odb: fixed streamed decompression reader ( specific tests would still be missing ) and added performance tests which are extremely promising

3 files changed, 107 insertions, 10 deletions

diff --git a/lib/git/odb/db.py b/lib/git/odb/db.py
index 5c50a512..e656b2b5 100644
--- a/lib/git/odb/db.py
+++ b/lib/git/odb/db.py
@@ -91,8 +91,8 @@ class iObjectDBW(object):
 			they where given.
 		:param iter_info: iterable yielding tuples containing the type_string
 			size_in_bytes and the steam with the content data.
-		:param dry_run: see ``to_obj``
-		:param sha_as_hex: see ``to_obj``
+		:param dry_run: see ``to_object``
+		:param sha_as_hex: see ``to_object``
 		:param max_threads: if < 1, any number of threads may be started while processing
 			the request, otherwise the given number of threads will be started.
 		:raise IOError: if data could not be written"""
diff --git a/lib/git/odb/utils.py b/lib/git/odb/utils.py
index 8a054201..1e4a8e9d 100644
--- a/lib/git/odb/utils.py
+++ b/lib/git/odb/utils.py
@@ -39,7 +39,7 @@ write = os.write
 close = os.close
 
 # ZLIB configuration
-# used when compressing objects
+# used when compressing objects - 1 to 9 ( slowest )
 Z_BEST_SPEED = 1
 
 #} END Routines
@@ -70,7 +70,7 @@ class FDCompressedSha1Writer(object):
 		bytes_written = write(self.fd, cdata)
 		if bytes_written != len(cdata):
 			raise self.exc
-		return bytes_written
+		return len(data)
 
 	def sha(self, as_hex = False):
 		""":return: sha so far
@@ -175,7 +175,7 @@ class DecompressMemMapReader(object):
 				self._br += size
 				return dat
 			else:
-				dat = self._buf.getvalue()		# ouch, duplicates data
+				dat = self._buf.read()		# ouch, duplicates data
 				size -= self._buflen
 				self._br += self._buflen
 				
@@ -195,28 +195,34 @@ class DecompressMemMapReader(object):
 		# copied once, and another copy of a part of it when it creates the unconsumed
 		# tail. We have to use it to hand in the appropriate amount of bytes durin g
 		# the next read.
-		if self._zip.unconsumed_tail:
+		tail = self._zip.unconsumed_tail
+		if tail:
 			# move the window, make it as large as size demands. For code-clarity, 
 			# we just take the chunk from our map again instead of reusing the unconsumed
 			# tail. The latter one would safe some memory copying, but we could end up
 			# with not getting enough data uncompressed, so we had to sort that out as well.
 			# Now we just assume the worst case, hence the data is uncompressed and the window
 			# needs to be as large as the uncompressed bytes we want to read.
-			self._cws = self._cwe - len(self._zip.unconsumed_tail)
+			self._cws = self._cwe - len(tail)
 			self._cwe = self._cws + size
+			
+			
 			indata = self._m[self._cws:self._cwe]		# another copy ... :(
+			# get the actual window end to be sure we don't use it for computations
+			self._cwe = self._cws + len(indata) 
 		else:
 			cws = self._cws
 			self._cws = self._cwe
 			self._cwe = cws + size 
 			indata = self._m[self._cws:self._cwe]		# ... copy it again :(
 		# END handle tail
-		
+			
 		dcompdat = self._zip.decompress(indata, size)
-		self._br += len(dcompdat)
 		
+		self._br += len(dcompdat)
 		if dat:
-			return dat + dcompdat
+			dcompdat = dat + dcompdat
+			
 		return dcompdat
 		
 #} END classes 
diff --git a/test/git/performance/test_streams.py b/test/git/performance/test_streams.py
new file mode 100644
index 00000000..15924c08
--- /dev/null
+++ b/test/git/performance/test_streams.py
@@ -0,0 +1,91 @@
+"""Performance data streaming performance"""
+
+from test.testlib import *
+from git.odb.db import *
+
+from array import array
+from cStringIO import StringIO
+from time import time
+import os
+import sys
+import stat
+import random
+
+
+from lib import (
+	TestBigRepoReadOnly
+	)
+
+
+
+def make_memory_file(size_in_bytes, randomize=False):
+	""":return: tuple(size_of_stream, stream)
+	:param randomize: try to produce a very random stream"""
+	actual_size = size_in_bytes / 4
+	producer = xrange(actual_size)
+	if randomize:
+		producer = list(producer)
+		random.shuffle(producer)
+	# END randomize
+	a = array('i', producer)
+	return actual_size*4, StringIO(a.tostring())
+
+
+class TestObjDBPerformance(TestBigRepoReadOnly):
+	
+	large_data_size_bytes = 1000*1000*10		# some MiB should do it
+	moderate_data_size_bytes = 1000*1000*1		# just 1 MiB
+	
+	@with_bare_rw_repo
+	def test_large_data_streaming(self, rwrepo):
+		ldb = LooseObjectDB(os.path.join(rwrepo.git_dir, 'objects'))
+		
+		for randomize in range(2):
+			desc = (randomize and 'random ') or ''
+			print >> sys.stderr, "Creating %s data ..." % desc
+			st = time()
+			size, stream = make_memory_file(self.large_data_size_bytes, randomize)
+			elapsed = time() - st
+			print >> sys.stderr, "Done (in %f s)" % elapsed
+			
+			# writing - due to the compression it will seem faster than it is 
+			st = time()
+			sha = ldb.to_object('blob', size, stream)
+			elapsed = time() - st
+			assert ldb.has_object(sha)
+			fsize_kib = os.path.getsize(ldb.readable_db_object_path(sha)) / 1000
+			
+			
+			size_kib = size / 1000
+			print >> sys.stderr, "Added %i KiB (filesize = %i KiB) of %s data to loose odb in %f s ( %f Write KiB / s)" % (size_kib, fsize_kib, desc, elapsed, size_kib / elapsed)
+			
+			# reading all at once
+			st = time()
+			type, size, shastream = ldb.object(sha)
+			shadata = shastream.read()
+			elapsed = time() - st
+			
+			stream.seek(0)
+			assert shadata == stream.getvalue()
+			print >> sys.stderr, "Read %i KiB of %s data at once from loose odb in %f s ( %f Read KiB / s)" % (size_kib, desc, elapsed, size_kib / elapsed)
+			
+			
+			# reading in chunks of 1 MiB
+			cs = 512*1000
+			chunks = list()
+			st = time()
+			type, size, shastream = ldb.object(sha)
+			while True:
+				data = shastream.read(cs)
+				chunks.append(data)
+				if len(data) < cs:
+					break
+			# END read in chunks
+			elapsed = time() - st
+			
+			stream.seek(0)
+			assert ''.join(chunks) == stream.getvalue()
+			
+			cs_kib = cs / 1000
+			print >> sys.stderr, "Read %i KiB of %s data in %i KiB chunks from loose odb in %f s ( %f Read KiB / s)" % (size_kib, desc, cs_kib, elapsed, size_kib / elapsed)
+		# END for each randomization factor