* changes:
Always attempt delta compression when reuseDeltas is false
Avoid TemporaryBuffer.Heap on very small deltas
Correct distribution of allowed delta size along chain length
Split remaining delta work on path boundaries
Replace DeltaWindow array with circularly linked list
Micro-optimize copy instructions in DeltaEncoder
Micro-optimize DeltaWindow primary loop
Micro-optimize DeltaWindow maxMemory test to be != 0
Mark DeltaWindowEntry methods final
If reuseObjects=true but reuseDeltas=false the caller wants attempt
a delta for every object in the input list. Test for reuseDeltas
to ensure every object passes through the searchInWindow() method.
If no delta is possible for an object and it will be stored whole
(non-delta format), PackWriter may still reuse its content from any
source pack. This avoids an inflate()-deflate() cycle to recompress
the object contents.
Change-Id: I845caeded419ef4551ef1c85787dd5ffd73235d9
TemporaryBuffer is great when the output size is not known, but must
be bound by a relatively large upper limit that fits in memory, e.g.
64 KiB or 20 MiB. The buffer gracefully supports growing storage by
allocating 8 KiB blocks and storing them in an ArrayList.
In a Git repository many deltas are less than 8 KiB. Typical tree
objects are well below this threshold, and their deltas must be
encoded even smaller.
For these much smaller cases avoid the 8 KiB minimum allocation used
by TemporaryBuffer. Instead allocate a very small OutputStream
writing to an array that is sized at the limit.
Change-Id: Ie25c6d3a8cf4604e0f8cd9a3b5b701a592d6ffca
Nicolas Pitre discovered a very simple rule for selecting between two
different delta base candidates:
- if based whole object, must be <= 50% of target
- if at end of a chain, must be <= 1/depth * 50% of target
The rule penalizes deltas near the end of the chain, requiring them to
be very small in order to be kept by the packer. This favors deltas
that are based on a shorter chain, where the read-time unpack cost is
much lower. Fewer bytes need to be consulted from the source pack
file, and less copying is required in memory to rebuild the object.
Junio Hamano explained Nico's rule to me today, and this commit fixes
DeltaWindow to implement it as described.
When no base has been chosen the computation is simply the statements
denoted above. However once a base with depth of 9 has been chosen
(e.g. when pack.depth is limited to 10), a non-delta source may
create a new delta that is up to 10x larger than the already selected
base. This reflects the intent of Nico's size distribution rule no
matter what order objects are visited in the DeltaWindow.
With this patch and my other patches applied, repacking JGit with:
[pack]
reuseObjects = false
reuseDeltas = false
depth = 50
window = 250
threads = 4
compression = 9
CGit (all) 5,711,735 bytes; real 0m13.942s user 0m47.722s [1]
JGit heads 5,718,295 bytes; real 0m11.880s user 0m38.177s [2]
rest 9,809 bytes
The improved JGit result for the head pack is only 6.4 KiB larger than
CGit's resulting pack. This patch allowed JGit to find an additional
39.7 KiB worth of space savings. JGit now also often runs 2s faster
than CGit, despite also creating bitmaps and pruning objects after the
head pack creation.
[1] time git repack -a -d -F --window=250 --depth=50
[2] time java -Xmx128m -jar jgit debug-gc
Change-Id: I5caec31359bf7248cabdd2a3254c84d4ee3cd96b
When an idle thread tries to steal work from a sibling's remaining
toSearch queue, always try to split along a path boundary. This
avoids missing delta opportunities in the current window of the
thread whose work is being taken.
The search order is reversed to walk further down the chain from
current position, avoiding the risk of splitting the list within
the path the thread is currently processing.
When selecting which thread to split from use an accurate estimate
of the size to be taken. This avoids selecting a thread that has
only one path remaining but may contain more pending entries than
another thread with several paths remaining.
As there is now a race condition where the straggling thread can
start the next path before the split can finish, the stealWork()
loop spins until it is able to acquire a split or there is only
one path remaining in the siblings.
Change-Id: Ib11ff99f90a4d9efab24bf4a85342cc63203dba5
PackWriter generally chooses the order for objects when it builds the
object lists. This ordering already depends on history information to
guide placing more recent objects first and historical objects last.
Allow PackWriter to make the basic ordering decisions, instead of
trying to override them. The old approach of sorting the list caused
DfsReader to override any ordering change PackWriter might have tried
to make when repacking a repository.
