When marking commits as uninteresting don't care if the tree exists
When during an ObjectWalk commits are marked as uninteresting we should
be tolerant against the situation that the commit exists in the repo but
the referenced tree is not exisiting. Since commit
c4797fe986 we are throwing
MissingObjectException in such a case. This semantic differs from native
git behaviour and may cause push operations to fail while they would
work in native git. See:
http://dev.eclipse.org/mhonarc/lists/egit-dev/msg03585.html
Bug: 445744
Change-Id: Ib7dec10fd2ef1adbb8adbabb9d3d5a64e554286a
Let ObjectWalk.markUninteresting also mark the root tree as
uninteresting
Using the ObjectWalk and marking a commit as uninteresting didn't mark
its root tree as uninteresting. This caused the "missing tree ..."
error in Gerrit under special circumstances. For example, if the
patch-set 2 changes only the commit message then the patch-set 1
and patch-set 2 share the same root-tree:
ps1 -> o o <- ps2
\ /
o root-tree
The transported pack will contain the ps2 commit but not the root-tree
object.
When using the BaseReceivePack.setCheckReferencedObjectsAreReachable
JGit will check the reachability of all referenced objects not provided
in the transported pack. Since the ps1 was advertised it will properly
be marked as uninteresting. However, the root-tree was reachable because
the ObjectWalk.markUninteresting missed to mark it as uninteresting.
JGit was then rejecting the pack with the "missing tree ..." exception.
Gerrit-issue: https://code.google.com/p/gerrit/issues/detail?id=1582
Change-Id: Iff2de8810f14ca304e6655fc8debeb8f3e20712b
Signed-off-by: Saša Živkov <sasa.zivkov@sap.com>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Rename RewriteTreeFilter to TreeRevFilter and make it public
The current behavior of passing a TreeFilter to RevWalk has limited
usefulness, since the RevFilter derived from the TreeFilter is always
ANDed together with any other RevFilters. It is also tied fairly
tightly to the parent rewriting mechanism.
Make TreeRevFilter a generic RevFilter that matches modified paths
against any TreeFilter. This allows for more complex logic like
(modified this path OR authored by this person).
Leave the rewrite flag logic in this class, since it's closely tied to
the parent comparison code, but hidden behind a protected constructor.
Change-Id: Ia72ef591a99415e6f340c5f64583a49c91f1b82f
Previously, setting any TreeFilter on a RevWalk triggered parent
rewriting, which in the current StartGenerator implementation ends up
buffering the entire commit history in memory. Aside from causing poor
performance on large histories, this does not match the default
behavior of `git rev-list`, which does not rewrite parent SHAs unless
asked to via --parents/--children.
Add a new method setRewriteParents() to RevWalk to disable this
behavior. Continue rewriting parents by default to maintain backwards
compatibility.
Change-Id: I1f38e05526071c75ca58095e312663de5e6f334d
Under certain circumstances isMergedInto() returned
false even though base is reachable from the tip.
This will hinder pushes and receives by falsely
detecting "non fast forward" merges.
o---o---o---o---o
/ \
/ o---o---A---o---M
/ /
---2---1-
if M (tip) was compared to 1 (base), the method
isMergedInto() could still return false, since
two mergeBases will be detected and the return
statement will only look at one of them:
return next() == base;
In most cases this would pass, but if "A" is
a commit with an old timestamp, the Generator
would walk down to "2" before completing the
walk pass "A" and first finding the other
merge base "1". In this case, the first call to
next() returns 2, which compared to base evaluates
as false.
This is fixed by iterating merge bases and
returning true if base is found among them.
Change-Id: If2ee1f4270f5ea4bee73ecb0e9c933f8234818da
Signed-off-by: Gustaf Lundh <gustaf.lundh@sonymobile.com>
Signed-off-by: Sven Selberg <sven.selberg@sonymobile.com>
Fix StackOverflowError in RevCommit.carryFlags on deep side graphs
Copying flags through a graph with deep side branches can cause
StackOverflowError. The recursive step to visit the 2nd parent of
a merge commit can overflow the stack if these are themselves very
deep histories with many branches.
