view mercurial/pvec.py @ 30442:41a8106789ca

util: implement zstd compression engine Now that zstd is vendored and being built (in some configurations), we can implement a compression engine for zstd! The zstd engine is a little different from existing engines. Because it may not always be present, we have to defer load the module in case importing it fails. We facilitate this via a cached property that holds a reference to the module or None. The "available" method is implemented to reflect reality. The zstd engine declares its ability to handle bundles using the "zstd" human name and the "ZS" internal name. The latter was chosen because internal names are 2 characters (by only convention I think) and "ZS" seems reasonable. The engine, like others, supports specifying the compression level. However, there are no consumers of this API that yet pass in that argument. I have plans to change that, so stay tuned. Since all we need to do to support bundle generation with a new compression engine is implement and register the compression engine, bundle generation with zstd "just works!" Tests demonstrating this have been added. How does performance of zstd for bundle generation compare? On the mozilla-unified repo, `hg bundle --all -t <engine>-v2` yields the following on my i7-6700K on Linux: engine CPU time bundle size vs orig size throughput none 97.0s 4,054,405,584 100.0% 41.8 MB/s bzip2 (l=9) 393.6s 975,343,098 24.0% 10.3 MB/s gzip (l=6) 184.0s 1,140,533,074 28.1% 22.0 MB/s zstd (l=1) 108.2s 1,119,434,718 27.6% 37.5 MB/s zstd (l=2) 111.3s 1,078,328,002 26.6% 36.4 MB/s zstd (l=3) 113.7s 1,011,823,727 25.0% 35.7 MB/s zstd (l=4) 116.0s 1,008,965,888 24.9% 35.0 MB/s zstd (l=5) 121.0s 977,203,148 24.1% 33.5 MB/s zstd (l=6) 131.7s 927,360,198 22.9% 30.8 MB/s zstd (l=7) 139.0s 912,808,505 22.5% 29.2 MB/s zstd (l=12) 198.1s 854,527,714 21.1% 20.5 MB/s zstd (l=18) 681.6s 789,750,690 19.5% 5.9 MB/s On compression, zstd for bundle generation delivers: * better compression than gzip with significantly less CPU utilization * better than bzip2 compression ratios while still being significantly faster than gzip * ability to aggressively tune compression level to achieve significantly smaller bundles That last point is important. With clone bundles, a server can pre-generate a bundle file, upload it to a static file server, and redirect clients to transparently download it during clone. The server could choose to produce a zstd bundle with the highest compression settings possible. This would take a very long time - a magnitude longer than a typical zstd bundle generation - but the result would be hundreds of megabytes smaller! For the clone volume we do at Mozilla, this could translate to petabytes of bandwidth savings per year and faster clones (due to smaller transfer size). I don't have detailed numbers to report on decompression. However, zstd decompression is fast: >1 GB/s output throughput on this machine, even through the Python bindings. And it can do that regardless of the compression level of the input. By the time you have enough data to worry about overhead of decompression, you have plenty of other things to worry about performance wise. zstd is wins all around. I can't wait to implement support for it on the wire protocol and in revlogs.
author Gregory Szorc <gregory.szorc@gmail.com>
date Fri, 11 Nov 2016 01:10:07 -0800
parents 983e93d88193
children 4462a981e8df
line wrap: on
line source

# pvec.py - probabilistic vector clocks for Mercurial
#
# Copyright 2012 Matt Mackall <mpm@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 (
    base85,
    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):
    bs = ""
    for p in xrange(l):
        bs = chr(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 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 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(base85.b85encode(bs))

class pvec(object):
    def __init__(self, hashorctx):
        if isinstance(hashorctx, str):
            self._bs = hashorctx
            self._depth, self._vec = _split(base85.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("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