This diff implements the standard dict copy() method for lrucachedicts, which
will be used in the pushrebase extension to make a copy of the manifestcache.
Contrary to the comment, I didn't see any evidence that we were copying
atime/mtime at all. This adds a parameter to copyfile to optionally copy it and
other stat data, with the default being to not copy it.
Many systems don't support changing the timestamp of a symlink, but we don't
need that in general anyway -- copystat is mostly useful for editors, most of
which will dereference symlinks anyway.
As part of attempting to more aggressively use the existing
lrucachedict, collections.deque operations were frequently
showing up in profiling output, negating benefits of caching.
Searching the internet seems to tell me that the most efficient
way to implement an LRU cache in Python is to have a dict indexing
the cached entries and then to use a doubly linked list to track
freshness of each entry. So, this patch replaces our existing
lrucachedict with a version using such a pattern.
The recently introduced perflrucachedict command reveals the
following timings for 10,000 operations for the following cache
sizes for the existing cache:
n=4 init=0.004079 gets=0.003632 sets=0.005188 mixed=0.005402
n=8 init=0.004045 gets=0.003998 sets=0.005064 mixed=0.005328
n=16 init=0.004011 gets=0.004496 sets=0.005021 mixed=0.005555
n=32 init=0.004064 gets=0.005611 sets=0.005188 mixed=0.006189
n=64 init=0.003975 gets=0.007684 sets=0.005178 mixed=0.007245
n=128 init=0.004121 gets=0.012005 sets=0.005422 mixed=0.009471
n=256 init=0.004143 gets=0.020295 sets=0.005227 mixed=0.013612
n=512 init=0.004039 gets=0.036703 sets=0.005243 mixed=0.020685
n=1024 init=0.004193 gets=0.068142 sets=0.005251 mixed=0.033064
n=2048 init=0.004070 gets=0.133383 sets=0.005160 mixed=0.050359
n=4096 init=0.004053 gets=0.265194 sets=0.004868 mixed=0.048352
n=8192 init=0.004087 gets=0.542218 sets=0.004562 mixed=0.032753
n=16384 init=0.004106 gets=1.064055 sets=0.004179 mixed=0.020367
n=32768 init=0.004034 gets=2.097620 sets=0.004260 mixed=0.013031
n=65536 init=0.004108 gets=4.106390 sets=0.004268 mixed=0.010191
As the data shows, the existing cache's retrieval performance
diminishes linearly with cache size. (Keep in mind the microbenchmark
is testing 100% cache hit rate.)
The new cache implementation reveals the following:
n=4 init=0.006665 gets=0.006541 sets=0.005733 mixed=0.006876
n=8 init=0.006649 gets=0.006374 sets=0.005663 mixed=0.006899
n=16 init=0.006570 gets=0.006504 sets=0.005799 mixed=0.007057
n=32 init=0.006854 gets=0.006459 sets=0.005747 mixed=0.007034
n=64 init=0.006580 gets=0.006495 sets=0.005740 mixed=0.006992
n=128 init=0.006534 gets=0.006739 sets=0.005648 mixed=0.007124
n=256 init=0.006669 gets=0.006773 sets=0.005824 mixed=0.007151
n=512 init=0.006701 gets=0.007061 sets=0.006042 mixed=0.007372
n=1024 init=0.006641 gets=0.007620 sets=0.006387 mixed=0.007464
n=2048 init=0.006517 gets=0.008598 sets=0.006871 mixed=0.008077
n=4096 init=0.006720 gets=0.010933 sets=0.007854 mixed=0.008663
n=8192 init=0.007383 gets=0.015969 sets=0.010288 mixed=0.008896
n=16384 init=0.006660 gets=0.025447 sets=0.011208 mixed=0.008826
n=32768 init=0.006658 gets=0.044390 sets=0.011192 mixed=0.008943
n=65536 init=0.006836 gets=0.082736 sets=0.011151 mixed=0.008826
Let's go through the results.
The new cache takes longer to construct. ~6.6ms vs ~4.1ms. However,
this is measuring 10,000 __init__ calls, so the difference is
~0.2us/instance. We currently only create lrucachedict for manifest
instances, so this regression is not likely relevant.