Rewrite the loop to iterate up to 500 recursive steps deep before
unwinding the stack and running the remaining parts of the graph
using a dynamically allocated FIFORevQueue.
This approach still allows simple graphs that mostly merge short
lived topic branches into master to copy flags with no dynamic
memory allocation, relying only on temporary stack extensions.
Larger more complex graphs only pay the allocation penalities
if copying has to extend outwards "very far" in the graph, which
is unlikely for many coloring based algorithms.
Change-Id: I1882e6832c916e27dd5f6b7602d9caf66fb39c84
Fix RevWalkUtils.findBranchesReachableFrom not finding some branches
The "cut off" optimization causes it to not include branches that
contain the specified commit but happen to share commits with a branch
that does not contain the commit.
An example:
-B foo
\
-A---C master
findBranchesReachableFrom for commit A with both branches as input may
not return master (depending on the order of the input). The reason is
that A is not contained in foo, and therefore the old code would put B
in the cutOff set. When then walking the master commits and B is
checked, it is found in the cutOff set and the walk is aborted, causing
master not to be returned even though it should.
Bug: 425674
Change-Id: I2c0c406ce5fcc9a03538b483473af930d4895d30
Signed-off-by: Robin Stocker <robin@nibor.org>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
In certain cases a JGit server updating an existing shallow client
selected a common ancestor that was behind the shallow edge of
the client. This allowed the server to assume the client had some
objects it did not have and allowed creation of pack deltas the
client could never inflate.
Any commit the client has advertised as shallow must be treated
by UploadPack server as though it has no parents. With no parents
the walker cannot visit graph history the client does not have,
and PackWriter cannot consider delta base candidates the client
is lacking.
Change-Id: I4922b9354df9f490966a586fb693762e897345a2
The comment about legacy Tag and Object types no longer applies,
though prior to Idb273d5a92849b42935ac14eed73b796b80aad50 the field
was still being used by RewriteTreeFilter.
Change-Id: I9ee5da8f8a3b61c9cf543817c03117ee0609dd8f
The various rename detection options are an inherent part of the
filter, similar to the path being followed.
This fixes a potential NPE when a RevWalk with a FollowFilter is
created without a Repository, since the old code path tried to get
the DiffConfig from the RevWalk's possibly-missing repository.
Change-Id: Idb273d5a92849b42935ac14eed73b796b80aad50
Support creating pack bitmap indexes in PackWriter.
Update the PackWriter to support writing out pack bitmap indexes,
a parallel ".bitmap" file to the ".pack" file.
Bitmaps are selected at commits every 1 to 5,000 commits for
each unique path from the start. The most recent 100 commits are
all bitmapped. The next 19,000 commits have a bitmaps every 100
commits. The remaining commits have a bitmap every 5,000 commits.
Commits with more than 1 parent are prefered over ones
with 1 or less. Furthermore, previously computed bitmaps are reused,
if the previous entry had the reuse flag set, which is set when the
bitmap was placed at the max allowed distance.
Bitmaps are used to speed up the counting phase when packing, for
requests that are not shallow. The PackWriterBitmapWalker uses
a RevFilter to proactively mark commits with RevFlag.SEEN, when
they appear in a bitmap. The walker produces the full closure
of reachable ObjectIds, given the collection of starting ObjectIds.
For fetch request, two ObjectWalks are executed to compute the
ObjectIds reachable from the haves and from the wants. The
ObjectIds needed to be written are determined by taking all the
resulting wants AND NOT the haves.
For clone requests, we get cached pack support for "free" since
it is possible to determine if all of the ObjectIds in a pack file
are included in the resulting list of ObjectIds to write.
On my machine, the best times for clones and fetches of the linux
kernel repository (with about 2.6M objects and 300K commits) are
tabulated below:
Operation Index V2 Index VE003
Clone 37530ms (524.06 MiB) 82ms (524.06 MiB)
Fetch (1 commit back) 75ms 107ms
Fetch (10 commits back) 456ms (269.51 KiB) 341ms (265.19 KiB)
Fetch (100 commits back) 449ms (269.91 KiB) 337ms (267.28 KiB)
Fetch (1000 commits back) 2229ms ( 14.75 MiB) 189ms ( 14.42 MiB)
Fetch (10000 commits back) 2177ms ( 16.30 MiB) 254ms ( 15.88 MiB)
Fetch (100000 commits back) 14340ms (185.83 MiB) 1655ms (189.39 MiB)
Change-Id: Icdb0cdd66ff168917fb9ef17b96093990cc6a98d
When a lot of commits are added to DateRevQueue, the
sort-on-insertion approach is very heavy on CPU cycles.