This now better matches with WindowCursor's implementation, where
PackWriter solely determines the object ordering.
Change-Id: Ic17ab5631ec539f0758b962966c3a1823735b814
Typical window sizes are 10 and 250 (although others are accepted).
In either case the pointer overhead of 1 pointer in an array or
2 pointers for a double linked list is trivial. A doubly linked
list as used here for window=250 is only another 1024 bytes on a
32 bit machine, or 2048 bytes on a 64 bit machine.
The critical search loops scan through the array in either the
previous direction or the next direction until the cycle is finished,
or some other scan abort condition is reached. Loading the next
object's pointer from a field in the current object avoids the
branch required to test for wrapping around the edge of the array.
It also saves the array bounds check on each access.
When a delta is chosen the window is shuffled to hoist the currently
selected base as an earlier candidate for the next object. Moving
the window entry is easier in a double-linked list than sliding a
group of array entries.
Change-Id: I9ccf20c3362a78678aede0f0f2cda165e509adff
The copy instruction formatter should not to compute the shifts and
masks twice. Instead compute them once and assume there is a register
available to store the temporary "b" for compare with 0.
Change-Id: Ic7826f29dca67b16903d8f790bdf785eb478c10d
javac and the JIT are more likely to understand a boolean being
used as a branch conditional than comparing int against 0 and 1.
Rewrite NEXT_RES and NEXT_SRC constants to be booleans so the
code is clarified for the JIT.
Change-Id: I1bdd8b587a69572975a84609c779b9ebf877b85d
Instead of using a compare-with-0 use a does not equal 0.
javac bytecode has a special instruction for this, as it
is very common in software. We can assume the JIT knows
how to efficiently translate the opcode to machine code,
and processors can do != 0 very quickly.
Change-Id: Idb84c1d744d2874517fd4bfa1db390e2dbf64eac
This class and all of its methods are only package visible.
Clarify the methods as final for the benefit of the JIT to
inline trivial code.
Change-Id: I078841f9900dbf299fbe6abf2599f0208ae96856
* changes:
Increase PackOutputStream copy buffer to 64 KiB
Tighten object header writing in PackOutuptStream
Skip main thread test in ThreadSafeProgressMonitor
Declare members of PackOutputStream final
Always allocate the PackOutputStream copyBuffer
Disable CRC32 computation when no PackIndex will be created
Steal work from delta threads to rebalance CPU load
Colby just pointed out to me the buffer was 16 KiB. This may
be very small for common objects. Increase to 64 KiB.
Change-Id: Ideecc4720655a57673252f7adb8eebdf2fda230d
Most objects are written as OFS_DELTA with the base in the pack,
that is why this case comes first in writeHeader(). Rewrite the
condition to always examine this first and cache the PackWriter's
formatting flag for use of OFS_DELTA headers, in modern Git networks
this is true more often then it it is false.
Assume the cost of write() is high, especially due to entering the
MessageDigest to update the pack footer SHA-1 computation. Combine
the OFS_DELTA information as part of the header buffer so that the
entire burst is a single write call, rather than two relatively
small ones. Most OFS_DELTA headers are <= 6 bytes, so this rewrite
tranforms 2 writes of 3 bytes each into 1 write of ~6 bytes.
Try to simplify the objectHeader code to reduce branches and use
more local registers. This shouldn't really be necessary if the
compiler is well optimized, but it isn't very hard to clarify data
usage to either javac or the JIT, which may make it easier for the
JIT to produce better machine code for this method.
Change-Id: I2b12788ad6866076fabbf7fa11f8cce44e963f35
update(int) is only invoked from a worker thread, in JGit's case
this is DeltaTask. The Javadoc of TSPM suggests update should only
ever be used by a worker thread.
Skip the main thread check, saving some cycles on each run of the
progress monitor.
Change-Id: I6cb9382d71b4cb3f8e8981c7ac382da25304dfcb
These methods cannot be sanely overridden anywhere. Most methods
are package visible only, or are private. A few public methods do
exist but there is no useful way to override them since creation
of PackOutputStream is managed by PackWriter and cannot be delegated.
Change-Id: I12cd3326b78d497c1f9751014d04d1460b46e0b0
The getCopyBuffer() is almost always used during output. All known
implementations of ObjectReuseAsIs rely on the buffer to be present,
and the only sane way to get good performance from PackWriter is to
reuse objects during packing.
Avoid a branch and test when obtaining this buffer by making sure
it is always populated.
