view mercurial/pvec.py @ 48068:bf8837e3d7ce

dirstate: Remove the flat Rust DirstateMap implementation Before this changeset we had two Rust implementations of `DirstateMap`. This removes the "flat" DirstateMap so that the "tree" DirstateMap is always used when Rust enabled. This simplifies the code a lot, and will enable (in the next changeset) further removal of a trait abstraction. This is a performance regression when: * Rust is enabled, and * The repository uses the legacy dirstate-v1 file format, and * For `hg status`, unknown files are not listed (such as with `-mard`) The regression is about 100 milliseconds for `hg status -mard` on a semi-large repository (mozilla-central), from ~320ms to ~420ms. We deem this to be small enough to be worth it. The new dirstate-v2 is still experimental at this point, but we aim to stabilize it (though not yet enable it by default for new repositories) in Mercurial 6.0. Eventually, upgrating repositories to dirsate-v2 will eliminate this regression (and enable other performance improvements). # Background The flat DirstateMap was introduced with the first Rust implementation of the status algorithm. It works similarly to the previous Python + C one, with a single `HashMap` that associates file paths to a `DirstateEntry` (where Python has a dict). We later added the tree DirstateMap where the root of the tree contains nodes for files and directories that are directly at the root of the repository, and nodes for directories can contain child nodes representing the files and directly that *they* contain directly. The shape of this tree mirrors that of the working directory in the filesystem. This enables the status algorithm to traverse this tree in tandem with traversing the filesystem tree, which in turns enables a more efficient algorithm. Furthermore, the new dirstate-v2 file format is also based on a tree of the same shape. The tree DirstateMap can access a dirstate-v2 file without parsing it: binary data in a single large (possibly memory-mapped) bytes buffer is traversed on demand. This allows `DirstateMap` creation to take `O(1)` time. (Mutation works by creating new in-memory nodes with copy-on-write semantics, and serialization is append-mostly.) The tradeoff is that for "legacy" repositories that use the dirstate-v1 file format, parsing that file into a tree DirstateMap takes more time. Profiling shows that this time is dominated by `HashMap`. For a dirstate containing `F` files with an average `D` directory depth, the flat DirstateMap does parsing in `O(F)` number of HashMap operations but the tree DirstateMap in `O(F × D)` operations, since each node has its own HashMap containing its child nodes. This slower costs ~140ms on an old snapshot of mozilla-central, and ~80ms on an old snapshot of the Netbeans repository. The status algorithm is faster, but with `-mard` (when not listing unknown files) it is typically not faster *enough* to compensate the slower parsing. Both Rust implementations are always faster than the Python + C implementation # Benchmark results All benchmarks are run on changeset 98c0408324e6, with repositories that use the dirstate-v1 file format, on a server with 4 CPU cores and 4 CPU threads (no HyperThreading). `hg status` benchmarks show wall clock times of the entire command as the average and standard deviation of serveral runs, collected by https://github.com/sharkdp/hyperfine and reformated. Parsing benchmarks are wall clock time of