Mercurial > hg
view mercurial/worker.py @ 39506:b66ea3fc3a86
sparse-revlog: set max delta chain length to on thousand
The new snapshot system used in the sparse-revlog case gave us some small size
benefit so far. However its most important property is to gracefully handle
harder limit on delta chainlength.
Long delta chain has a very detrimental impact on read (and write) performance
in revlog. Being able to shorter them provide a great boost. However, shorting
delta used to result significantly lower compression ratio. The intermediate
snapshots effectively suppress most of this effect (even all in some case).
# Effect on the test repository
The repository we use for test is not "realistic" but can still show this in
action using an unreasonably low chain limit. Limiting the chain length show a
sizeable increase but stay under control: +6% for limit=15; +15% for limit=10.
Without the snapshot system the increase is significantly bigger: +45% for
limit=15; +80% for limit=10. Even slightly larger than without delta chain
limit, the resulting size is still smaller than before we started doing
snapshots.
Here is a table for comparison. *Since the repository is not branchy, the
initial sparse-revlog version does not bring much benefit compare to the
non-sparse one):
chain length limit | none | limit=15 | limit=10 |
without sparse-revlog | 62 818 987 | 112 664 615 | 131 222 574 |
without snapshot | 74 365 490 | 108 211 410 | 133 857 764 |
with snapshot | 59 230 936 | 63 002 924 | 68 415 329 |
# Effect On Real Life Repositories
The series provides significant benefits on all kind of repositories.
Using `hg debugupgraderepo -o redeltaparent --run`, we recomputed delta chain
for various repositories with different settings:
- delta chain length: unlimited or 1000 limit
- sparse-revlog: enabled or disabled
- this series: applied or not applied
We can observe multiple types of effect:
- On very branchy repositories:
* The delta chain limit as low impact on the repo size.
* Intermediate snapshot greatly reduces manifest size:
- pypy: -80%
- netbeans: -95%
* The delta chain limit is effective, without a size impact:
- netbeans average: 613 -> 282
- private #1 average: 1 068 -> 307
- On more linear repository:
* Intermediate snapshot limit the impact of delta chain limit:
- mozilla:
without the series: +360%
with the series: +25%
* The delta chain limit provides large improvement:
- mozilla's average chain length:
unlimited: 15 338
limited: 469
* Despite the chain length limit, the manifest size is reduced:
- mercurial: -25%
- mozilla: -30%
It is clear that the use of chains of intermediate snapshots provide large
benefits both in storage size and delta chains quality. We should now switch our
effort toward making sure the write performance are acceptable. Then,
`sparse-revlog` will be a suitable format for all new repository.
# Raw Statistic
* no-sparse: general delta repository not using sparse-revlog
* no-snapshot: sparse-revlog repository not using this series
* snapshot: sparse-revlog repository using this series
mercurial
Manifest Size:
limit | none | 1000
------------|-------------|------------
no-sparse | 8 021 373 | 8 199 366
no-snapshot | 8 103 561 | 8 259 719
snapshot | 6 137 116 | 6 126 433
Manifest Chain length data
limit || none || 1000 ||
value || average | max || average | max ||
------------||---------|---------||---------|---------||
no-sparse || 307 | 1456 || 279 | 1000 ||
no-snapshot || 312 | 1456 || 283 | 1000 ||
snapshot || 248 | 1208 || 241 | 1000 ||
Full Store Size
limit | none | 1000
------------|-------------|------------
no-sparse | 51 013 198 | 51 201 574
no-snapshot | 50 930 795 | 51 141 006
snapshot | 48 072 037 | 48 093 572
pypy
Manifest Size:
limit | none | 1000
------------|-------------|------------
no-sparse | 193 987 784 | 193 987 784
no-snapshot | 163 171 745 | 163 312 229
snapshot | 34 605 900 | 34 600 750
Manifest Chain length data
limit || none || 1000 ||
value || average | max || average | max ||
------------||---------|---------||---------|---------||
no-sparse || 101 | 692 || 101 | 692 ||
no-snapshot || 151 | 1307 || 148 | 1000 ||
snapshot || 128 | 1309 || 125 | 1000 ||
Full Store Size
limit | none | 1000
------------|-------------|------------
no-sparse | 495 931 473 | 495 931 473
no-snapshot | 465 441 017 | 465 581 501