One approach to fix this was made by Dave Borowitz:
https://git.eclipse.org/r/#/c/5491/
But using Java's PriorityQueue seems to have brought some
extra overhead, and the desired performance could not be
reached.
This fix takes another approach to the insertion problem,
without changing the expected behaviour or bringing extra
memory overhead:
If we detect over 1000 commits in the DateRevQueue, a
"seek-index" is rebuilt every 1000th added commit.
The index keeps track of every 100th commit in the
DateRevQueue. During insertions, it will be used for a
preliminary scanning (binary search) of the queue, with
the intention of helping add() find a good starting point
to start walking from. After finding this starting point,
add() will step commit-by-commit until the correct
insertion place in the queue is found (today, the queue
is expected to be sorted at all times).
When applied to repositories with many refs, this approach
has proven to bring huge performance gains and scales quite
well.
For instance, in a repository with close to 80000 refs,
we could cut down the time a typical Gerrit replication
of 1 commit would take (just a push from JGit's point of
view) from 32sec down to 3.5sec.
Below you see some typical times to add a specific amount
of commits (with random commit times) to the DateRevQueue
and the difference the preliminary seek-index makes:
Commits | Index | No Index
1024 8ms 8ms
2048 13ms 9ms
4096 5ms 59ms
8192 11ms 595ms
16384 22ms 3058ms
32768 64ms 13811ms
65536 201ms 62677ms
131072 783ms 331585ms
Only one extra reference is needed for every 100 inserted
commits (and only when we see more than 1000 commits in
the queue), so the memory overhead should be negligible.
Various index-stepping values were tested, and 100 seemed to
scale very well and be effective from start.
In the future, it should probably be dynamic and based on
the number of refs in the queue, but this should serve well
as a starting point.
Note: While other fundamentally different data structures may
be more suitable, the DateRevQueue is extremely central to
many of the Git core operations. This approach was chosen,
since the effect of the patch is easy to predict in conjuction
with the current implementation. A totally new data structure
will make it harder to predict behaviour in many common and
uncommon cases (in terms of breaking ties, memory usage, cost
when using few elements, object creation/disposing overhead,
etc).
Change-Id: Ie7b99f40eacf6324bfb4716d82073adeda64d10f
A few classes such as Constanrs are marked with @SuppressWarnings, as are
toString() methods with many liternal, but otherwise $NLS-n$ is used for
string containing text that should not be translated. A few literals may
fall into the gray zone, but mostly I've tried to only tag the obvious
ones.
Change-Id: I22e50a77e2bf9e0b842a66bdf674e8fa1692f590
StartGenerator now processes .git/shallow to have the
RevWalk stop for shallow commits.
See RevWalkShallowTest for tests.
Bug: 394543
CQ: 6908
Change-Id: Ia5af1dab3fe9c7888f44eeecab1e1bcf2e8e48fe
Signed-off-by: Chris Aniszczyk <zx@twitter.com>
These appear as descriptions in the index, see here (currently empty):
http://download.eclipse.org/jgit/docs/latest/apidocs/
Change-Id: If7996deef30ae688bade8b3ad6b19547ca3d8b50
Signed-off-by: Chris Aniszczyk <zx@twitter.com>
Refactored method to find branches from which a commit is reachable
The method uses some heuristics to obtain much better performance
than isMergeBase.
Since I wrote the relevant code in the method I approve the license
change from EPL to EDL implied by the move.
Change-Id: Ic4a7584811a2b0bf24e4f6b3eab2a7c022eabee8
Signed-off-by: Robin Rosenberg <robin.rosenberg@dewire.com>
This is needed to allow jumping to a selected commit when loading
history incrementally.
Change-Id: Id3b97d88d3b4b2d67561b11f8810cb88fe040823
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
This is needed to allow jumping to a selected commit when loading
history incrementally.