Change-Id: I200baa0bde5dcdd11bab7787291ad64535c9f7fb
If a server is streaming 3GiB worth of pack data to a client there
is no reason to compute the CRC32 checksum on the objects. The
CRC32 code computed by PackWriter is used only in the new index
created by writeIndex(), which is never invoked for the native Git
network protocols.
Object reuse may still compute its own CRC32 to verify the data
being copied from an existing pack has not been corrupted. This
check is done by the ObjectReader that implements ObjectReuseAsIs
and has no relationship to the CRC32 being skipped during output.
Change-Id: I05626f2e0d6ce19119b57d8a27193922636d60a7
If the configuration wants to run 4 threads the delta search work
is initially split somewhat evenly across the 4 threads. During
execution some threads will finish early due to the work not being
split fairly, as the initial partitions were based on object count
and not cost to inflate or size of DeltaIndex.
When a thread finishes early it now tries to take 50% of the work
remaining on a sibling thread, and executes that before exiting.
This repeats as each thread completes until a thread has only 1
object remaining.
Repacking Blink, Chromium's new fork of WebKit (2.2M objects 3.9G):
[pack]
reuseDeltas = false
reuseObjects = false
depth = 50
threads = 8
window = 250
windowMemory = 800m
before: ~105% CPU after 80%
after: >780% CPU to 100%
Change-Id: I65e45422edd96778aba4b6e5a0fd489ea48e8ca3
* changes:
Fix plugin provider names to conform with release train requirement
Add missing @since tags for new API methods
DfsReaderOptions are options for a DFS stored repository
According to release train requirements [1] the provider name for all
artifacts of Eclipse projects is "Eclipse <project name>".
[1] http://wiki.eclipse.org/Development_Resources/HOWTO/Release_Reviews#Checklist
Change-Id: I8445070d1d96896d378bfc49ed062a5e7e0f201f
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Some packs built by JGit have incredibly long delta chains due to a
long standing bug in PackWriter. Google has packs created by JGit's
DfsGarbageCollector with chains of 6000 objects long, or more.
Inflating objects at the end of this 6000 long chain is impossible
to complete within a reasonable time bound. It could take a beefy
system hours to perform even using the heavily optimized native C
implementation of Git, let alone with JGit.
Enable pack.cutDeltaChains to be set in a configuration file to
permit the PackWriter to determine the length of each delta chain
and clip the chain at arbitrary points to fit within pack.depth.
Delta chain cycles are still possible, but no attempt is made to
detect them. A trivial chain of A->B->A will iterate for the full
pack.depth configured limit (e.g. 50) and then pick an object to
store as non-delta.
When cutting chains the object list is walked in reverse to try
and take advantage of existing chain computations. The assumption
here is most deltas are near the end of the list, and their bases
are near the front of the list. Going up from the tail attempts to
reuse chainLength computations by relying on the memoized value in
the delta base.
The chainLength field in ObjectToPack is overloaded into the depth
field normally used by DeltaWindow. This is acceptable because the
chain cut happens before delta search, and the chainLength is reset
to 0 if delta search will follow.
Change-Id: Ida4fde9558f3abbbb77ade398d2af3941de9c812
This switch is called mostly for OBJ_TREE and OBJ_BLOB types, which
typically make up 66% of the objects in a repository. Simplify the
test for these common types by testing for the one bit they have in
common and returning early.
Object type 5 is currently undefined. In the old code it would hit
the default and return true. In the new code it will match the early
case and also return true. In either implementation 5 should never show
up as it is not a valid type known to Git.
Object type 6 OFS_DELTA is not permitted to be supplied here.
Object type 7 REF_DELTA is not permitted to be supplied here.
Change-Id: I0ede8acee928bb3e73c744450863942064864e9c
Now that WANT_WRITE is gone renumber the flags to move the unused
bit next to the type. Recluster AS_IS and DELTA_ATTEMPTED to be
next to each other since these bits are tested as a pair.
Change-Id: I42994b5ff1f67435e15c3f06d02e3b82141e8f08
Free up the WANT_WRITE flag in ObjectToPack by switching the test
to use the special offset value of 1. The Git pack file format
calls for the first 4 bytes to be 'PACK', which means any object
must start at an offset >= 4. Current versions require another 8
bytes in the header, placing the first object at offset = 12.
So offset = 1 is an invalid location for an object, and can be
used as a marker signal to indicate the writing loop has tried
to write the object, but recursed into the base first. When an
object is visited with offset == 1 it means there is a cycle in
the delta base path, and the cycle must be broken.