the Rust function that converts a bytes buffer of the dirstate file into the `DirstateMap` data structure as used by the status algorithm. A single run each, collected by running `hg status` this environment variable: RUST_LOG=hg::dirstate::dirstate_map=trace,hg::dirstate_tree::dirstate_map=trace Benchmark 1: Rust flat DirstateMap → Rust tree DirstateMap hg status mozilla-clean 562.3 ms ± 2.0 ms → 462.5 ms ± 0.6 ms 1.22 ± 0.00 times faster mozilla-dirty 859.6 ms ± 2.2 ms → 719.5 ms ± 3.2 ms 1.19 ± 0.01 times faster mozilla-ignored 558.2 ms ± 3.0 ms → 457.9 ms ± 2.9 ms 1.22 ± 0.01 times faster mozilla-unknowns 859.4 ms ± 5.7 ms → 716.0 ms ± 4.7 ms 1.20 ± 0.01 times faster netbeans-clean 336.5 ms ± 0.9 ms → 339.5 ms ± 0.4 ms 0.99 ± 0.00 times faster netbeans-dirty 491.4 ms ± 1.6 ms → 475.1 ms ± 1.2 ms 1.03 ± 0.00 times faster netbeans-ignored 343.7 ms ± 1.0 ms → 347.8 ms ± 0.4 ms 0.99 ± 0.00 times faster netbeans-unknowns 484.3 ms ± 1.0 ms → 466.0 ms ± 1.2 ms 1.04 ± 0.00 times faster hg status -mard mozilla-clean 317.3 ms ± 0.6 ms → 422.5 ms ± 1.2 ms 0.75 ± 0.00 times faster mozilla-dirty 315.4 ms ± 0.6 ms → 417.7 ms ± 1.1 ms 0.76 ± 0.00 times faster mozilla-ignored 314.6 ms ± 0.6 ms → 417.4 ms ± 1.0 ms 0.75 ± 0.00 times faster mozilla-unknowns 312.9 ms ± 0.9 ms → 417.3 ms ± 1.6 ms 0.75 ± 0.00 times faster netbeans-clean 212.0 ms ± 0.6 ms → 283.6 ms ± 0.8 ms 0.75 ± 0.00 times faster netbeans-dirty 211.4 ms ± 1.0 ms → 283.4 ms ± 1.6 ms 0.75 ± 0.01 times faster netbeans-ignored 211.4 ms ± 0.9 ms → 283.9 ms ± 0.8 ms 0.74 ± 0.01 times faster netbeans-unknowns 211.1 ms ± 0.6 ms → 283.4 ms ± 1.0 ms 0.74 ± 0.00 times faster Parsing mozilla-clean 38.4ms → 177.6ms mozilla-dirty 38.8ms → 177.0ms mozilla-ignored 38.8ms → 178.0ms mozilla-unknowns 38.7ms → 176.9ms netbeans-clean 16.5ms → 97.3ms netbeans-dirty 16.5ms → 98.4ms netbeans-ignored 16.9ms → 97.4ms netbeans-unknowns 16.9ms → 96.3ms Benchmark 2: Python + C dirstatemap → Rust tree DirstateMap hg status mozilla-clean 1261.0 ms ± 3.6 ms → 461.1 ms ± 0.5 ms 2.73 ± 0.00 times faster mozilla-dirty 2293.4 ms ± 9.1 ms → 719.6 ms ± 3.6 ms 3.19 ± 0.01 times faster mozilla-ignored 1240.4 ms ± 2.3 ms → 457.7 ms ± 1.9 ms 2.71 ± 0.00 times faster mozilla-unknowns 2283.3 ms ± 9.0 ms → 719.7 ms ± 3.8 ms 3.17 ± 0.01 times faster netbeans-clean 879.7 ms ± 3.5 ms → 339.9 ms ± 0.5 ms 2.59 ± 0.00 times faster netbeans-dirty 1257.3 ms ± 4.7 ms → 474.6 ms ± 1.6 ms 2.65 ± 0.01 times faster netbeans-ignored 943.9 ms ± 1.9 ms → 347.3 ms ± 1.1 ms 2.72 ± 0.00 times faster netbeans-unknowns 1188.1 ms ± 5.0 ms → 465.2 ms ± 2.3 ms 2.55 ± 0.01 times faster hg status -mard mozilla-clean 903.2 ms ± 3.6 ms → 423.4 ms ± 2.2 ms 2.13 ± 0.01 times faster mozilla-dirty 884.6 ms ± 4.5 ms → 417.3 ms ± 1.4 ms 2.12 ± 0.01 times faster mozilla-ignored 881.9 ms ± 1.3 ms → 417.3 ms ± 0.8 ms 2.11 ± 0.00 times faster mozilla-unknowns 878.5 ms ± 1.9 ms → 416.4 ms ± 0.9 ms 2.11 ± 0.00 times faster netbeans-clean 434.9 ms ± 1.8 ms → 284.0 ms ± 0.8 ms 1.53 ± 0.01 times faster netbeans-dirty 434.1 ms ± 0.8 ms → 283.1 ms ± 0.8 ms 1.53 ± 0.00 times faster netbeans-ignored 431.7 ms ± 1.1 ms → 283.6 ms ± 1.8 ms 1.52 ± 0.01 times faster netbeans-unknowns 433.0 ms ± 1.3 ms → 283.5 ms ± 0.7 ms 1.53 ± 0.00 times faster Differential Revision: https://phab.mercurial-scm.org/D11516
author Simon Sapin <simon.sapin@octobus.net>
date Mon, 27 Sep 2021 12:09:15 +0200
parents d4ba4d51f85f
children 6000f5b25c9b
line wrap: on
line source