Change-Id: Id3b97d88d3b4b2d67561b11f8810cb88fe040823
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Use Integer, Character, and Long valueOf methods when
passing parameters to MessageFormat and other places
that expect objects instead of primitives
Change-Id: I5942fbdbca6a378136c00d951ce61167f2366ca4
It returns the number of commits that are in start and not in end.
Useful for calculating how much a branch is ahead of another one.
Change-Id: I09f7d9b049beea417da7ff32c9f8bf0d4ed46a7f
Signed-off-by: Robin Stocker <robin@nibor.org>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Allow detecting which files were renamed during a revwalk
The egit history view shows the files associated with a commit by using
a PathFilter. When following renames with a FollowFilter, the PathFilter
cannot be configured anymore because the affected files are simply not
known.
Thus, it should be possible to get to know which files are renamed.
Bug: 302549
Change-Id: I4761e9f5cfb4f0ef0b0e1e38991401a1d5003bea
This implements the server side of shallow clones only (i.e.
git-upload-pack), not the client side.
CQ: 5517
Bug: 301627
Change-Id: Ied5f501f9c8d1fe90ab2ba44fac5fa67ed0035a4
Signed-off-by: Chris Aniszczyk <caniszczyk@gmail.com>
The "Counting objects" phase of packing is the most time consuming
part for any server providing access to Git repositories. Scanning
through the entire project history, including every revision of
every tree that has ever existed is expensive and takes an incredible
amount of CPU time.
Inline the tree parsing logic, unroll a number of loops, and setup
to better handle the common case of seeing another occurrence of
an object that was already marked SEEN.
This change boosts the "Counting objects" phase when JGit is acting
as a server and is packing the linux-2.6 repository for its client.
Compared to CGit on the same hardware, a JGit daemon server is now
21883 objects/sec faster:
CGit:
Counted 2058062 objects in 38981 ms at 52796.54 objects/sec
Counted 2058062 objects in 38920 ms at 52879.29 objects/sec
Counted 2058062 objects in 39059 ms at 52691.11 objects/sec
JGit (before):
Counted 2058062 objects in 31529 ms at 65275.21 objects/sec
Counted 2058062 objects in 30359 ms at 67790.84 objects/sec
Counted 2058062 objects in 30033 ms at 68526.69 objects/sec
JGit (this commit):
Counted 2058062 objects in 28726 ms at 71644.57 objects/sec
Counted 2058062 objects in 27652 ms at 74427.24 objects/sec
Counted 2058062 objects in 27528 ms at 74762.50 objects/sec
Above the first run was a "cold server". For JGit the JVM had just
started up with `jgit daemon`, and for CGit we hadn't touched the
repository "recently" (but it was certainly in kernel buffer cache).
The second and third runs were against the running JGit JVM, allowing
timing tests to better reflect the benefits of JGit's pack and index
caching, as well as any optimizations the JIT may have performed.
The timings are fair. CGit is opening, checking and mmap'ing both
the pack and index during the timer. JGit is opening, checking
and malloc+read'ing the pack and index data into its Java heap
during the timer. Both processes are walking the same graph space,
and are computing the "path hash" necessary to sort objects in the
object table for delta compression. Since this commit only impacts
the "Counting objects" phase, delta compression was obviously not
included in the timings and JGit may still be performing delta
compression slower than CGit, resulting in an overall slower server
experience for clients.
Change-Id: Ieb184bfaed8475d6960a494b1f3c870e0382164a
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
ObjectIdOwnerMap: More lightweight map for ObjectIds
OwnerMap is about 200 ms faster than SubclassMap, more friendly to the
GC, and uses less storage: testing the "Counting objects" part of
PackWriter on 1886362 objects:
ObjectIdSubclassMap:
load factor 50%
table: 4194304 (wasted 2307942)
ms spent 36998 36009 34795 34703 34941 35070 34284 34511 34638 34256
ms avg 34800 (last 9 runs)
ObjectIdOwnerMap:
load factor 100%
table: 2097152 (wasted 210790)
directory: 1024
ms spent 36842 35112 34922 34703 34580 34782 34165 34662 34314 34140
ms avg 34597 (last 9 runs)
The major difference with OwnerMap is entries must extend from
ObjectIdOwnerMap.Entry, where the OwnerMap has injected its own
private "next" field into each object. This allows the OwnerMap to use
a singly linked list for chaining collisions within a bucket. By
putting collisions in a linked list, we gain the entire table back for
the SHA-1 bits to index their own "private" slot.