Change-Id: I2d05b9017c5f9bd9464b91d43e8d4b4a085e55bc
Garbage is randomly ordered and unlikely to delta compress against
other garbage. Disable delta compression allowing objects to switch
to whole form when moving to the garbage pack.
Because the garbage is not well compressed assume deltas were not
attempted during a normal GC cycle.
Override the reuse settings, garbage that can be reused should be
reused as-is into the garbage pack rather than switching something
like the compression level during a GC. It is intended that garbage
will eventually be removed from the repository so expending CPU
time on a compression switch is not worthwhile.
Change-Id: I0e8e58ee99e5011d375d3d89c94f2957de8402b9
This implementation has been proven to deadlock in production server
loads. Google has been running with it disabled for a quite a while,
as the bugs have been difficult to identify and fix.
Instead of suggesting it works and is useful, drop the code. JGit
should not advertise support for functionality that is known to
be broken.
In a few of the places where read-ahead was enabled by DfsReader
there is more information about what blocks should be loaded when.
During object representation selection, or size lookup, or sending
object as-is to a PackWriter, or sending an entire pack as-is the
reader knows exactly which blocks are required in the cache, and it
also can compute when those will be needed. The broken read-ahead
code was stupid and just read a fixed amount ahead of the current
offset, which can waste IOs if more precise data was available.
DFS systems are usually slow to respond so read-ahead is still
a desired feature, but it needs to be rebuilt from scratch and
make better use of the offset information.
Change-Id: Ibaed8288ec3340cf93eb269dc0f1f23ab5ab1aea
String.valueOf is an overloaded and the compiler unfortunately picks
the wrong one since null contains no type information.
Change-Id: Icd197eaa046421f3cfcc5bf3e7601dc5bc7486b6
Rewrite this complicated logic to examine each pack file exactly
once. This reduces thrashing when there are many large pack files
present and the reader needs to locate each object's header.
The intermediate temporary list is now smaller, it is bounded to
the same length as the input object list. In the prior version of
this code the list contained one entry for every representation of
every object being packed.
Only one representation object is allocated, reducing the overall
memory footprint to be approximately one reference per object found
in the current pack file (the pointer in the BlockList). This saves
considerable working set memory compared to the prior version that
made and held onto a new representation for every ObjectToPack.
Change-Id: I2c1f18cd6755643ac4c2cf1f23b5464ca9d91b22
Clip the configured limit to Integer.MAX_VALUE at the top of the
loop, saving a compare branch per object considered. This can cut
2M branches out of a repacking of the Linux kernel.
Rewrite the logic so the primary path is to match the conditional;
most objects are larger than BLKSZ (16 bytes) and less than limit.
This may help branch prediction on CPUs if the CPU tries to assume
execution takes the side of the branch and not the second.
Change-Id: I5133d1651640939afe9fbcfd8cfdb59965c57d5a
There is no reasonable way for a subclass to correctly override and
implement these methods. They depend on internal state that cannot
otherwise be managed.
Most of these methods are also in critical paths of PackWriter.
Declare them final so subclasses do not try to replace them,
and so the JIT knows the smaller ones can be safely inlined.
Change-Id: I9026938e5833ac0b94246d21c69a143a9224626c
None of these methods should ever be overridden at runtime by an
extension class. Given how small they are the JIT should perform
inlining where reasonable. Hint this is possible by marking all
methods final so its clear no replacement can be loaded later on.
Change-Id: Ia75a5d36c6bd25b24169e2bdfa360c8f52b669cd
This flag is never checked on its own. It is only checked as part
of a pair through the doNotAttemptDelta() method. Delete the method
so there is less confusion about the flag being used on its own.
Change-Id: Id7088caa649599f4f11d633412c2a2af0fd45dd8
This method is only invoked with true as the argument.
Remove the unnecessary parameter and branch, making
the code easier for the JIT to optimize.
Change-Id: I68a9cd82f197b7d00a524ea3354260a0828083c6
Added also tests and the associated option for the command line Merge
command.
Bug: 335091
Change-Id: Ie321c572284a6f64765a81674089fc408a10d059
Signed-off-by: Christian Halstrick <christian.halstrick@sap.com>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Robin Rosenberg
Shawn Pearce
Shawn Pearce
Matthias Sohn
Christian Halstrick
Tomasz Zarna