snapshot | 355 467 301 | 355 472 451
Mozilla
Manifest Size:
limit | none | 1000
------------|----------------|---------------
no-sparse | 416 757 148 | 1 869 009 668
no-snapshot | 401 592 370 | 1 843 493 795
snapshot | 224 359 521 | 284 615 500
Manifest Chain length data
limit || none || 1000 ||
value || average | max || average | max ||
------------||---------|---------||---------|---------||
no-sparse || 15 333 | 58 980 || 468 | 1 000 ||
no-snapshot || 15 336 | 58 980 || 469 | 1 000 ||
snapshot || 15 338 | 58 983 || 469 | 1 000 ||
Full Store Size
limit | none | 1000
------------|----------------|---------------
no-sparse | 2 712 477 887 | 4 164 995 451
no-snapshot | 2 698 887 835 | 4 141 054 304
snapshot | 2 518 130 385 | 2 578 587 596
Netbeans
Manifest Size:
limit | none | 1000
------------|----------------|---------------
no-sparse | 4 766 794 101 | 4 870 642 687
no-snapshot | 4 334 806 082 | 4 428 681 309
snapshot | 232 659 666 | 240 330 665
Manifest Chain length data
limit || none || 1000 ||
value || average | max || average | max ||
------------||---------|---------||---------|---------||
no-sparse || 597 | 6802 || 254 | 1 000 ||
no-snapshot || 648 | 6 802 || 305 | 1 000 ||
snapshot || 613 | 6 804 || 282 | 1 000 ||
Full Store Size
limit | none | 1000
------------|----------------|---------------
no-sparse | 5 807 347 998 | 5 911 196 584
no-snapshot | 5 375 398 602 | 5 469 273 829
snapshot | 1 282 519 928 | 1 290 190 927
Private repo #1
Manifest Size:
limit | none | 1000
------------|-----------------|---------------
no-sparse | 41 389 010 840 | 41 398 162 091
no-snapshot | 9 737 319 435 | 10 223 773 150
snapshot | 744 215 807 | 747 961 822
Manifest Chain length data
limit || none || 1000 ||
value || average | max || average | max ||
------------||---------|---------||---------|---------||
no-sparse || 245 | 8 885 || 81 | 1 000 ||
no-snapshot || 1 225 | 8 885 || 336 | 1 000 ||
snapshot || 1 068 | 7 909 || 307 | 1 000 ||
Full Store Size
limit | none | 1000
------------|----------------|---------------
no-sparse | 49 646 065 126 | 49 655 216 377
no-snapshot | 17 924 862 856 | 18 411 316 571
snapshot | 9 009 024 710 | 9 012 770 725
Private repo #2
We currently have less data available for that repository.
* Before is a sparse-revlog repository without this series
* After is a sparse-revlog repository with this series + 1000 chain limit
Manifest Size:
Before: 1 531 485 040 bytes
After: 1 091 422 451 bytes
Manifest Chain:
Before: 2 218 avg; 6 575 Max
After: 442 avg; 1 000 Max
Full Store Size
Before: 15 203 955 615
after: 8 207 180 693
author | Boris Feld <boris.feld@octobus.net> |
---|---|
date | Fri, 07 Sep 2018 11:18:45 -0400 |
parents | c08ea1e219c0 |
children | 909c31805f54 03f7d0822ec1 |
line wrap: on
line source
# worker.py - master-slave parallelism support # # Copyright 2013 Facebook, Inc. # # This software may be used and distributed according to the terms of the # GNU General Public License version 2 or any later version. from __future__ import absolute_import import errno import os import signal import sys import threading import time try: import selectors selectors.BaseSelector except ImportError: from .thirdparty import selectors2 as selectors from .i18n import _ from . import ( encoding, error, pycompat, scmutil, util, ) def countcpus(): '''try to count the number of CPUs on the system''' # posix try: n = int(os.sysconf(r'SC_NPROCESSORS_ONLN')) if n > 0: return n except (AttributeError, ValueError): pass # windows try: n = int(encoding.environ['NUMBER_OF_PROCESSORS']) if n > 0: return n except (KeyError, ValueError): pass return 1 def _numworkers(ui): s = ui.config('worker', 'numcpus') if s: try: n = int(s) if n >= 1: return n except ValueError: raise error.Abort(_('number of cpus must be an integer')) return min(max(countcpus(), 4), 32) if pycompat.isposix or pycompat.iswindows: _STARTUP_COST = 0.01 # The Windows worker is thread based. If tasks are CPU bound, threads # in the presence of the GIL result in excessive context switching and # this overhead can slow down execution. _DISALLOW_THREAD_UNSAFE = pycompat.iswindows else: _STARTUP_COST = 1e30 _DISALLOW_THREAD_UNSAFE = False def worthwhile(ui, costperop, nops, threadsafe=True): '''try to determine whether the benefit of multiple processes can outweigh the cost of starting them''' if not threadsafe