The new cache is slightly slower for retrievals for cache sizes
< 1024. It's worth noting that the only existing use of lurcachedict
is in manifest.py and the default cache size is 4. This regression
is worrisome. However, for n=4, the delta is ~2.9s for 10,000 lookups,
or ~0.29us/op. Again, this is a marginal regression and likely not
relevant in the real world. Timing `hg log -p -l 100` for
mozilla-central reveals that cache lookup times are dominated by
decompression and fulltext resolution (even with lz4 manifests).
The new cache is significantly faster for retrievals at larger
capacities. Whereas the old implementation has retrieval performance
linear with cache capacity, the new cache is constant time until much
larger values. And, when it does start to increase significantly, it
is a few magnitudes faster than the current cache.
The new cache does appear to be slower for sets when capacity is large.
However, performance is similar for smaller capacities. Of course,
caches should generally be optimized for retrieval performance because
if a cache is getting more sets than gets, it doesn't really make
sense to cache. If this regression is worrisome, again, taking the
largest regression at n=65536 of ~6.9ms for 10,000 results in a
regression of ~0.68us/op. This is not significant in the grand scheme
of things.
Overall, the new cache is performant at retrievals at much larger
capacity values which makes it a generally more useful cache backend.
While there are regressions, their absolute value is extremely small.
Since we aren't using lrucachedict aggressively today, these
regressions should not be relevant. The improved scalability of
lrucachedict should enable us to more aggressively utilize
lrucachedict for more granular caching (read: higher capacity caches)
in the near future. The impetus for this patch is to establish a cache
of decompressed revlog revisions, notably manifest revisions. And since
delta chains can grow to >10,000 and cache hit rate can be high, the
improved retrieval performance of lrucachedict should be relevant.
hashlib was added in Python 2.5. As far as I can tell, SHA-512 is always
available in 2.6+. So move the hashlib import to the top of the file and
remove the one-off handling of SHA-512.
We were computing the quotient and remainder of a division operation
separately. The built-in divmod() function allows us to do this with
a single function call. Do that.
We've globablly forced stat to return integer times which agrees with
our extension code, so this is no longer needed.
This speeds up status on mozilla-central substantially:
$ hg perfstatus
! wall 0.190179 comb 0.180000 user 0.120000 sys 0.060000 (best of 53)
$ hg perfstatus
! wall 0.275729 comb 0.270000 user 0.210000 sys 0.060000 (best of 36)
Alternate fix for this issue which avoids putting extra function calls
and exception handling in the fast path.
For almost all purposes, integer timestamps are preferable to
Mercurial. It stores integer timestamps in the dirstate and would thus
like to avoid doing any float/int comparisons or conversions. We will
continue to have to deal with 1-second granularity on filesystems for
quite some time, so this won't significantly hinder our capabilities.
This has some impact on our file cache validation code in that it
lowers timestamp resolution. But as we still have to deal with
low-resolution filesystems, we're not relying on this anyway.
An alternate approach is to use stat[ST_MTIME], which is guaranteed to
be an integer. But since this support isn't already in our extension,
we can't depend on it being available without adding a hard Python->C
API dependency that's painful for people like yours truly who have
bisect regularly and people without compilers.
The home of 'Abort' is 'error' not 'util' however, a lot of code seems to be
confused about that and gives all the credit to 'util' instead of the
hardworking 'error'. In a spirit of equity, we break the cycle of injustice and
give back to 'error' the respect it deserves. And screw that 'util' poser.
For great justice.
Because st.st_mtime is computed as 'sec + 1e-9 * nsec' and double is too narrow
to represent nanoseconds, int(st.st_mtime) can be 'sec + 1'. Therefore, that
value could be different from the one got by osutils.listdir().
This patch fixes the problem by accessing to raw st_mtime by tuple index.
It catches TypeError to fall back to st.st_mtime because our osutil.stat does
not support tuple index. In dirstate.normal(), 'st' is always a Python stat,
but in dirstate.status(), it can be either a Python stat or an osutil.stat.