# pvec.py - probabilistic vector clocks for Mercurial
#
# Copyright 2012 Olivia Mackall <olivia@selenic.com>
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.

'''
A "pvec" is a changeset property based on the theory of vector clocks
that can be compared to discover relatedness without consulting a
graph. This can be useful for tasks like determining how a
disconnected patch relates to a repository.

Currently a pvec consist of 448 bits, of which 24 are 'depth' and the
remainder are a bit vector. It is represented as a 70-character base85
string.

Construction:

- a root changeset has a depth of 0 and a bit vector based on its hash
- a normal commit has a changeset where depth is increased by one and
  one bit vector bit is flipped based on its hash
- a merge changeset pvec is constructed by copying changes from one pvec into
  the other to balance its depth

Properties:

- for linear changes, difference in depth is always <= hamming distance
- otherwise, changes are probably divergent
- when hamming distance is < 200, we can reliably detect when pvecs are near

Issues:

- hamming distance ceases to work over distances of ~ 200
- detecting divergence is less accurate when the common ancestor is very close
  to either revision or total distance is high
- this could probably be improved by modeling the relation between
  delta and hdist

Uses:

- a patch pvec can be used to locate the nearest available common ancestor for
  resolving conflicts
- ordering of patches can be established without a DAG
- two head pvecs can be compared to determine whether push/pull/merge is needed
  and approximately how many changesets are involved
- can be used to find a heuristic divergence measure between changesets on
  different branches
'''

from __future__ import absolute_import

from .node import nullrev
from . import (
    pycompat,
    util,
)

_size = 448  # 70 chars b85-encoded
_bytes = _size // 8
_depthbits = 24
_depthbytes = _depthbits // 8
_vecbytes = _bytes - _depthbytes
_vecbits = _vecbytes * 8
_radius = (_vecbits - 30) // 2  # high probability vectors are related


def _bin(bs):
    '''convert a bytestring to a long'''
    v = 0
    for b in bs:
        v = v * 256 + ord(b)
    return v


def _str(v, l):
    # type: (int, int) -> bytes
    bs = b""
    for p in pycompat.xrange(l):
        bs = pycompat.bytechr(v & 255) + bs
        v >>= 8
    return bs


def _split(b):
    '''depth and bitvec'''
    return _bin(b[:_depthbytes]), _bin(b[_depthbytes:])


def _join(depth, bitvec):
    return _str(depth, _depthbytes) + _str(bitvec, _vecbytes)


def _hweight(x):
    c = 0
    while x:
        if x & 1:
            c += 1
        x >>= 1
    return c


_htab = [_hweight(x) for x in pycompat.xrange(256)]


def _hamming(a, b):
    '''find the hamming distance between two longs'''
    d = a ^ b
    c = 0
    while d:
        c += _htab[d & 0xFF]
        d >>= 8
    return c


def _mergevec(x, y, c):
    # Ideally, this function would be x ^ y ^ ancestor, but finding
    # ancestors is a nuisance. So instead we find the minimal number
    # of changes to balance the depth and hamming distance

    d1, v1 = x
    d2, v2 = y
    if d1 < d2:
        d1, d2, v1, v2 = d2, d1, v2, v1

    hdist = _hamming(v1, v2)
    ddist = d1 - d2
    v = v1
    m = v1 ^ v2  # mask of different bits
    i = 1

    if hdist > ddist:
        # if delta = 10 and hdist = 100, then we need to go up 55 steps
        # to the ancestor and down 45
        changes = (hdist - ddist + 1) // 2
    else:
        # must make at least one change
        changes = 1
    depth = d1 + changes

    # copy changes from v2
    if m:
        while changes:
            if m & i:
                v ^= i
                changes -= 1
            i <<= 1
    else:
        v = _flipbit(v, c)

    return depth, v


def _flipbit(v, node):
    # converting bit strings to longs is slow
    bit = (hash(node) & 0xFFFFFFFF) % _vecbits
    return v ^ (1 << bit)


def ctxpvec(ctx):
    '''construct a pvec for ctx while filling in the cache'''
    r = ctx.repo()
    if not util.safehasattr(r, "_pveccache"):
        r._pveccache = {}
    pvc = r._pveccache
    if ctx.rev() not in pvc:
        cl = r.changelog
        for n in pycompat.xrange(ctx.rev() + 1):
            if n not in pvc:
                node = cl.node(n)
                p1, p2 = cl.parentrevs(n)
                if p1 == nullrev:
                    # start with a 'random' vector at root
                    pvc[n] = (0, _bin((node * 3)[:_vecbytes]))
                elif p2 == nullrev:
                    d, v = pvc[p1]
                    pvc[n] = (d + 1, _flipbit(v, node))
                else:
                    pvc[n] = _mergevec(pvc[p1], pvc[p2], node)
    bs = _join(*pvc[ctx.rev()])
    return pvec(util.b85encode(bs))


class pvec(object):
    def __init__(self, hashorctx):
        if isinstance(hashorctx, bytes):
            self._bs = hashorctx
            self._depth, self._vec = _split(util.b85decode(hashorctx))
        else:
            self._vec = ctxpvec(hashorctx)

    def __str__(self):
        return self._bs

    def __eq__(self, b):
        return self._vec == b._vec and self._depth == b._depth

    def __lt__(self, b):
        delta = b._depth - self._depth
        if delta < 0:
            return False  # always correct
        if _hamming(self._vec, b._vec) > delta:
            return False
        return True

    def __gt__(self, b):
        return b < self

    def __or__(self, b):
        delta = abs(b._depth - self._depth)
        if _hamming(self._vec, b._vec) <= delta:
            return False
        return True

    def __sub__(self, b):
        if self | b:
            raise ValueError(b"concurrent pvecs")
        return self._depth - b._depth

    def distance(self, b):
        d = abs(b._depth - self._depth)
        h = _hamming(self._vec, b._vec)
        return max(d, h)

    def near(self, b):
        dist = abs(b.depth - self._depth)
        if dist > _radius or _hamming(self._vec, b._vec) > _radius:
            return False