and _DISALLOW_THREAD_UNSAFE: return False linear = costperop * nops workers = _numworkers(ui) benefit = linear - (_STARTUP_COST * workers + linear / workers) return benefit >= 0.15 def worker(ui, costperarg, func, staticargs, args, threadsafe=True): '''run a function, possibly in parallel in multiple worker processes. returns a progress iterator costperarg - cost of a single task func - function to run staticargs - arguments to pass to every invocation of the function args - arguments to split into chunks, to pass to individual workers threadsafe - whether work items are thread safe and can be executed using a thread-based worker. Should be disabled for CPU heavy tasks that don't release the GIL. ''' enabled = ui.configbool('worker', 'enabled') if enabled and worthwhile(ui, costperarg, len(args), threadsafe=threadsafe): return _platformworker(ui, func, staticargs, args) return func(*staticargs + (args,)) def _posixworker(ui, func, staticargs, args): workers = _numworkers(ui) oldhandler = signal.getsignal(signal.SIGINT) signal.signal(signal.SIGINT, signal.SIG_IGN) pids, problem = set(), [0] def killworkers(): # unregister SIGCHLD handler as all children will be killed. This # function shouldn't be interrupted by another SIGCHLD; otherwise pids # could be updated while iterating, which would cause inconsistency. signal.signal(signal.SIGCHLD, oldchldhandler) # if one worker bails, there's no good reason to wait for the rest for p in pids: try: os.kill(p, signal.SIGTERM) except OSError as err: if err.errno != errno.ESRCH: raise def waitforworkers(blocking=True): for pid in pids.copy(): p = st = 0 while True: try: p, st = os.waitpid(pid, (0 if blocking else os.WNOHANG)) break except OSError as e: if e.errno == errno.EINTR: continue elif e.errno == errno.ECHILD: # child would already be reaped, but pids yet been # updated (maybe interrupted just after waitpid) pids.discard(pid) break else: raise if not p: # skip subsequent steps, because child process should # be still running in this case continue pids.discard(p) st = _exitstatus(st) if st and not problem[0]: problem[0] = st def sigchldhandler(signum, frame): waitforworkers(blocking=False) if problem[0]: killworkers() oldchldhandler = signal.signal(signal.SIGCHLD, sigchldhandler) ui.flush() parentpid = os.getpid() pipes = [] for pargs in partition(args, workers): # Every worker gets its own pipe to send results on, so we don't have to # implement atomic writes larger than PIPE_BUF. Each forked process has # its own pipe's descriptors in the local variables, and the parent # process has the full list of pipe descriptors (and it doesn't really # care what order they're in). rfd, wfd = os.pipe() pipes.append((rfd, wfd)) # make sure we use os._exit in all worker code paths. otherwise the # worker may do some clean-ups which could cause surprises like # deadlock. see sshpeer.cleanup for example. # override error handling *before* fork. this is necessary because # exception (signal) may arrive after fork, before "pid =" assignment # completes, and other exception handler (dispatch.py) can lead to # unexpected code path without os._exit. ret = -1 try: pid = os.fork() if pid == 0: signal.signal(signal.SIGINT, oldhandler) signal.signal(signal.SIGCHLD, oldchldhandler) def workerfunc(): for r, w in pipes[:-1]: os.close(r) os.close(w) os.close(rfd) for result in func(*(staticargs + (pargs,))): os.write(wfd, util.pickle.dumps(result)) return 0 ret = scmutil.callcatch(ui, workerfunc) except: # parent re-raises, child never returns if os.getpid() == parentpid: raise exctype = sys.exc_info()[0] force = not issubclass(exctype, KeyboardInterrupt) ui.traceback(force=force) finally: if os.getpid() != parentpid: try: ui.flush() except: # never returns, no re-raises pass finally: os._exit(ret & 255) pids.add(pid) selector = selectors.DefaultSelector() for rfd, wfd in pipes: os.close(wfd) selector.register(os.fdopen(rfd, r'rb', 0), selectors.EVENT_READ) def cleanup(): signal.signal(signal.SIGINT, oldhandler) waitforworkers() signal.signal(signal.SIGCHLD, oldchldhandler) selector.close() status = problem[0] if status: if status < 0: os.kill(os.getpid(), -status) sys.exit(status) try: openpipes = len(pipes) while openpipes > 0: for key, events in selector.select(): try: yield util.pickle.load(key.fileobj) except EOFError: selector.unregister(key.fileobj) key.fileobj.close() openpipes -= 1 except IOError as e: if e.errno == errno.EINTR: continue raise except: # re-raises killworkers() cleanup() raise cleanup() def _posixexitstatus(code): '''convert a posix exit status into the same form returned by os.spawnv returns None if the process was stopped instead of exiting''' if os.WIFEXITED(code): return os.WEXITSTATUS(code) elif os.WIFSIGNALED(code): return -os.WTERMSIG(code) def _windowsworker(ui, func, staticargs, args): class Worker(threading.Thread): def __init__(self, taskqueue, resultqueue, func, staticargs, group=None, target=None, name=None, verbose=None): threading.Thread.__init__(self, group=group, target=target, name=name, verbose=verbose) self._taskqueue = taskqueue self._resultqueue = resultqueue self._func = func self._staticargs = staticargs self._interrupted = False self.daemon = True self.exception = None def interrupt(self): self._interrupted = True def run(self): try: while not self._taskqueue.empty(): try: args = self._taskqueue.get_nowait() for res in self._func(*self._staticargs + (args,)): self._resultqueue.put(res) # threading doesn't provide a native way to # interrupt execution. handle it manually at every # iteration. if self._interrupted: return except pycompat.queue.Empty: break except Exception as e: # store the exception such that the main thread can resurface # it as if the func was running without workers. self.exception = e raise threads = [] def trykillworkers(): # Allow up to 1 second to clean worker threads nicely cleanupend = time.time() + 1 for t in threads: t.interrupt() for t in threads: remainingtime = cleanupend - time.time() t.join(remainingtime) if t.is_alive(): # pass over the workers joining failure. it is more # important to surface the inital exception than the # fact that one of workers may be processing a large # task and does not get to handle the interruption. ui.warn(_("failed to kill worker threads while " "handling an exception\n")) return workers = _numworkers(ui) resultqueue = pycompat.queue.Queue() taskqueue = pycompat.queue.Queue() # partition work to more pieces than workers to minimize the chance # of uneven distribution of large tasks between the workers for pargs in partition(args, workers * 20): taskqueue.put(pargs) for _i in range(workers): t = Worker(taskqueue, resultqueue, func, staticargs) threads.append(t) t.start() try: while len(threads) > 0: while not resultqueue.empty(): yield resultqueue.get() threads[0].join(0.05) finishedthreads = [_t for _t in threads if not _t.is_alive()] for t in finishedthreads: if t.exception is not None: raise t.exception threads.remove(t) except (Exception, KeyboardInterrupt): # re-raises trykillworkers() raise while not resultqueue.empty(): yield resultqueue.get() if pycompat.iswindows: _platformworker = _windowsworker else: _platformworker = _posixworker _exitstatus = _posixexitstatus def partition(lst, nslices): '''partition a list into N slices of roughly equal size The current strategy takes every Nth element from the input. If we ever write workers that need to preserve grouping in input we should consider allowing callers to specify a partition strategy. mpm is not a fan of this partitioning strategy when files are involved. In his words: Single-threaded Mercurial makes a point of creating and visiting files in a fixed order (alphabetical). When creating files in order, a typical filesystem is likely to allocate them on nearby regions on disk. Thus, when revisiting in the same order, locality is maximized and various forms of OS and disk-level caching and read-ahead get a chance to work. This effect can be quite significant on spinning disks. I discovered it circa Mercurial v0.4 when revlogs were named by hashes of filenames. Tarring a repo and copying it to another disk effectively randomized the revlog ordering on disk by sorting the revlogs by hash and suddenly performance of my kernel checkout benchmark dropped by ~10x because the "working set" of sectors visited no longer fit in the drive's cache and the workload switched from streaming to random I/O. What we should really be doing is have workers read filenames from a ordered queue. This preserves locality and also keeps any worker from getting more than one file out of balance. ''' for i in range(nslices): yield lst[i::nslices]