Unfortunately this means that each object can appear in at most ONE
OwnerMap, as there is only one "next" field within the object instance
to thread into the map. For types that are very object map heavy like
RevWalk (entity RevObject) and PackWriter (entity ObjectToPack) this
is sufficient, these entity types are only put into one map by their
container. By introducing a new map type, we don't break existing
applications that might be trying to use ObjectIdSubclassMap to track
RevCommits they obtained from a RevWalk.
The OwnerMap uses less memory. Each object uses 1 reference more (so
we're up 1,886,362 references), but the table is 1/2 the size (2^20
rather than 2^21). The table itself wastes only 210,790 slots, rather
than 2,307,942. So OwnerMap is wasting 200k fewer references.
OwnerMap is more friendly to the GC, because it hardly ever generates
garbage. As the map reaches its 100% load factor target, it doubles in
size by allocating additional segment arrays of 2048 entries. (So the
first grow allocates 1 segment, second 2 segments, third 4 segments,
etc.) These segments are hooked into the pre-allocated directory of
1024 spaces. This permits the map to grow to 2 million objects before
the directory itself has to grow. By using segments of 2048 entries,
we are asking the GC to acquire 8,204 bytes in a 32 bit JVM. This is
easier to satisfy then 2,307,942 bytes (for the 512k table that is
just an intermediate step in the SubclassMap). By reusing the
previously allocated segments (they are re-hashed in-place) we don't
release any memory during a table grow.
When the directory grows, it does so by discarding the old one and
using one that is 4x larger (so the directory goes to 4096 entries on
its first grow). A directory of size 4096 can handle up to 8 millon
objects. The second directory grow (16384) goes to 33 million objects.
At that point we're starting to really push the limits of the JVM
heap, but at least its many small arrays. Previously SubclassMap would
need a table of 67108864 entries to handle that object count, which
needs a single contiguous allocation of 256 MiB. That's hard to come
by in a 32 bit JVM. Instead OwnerMap uses 8192 arrays of about 8 KiB
each. This is much easier to fit into a fragmented heap.
Change-Id: Ia4acf5cfbf7e9b71bc7faa0db9060f6a969c0c50
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
In bc1af8459e ("RevWalk: Don't reset ObjectReader when stopping") we
stopped releasing the reader when the current log traversal is over.
This should have also been applied to the merge base logic that is
buried within MergeGenerator, but got missed.
Change-Id: I8328f43f02cba06fd545e22134872e781b9d4d36
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
RevWalk: Avoid unnecessary re-parsing of commit bodies
If the RevFilter doesn't actually require the commit body,
we shouldn't reparse it if the body was disposed. This happens
often inside of UploadPack during common ancestor negotation, the
RevWalk is reset and re-run over roughly the same commit space,
but the bodies are discarded because the commit message is not
relevant to the process.
Change-Id: I87b6b6a5fb269669867047698abf718d366bd002
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Applications like UploadPack reset() and reuse the same RevWalk
multiple times in very rapid succession. Releasing the ObjectReader's
internal state on each use, only to allocate it again on the next
cycle kills performance if the ObjectReader has internal caches, or
even if the Inflater gets returned and pulled from the InflaterCache
too frequently.
Making releasing the ObjectReader the application's responsibility
when it is done with the RevWalk, which most already do by wrapping
their loop in a try/finally block.
Change-Id: I3ad188a719e8d7f6bf27d1a7ca16d465534713f4
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
When UploadPack has computed the merge base between the client's have
set and the want set, its already loaded and parsed all of the
interesting commits that PackWriter needs to transmit to the client.
Switching the RevWalk and its object pool over to be an ObjectWalk
saves PackWriter from needing to re-parse these same commits from the
ObjectDatabase, reducing the startup latency for the enumeration
phase of packing.
