return write_tar_gzip_archive(ar, args);
+
 	strbuf_addstr(&cmd, ar->data);
 	if (args->compression_level >= 0)
 		strbuf_addf(&cmd, " -%d", args->compression_level);
--
2.21.0