hash-object: don't pointlessly zlib compress without -w
PATCH 1/1 hash-object: don't pointlessly zlib compress without -w
When hash-object hashes something the size of core.bigFileThreshold or larger (512MB by default) it'll be streamed through stream_to_pack().
That added in 568508e765 ("bulk-checkin: replace fast-import based implementation", 2011-10-28) would compress the file with zlib, but was oblivious as to whether the content would actually be written out to disk, which isn't the case unless hash-object is called with the "-w" option.
Hashing is much slower if we need to compress the content, so let's check if the HASH_WRITE_OBJECT flag has been given.
An accompanying perf test shows how much this improves things. With CFLAGS=-O3 and OPENSSL_SHA1=Y the relevant change is (manually reformatted to avoid long lines):
1007.6: 'git hash-object <file>' with threshold=32M
-> 1.57(1.55+0.01) 0.09(0.09+0.00) -94.3%
1007.7: 'git hash-object --stdin < <file>' with threshold=32M
-> 1.57(1.57+0.00) 0.09(0.07+0.01) -94.3%
1007.8: 'echo <file> | git hash-object --stdin-paths' threshold=32M
-> 1.59(1.56+0.00) 0.09(0.08+0.00) -94.3%The same tests using "-w" still take that long, since those will need to zlib compress the relevant object. With the sha1collisiondetection library (our default) there's less of a difference since the hashing itself is slower, or respectively:
1.71(1.65+0.01) 0.19(0.18+0.01) -88.9%
1.70(1.66+0.02) 0.19(0.19+0.00) -88.8%
1.69(1.66+0.00) 0.19(0.18+0.00) -88.8%--- bulk-checkin.c | 3 ++- t/perf/p1007-hash-object.sh | 53 +++++++++++++++++++++++++++++++++++++ 2 files changed, 55 insertions(+), 1 deletion(-) create mode 100755 t/perf/p1007-hash-object.sh
bulk-checkin.c +2 -1
@@ -105,8 +105,9 @@ static int stream_to_pack(struct bulk_checkin_state *state,int status =Z_OK ;int write_object = (flags &HASH_WRITE_OBJECT );off_t offset =0 ; +int level = write_object ? pack_compression_level :Z_NO_COMPRESSION ; -git_deflate_init (&s, pack_compression_level); +git_deflate_init (&s, level); hdrlen =encode_in_pack_object_header (obuf,sizeof (obuf), type, size); s.next_out = obuf + hdrlen;
t/perf/p1007-hash-object.sh new file +53 -0
53 lines not on this page · raw message
Signature · 1 line
2.21.0.1020.gf2820cf01a
Quoted 5 lines · Ævar Arnfjörð Bjarmason
That added in
568508e765("bulk-checkin: replace fast-import based implementation", 2011-10-28) would compress the file with zlib, but was oblivious as to whether the content would actually be written out to disk, which isn't the case unless hash-object is called with the "-w" option.
Yeah, this seems like an obvious and easy improvement. I'd guess that not many people use hash-object without "-w", so nobody ever really cared much.
bulk-checkin.c static int stream_to_pack(struct bulk_checkin_state *state, · line 110
Quoted 12 lines
bulk-checkin.c +2
-1@@ -105,8 +105,9 @@ static int stream_to_pack(struct bulk_checkin_state *state,int status =Z_OK ;int write_object = (flags &HASH_WRITE_OBJECT );off_t offset =0 ; +int level = write_object ? pack_compression_level :Z_NO_COMPRESSION ; -git_deflate_init (&s, pack_compression_level); +git_deflate_init (&s, level);
Hmm. It really seems like if we're not writing, that we could avoid this whole pack-encoding business altogether, and just stream the contents through the hash, no matter the size. E.g., we're paying extra here to compute a pack crc32 that nobody cares about, not to mention the zlib crc.
That said, this is just not that common a case, and you've regained the vast majority of the performance with this simple change. So it may not be worth pursuing further.
bulk-checkin.c static int stream_to_pack(struct bulk_checkin_state *state,
Quoted 8 lines
bulk-checkin.c +8
-0+test_lazy_prereq SHA1SUM_AND_SANE_DD_AND_URANDOM ' + >empty && + sha1sum empty >empty.sha1sum && + grep -q -w da39a3ee5e6b4b0d3255bfef95601890afd80709 empty.sha1sum && + dd if=/dev/urandom of=random.test bs=1024 count=1 && + stat -c %s random.test >random.size && + grep -q -x 1024 random.size +'
I don't think this needs to be cryptographically random. You could just use "test-tool genrandom" instead, which gets rid of the need for dd and /dev/urandom (and as a bonus, is more deterministic since it's a seeded LCG).
bulk-checkin.c
Quoted 3 lines
bulk-checkin.c +3
-0+test_perf 'sha1sum(1) on file.random (for comparison)' ' + sha1sum file.random +'
I'm not sure this is worth the prereq it entails. Who really cares how fast sha1sum is? What we generally care about in the perf suite is how Git speeds compare. Primarily between versions (to show off improvements, or find regressions), but sometimes between two builds (but sadly the perf output is not good at comparing those numbers automatically).
I know this sets a baseline, but it feels a bit like apples to oranges.
bulk-checkin.c
Quoted 5 lines
bulk-checkin.c +5
-0+for threshold in 32M 64M +do + for write in '' ' -w' + do + for literally in ' --literally -t commit' ''
Is this "-t commit" worth it? I guess it proves that --literally is being used, since we'd choke otherwise. But for a perf test, it seems like you'd want to change as few variables as possible.
bulk-checkin.c
Quoted 1 line
bulk-checkin.c +1
-0+ test_perf "'git hash-object$write$literally --stdin < <file>' with threshold=$threshold" "
This test title violates our shell style (whitespace after "<"). ;)
(Yes, I know it was probably to avoid "<<").
bulk-checkin.c
Quoted 2 lines
bulk-checkin.c +2
-0+ test_perf "'echo <file> | git hash-object$write$literally --stdin-paths' threshold=$threshold" " + echo file.random | git -c core.bigFileThreshold=$threshold hash-object$write$literally --stdin-paths
You have to bunch up the option variables in the title to get the whitespace right. But in the command itself, you can do:
hash-object $write $literally ...
and the shell's whitespace handling will make it work. Maybe worth it to make the source a bit more readable.
-Peff