UploadPack doesn't want to use an ObjectWalk for the okToGiveUp()
tests because its slower, during each commit popped it needs to cache
the tree into the pendingObjects list, and during each reset() it
discards a bunch of ObjectWalk specific state and reallocates some
internal collections. ObjectWalk was never meant to be rapidly
reset() like UploadPack does, so its perhaps somewhat cleaner to allow
"upgrading" a RevWalk to an ObjectWalk.
Bug: 301639
Change-Id: I97ef52a0b79d78229c272880aedb7f74d0f7532f
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
UploadPack: Avoid walking the entire project history
If the client presents a common commit on a side branch, and there is
a want for a disconnected branch UploadPack was walking back on the
entire history of the disconnected branch because it never would find
the common commit.
Limit our search back along any given want to be no earlier than the
oldest common commit received via a "have" line from our client. This
prevents us from looking at all of the project history.
Bug: 301639
Change-Id: Iffaaa2250907150d6efa1cf2f2fcf59851d5267d
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Chris Aniszczyk <caniszczyk@gmail.com>
There is no point in pushing all of the files within the edge
commits into the delta search when making a thin pack. This floods
the delta search window with objects that are unlikely to be useful
bases for the objects that will be written out, resulting in lower
data compression and higher transfer sizes.
Instead observe the path of a tree or blob that is being pushed
into the outgoing set, and use that path to locate up to WINDOW
ancestor versions from the edge commits. Push only those objects
into the edgeObjects set, reducing the number of objects seen by the
search window. This allows PackWriter to only look at ancestors
for the modified files, rather than all files in the project.
Limiting the search to WINDOW size makes sense, because more than
WINDOW edge objects will just skip through the window search as
none of them need to be delta compressed.
To further improve compression, sort edge objects into the front
of the window list, rather than randomly throughout. This puts
non-edges later in the window and gives them a better chance at
finding their base, since they search backwards through the window.
These changes make a significant difference in the thin-pack:
Before:
remote: Counting objects: 144190, done
remote: Finding sources: 100% (50275/50275)
remote: Getting sizes: 100% (101405/101405)
remote: Compressing objects: 100% (7587/7587)
Receiving objects: 100% (50275/50275), 24.67 MiB | 9.90 MiB/s, done.
Resolving deltas: 100% (40339/40339), completed with 2218 local objects.
real 0m30.267s
After:
remote: Counting objects: 61549, done
remote: Finding sources: 100% (50275/50275)
remote: Getting sizes: 100% (18862/18862)
remote: Compressing objects: 100% (7588/7588)
Receiving objects: 100% (50275/50275), 11.04 MiB | 3.51 MiB/s, done.
Resolving deltas: 100% (43160/43160), completed with 5014 local objects.
real 0m22.170s
The resulting pack is 13.63 MiB smaller, even though it contains the
same exact objects. 82,543 fewer objects had to have their sizes
looked up, which saved about 8s of server CPU time. 2,796 more
objects from the client were used as part of the base object set,
which contributed to the smaller transfer size.
Change-Id: Id01271950432c6960897495b09deab70e33993a9
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Sigend-off-by: Chris Aniszczyk <caniszczyk@gmail.com>
Non-commits are added to a pending queue, but duplicates are
removed by checking a flag. During a reset that flag must be
stripped off the old roots, otherwise the caller cannot reuse
the old roots after the reset.
RevWalk already does this correctly for commits, but ObjectWalk
failed to handle the non-commit case itself.
Change-Id: I99e1832bf204eac5a424fdb04f327792e8cded4a
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Revert "Teach PackWriter how to reuse an existing object list"
This reverts commit f5fe2dca3c.
I regret adding this feature to the public API. Caches aren't always
the best idea, as they require work to maintain. Here the cache is
redundant information that must be computed, and when it grows stale
must be removed. The redundant information takes up more disk space,
about the same size as the pack-*.idx files are. For the linux-2.6
repository, that's more than 40 MB for a 400 MB repository. So the
cache is a 10% increase in disk usage.
The entire point of this cache is to improve PackWriter performance,
and only PackWriter performance, and only when sending an initial
clone to a new client. There may be better ways to optimize this, and
until we have a solid solution, we shouldn't be using a separate cache
in JGit.