Thanks to vgatien-baron@janestreet.com for finding the root cause of this
subtle problem.
Previously, a read(N) where N was less than the length of the first
available chunk would mutate the deque instance twice and allocate a new
str from the slice of the existing chunk. Profiling drawed my attention
to these as a potential hot spot during changegroup reading.
This patch makes the code more complicated in order to avoid the
aforementioned 3 operations.
On a pre-generated mozilla-central gzip bundle, this series has the
following impact on `hg unbundle` performance on my MacBook Pro:
before: 358.21 real 317.69 user 38.49 sys
after: 301.57 real 262.69 user 37.11 sys
delta: -56.64 real -55.00 user -1.38 sys
The new code results in simpler logic within the while loop. It is also
faster since we avoid performing operations on the queue and buf
collections. However, there shouldn't be any super hot loops for this
since the whole point of chunkbuffer is to avoid reading large amounts
of data at once. This does, however, make it easier to optimize
chunkbuffer in a subsequent patch.
For "space saving", bundle1 "strip" the first two bytes of the BZ stream since
they always are 'BZ'. So the current code boostrap the uncompressor with 'BZ'.
This hack is impractical in more generic case so we move it in a dedicated
"decompression".
The assumption that dynamically computing the length of the buffer was N^2, but
negligible because fast was False. So we drop the dynamic computation and
manually keep track of the buffer length.
This comment is the remains of a intermediate implementation using
self._buffer += data
This implementation never made it to the repository and we can safely drop the
comment.
Python 2.6 introduced the "except type as instance" syntax, replacing
the "except type, instance" syntax that came before. Python 3 dropped
support for the latter syntax. Since we no longer support Python 2.4 or
2.5, we have no need to continue supporting the "except type, instance".
This patch mass rewrites the exception syntax to be Python 2.6+ and
Python 3 compatible.
This patch was produced by running `2to3 -f except -w -n .`.
We'll use it to detect when a sshpeer have server output to be displayed.
The implementation is super basic because all case support is not the focus of
this series.
To restore real time server output through ssh, we need to using polling feature
(like select) on the pipes used to communicate with the ssh client. However
we cannot use select alongside python level buffering of these pipe (because we
need to know if the buffer is non-empty before calling select).
However, unbuffered performance are terrible, presumably because the 'readline'
call is issuing 'read(1)' call until it find a '\n'. To work around that we
introduces our own overlay that do buffering by hand, exposing the state of the
buffer to the outside world.
The usage of polling IO will be introduced later in the 'sshpeer' module. All
its logic will be very specific to the way mercurial communicate over ssh and
does not belong to the generic 'util' module.
We will need unbuffered IO to restore real time output with ssh peer.
Changeset b61b215fcfa8 seems to indicate playing with this value could be
dangerous, but does not indicate why.
One case where that would happen is while trying to resolve a subrepo, if the
path to the subrepo was actually a broken symlink. This bug was exposed by an
hg-git test.
According to 6b1369445b7b introducing "windows._removedirs()":
If a hg repository including working directory is a reparse point
(directory symlinked or a junction point), then using
os.removedirs will remove the reparse point erroneously.
"windows._removedirs()" should be used instead of "os.removedirs()" on
Windows.
This patch adds "removedirs" as platform depending function to replace
"os.removedirs()" invocations for portability and safety
An upcoming commit requires that match.py be able to call scmutil.dirs(), but
when match.py imports scmutil, a dependency cycle is created. This commit
avoids the cycle by moving dirs() and its related finddirs() function from
scmutil to util, which match.py already depends on.
Hi there!
Fixed date names are helpful for automated systems. So it is possible to
use english date parameter even if the underlying system uses another
locale.
We have here a jenkins with build jobs on different slaves that will do
some operations with "dates" parameter. Some systems uses English locale
and some systems uses German locale. So we needed to configure the job to
uses other date names.
As this is really annoying to keep the systems locale in mind for some
operations I looked into util.py. It would be helpful for automated systems
if the "default English" date names would even usable on other
locales.
I attached a simple patch for this.
Best regards
André Klitzing