]> Cypherpunks.ru repositories - gostls13.git/commit
index/suffixarray: index 3-10X faster in half the memory
authorRuss Cox <rsc@golang.org>
Fri, 26 Apr 2019 14:32:15 +0000 (10:32 -0400)
committerRuss Cox <rsc@golang.org>
Mon, 13 May 2019 18:50:32 +0000 (18:50 +0000)
commit6ca324f2837db696dff8e7d7342280dd5cdf6bca
tree7ef986695b068450a88642dd2fbb0862aef46c8b
parent7a43f8a5fb355a9cff73314d234da9a817695cbd
index/suffixarray: index 3-10X faster in half the memory

This CL changes the index/suffixarray construction algorithm from QSufSort to SAIS.

For an N-byte input, QSufSort runs in O(N log N) time and requires
an N-int temporary work space in addition to the N-int output.

In contrast, SAIS runs in O(N) time and, for essentially all real inputs,
is able to use the N-int output buffer as its temporary work space.
(In pathological cases, SAIS must allocate a temporary work space of
at most N/2 ints. There exist more complex variants that guarantee
to avoid the work space in all cases, but they hardly seem worth the cost
given how rare these pathological cases are.)

The SAIS code therefore uses 50% of the memory across the board.
It also runs 3-10X faster on real input text.

This CL also adds more extensive algorithmic tests, including an
exhaustive test over small inputs to catch corner case problems.

name                                   old speed      new speed        delta
New/text=opticks/size=100K/bits=32-12  6.15MB/s ± 1%   26.79MB/s ± 1%   +335.89%  (p=0.008 n=5+5)
New/text=opticks/size=100K/bits=64-12  5.90MB/s ± 2%   27.29MB/s ± 2%   +362.23%  (p=0.008 n=5+5)
New/text=opticks/size=500K/bits=32-12  4.99MB/s ± 3%   25.37MB/s ± 2%   +408.01%  (p=0.008 n=5+5)
New/text=opticks/size=500K/bits=64-12  4.88MB/s ± 1%   25.66MB/s ± 4%   +425.52%  (p=0.008 n=5+5)
New/text=go/size=100K/bits=32-12       5.81MB/s ± 1%   26.49MB/s ± 2%   +355.85%  (p=0.008 n=5+5)
New/text=go/size=100K/bits=64-12       5.76MB/s ± 2%   26.65MB/s ± 3%   +362.60%  (p=0.008 n=5+5)
New/text=go/size=500K/bits=32-12       4.91MB/s ± 1%   25.12MB/s ± 2%   +411.86%  (p=0.008 n=5+5)
New/text=go/size=500K/bits=64-12       4.83MB/s ± 2%   25.79MB/s ± 2%   +434.44%  (p=0.008 n=5+5)
New/text=go/size=1M/bits=32-12         4.62MB/s ± 2%   24.87MB/s ± 2%   +438.78%  (p=0.008 n=5+5)
New/text=go/size=1M/bits=64-12         4.39MB/s ± 2%   24.61MB/s ± 2%   +460.68%  (p=0.008 n=5+5)
New/text=go/size=5M/bits=32-12         2.85MB/s ± 2%   24.78MB/s ± 7%   +768.33%  (p=0.008 n=5+5)
New/text=go/size=5M/bits=64-12         2.28MB/s ± 1%   18.70MB/s ± 7%   +719.63%  (p=0.008 n=5+5)
New/text=go/size=10M/bits=32-12        2.08MB/s ± 1%   21.04MB/s ± 6%   +909.60%  (p=0.008 n=5+5)
New/text=go/size=10M/bits=64-12        1.83MB/s ± 1%   16.64MB/s ± 2%   +809.18%  (p=0.008 n=5+5)
New/text=go/size=50M/bits=32-12        1.51MB/s ± 0%   10.58MB/s ± 1%   +602.52%  (p=0.008 n=5+5)
New/text=go/size=50M/bits=64-12        1.34MB/s ± 4%    9.00MB/s ± 1%   +569.35%  (p=0.008 n=5+5)
New/text=zero/size=100K/bits=32-12     4.17MB/s ± 0%  157.56MB/s ± 1%  +3678.42%  (p=0.016 n=4+5)
New/text=zero/size=100K/bits=64-12     4.19MB/s ± 2%  162.72MB/s ± 2%  +3783.63%  (p=0.008 n=5+5)
New/text=zero/size=500K/bits=32-12     3.72MB/s ± 5%  159.17MB/s ± 1%  +4176.57%  (p=0.008 n=5+5)
New/text=zero/size=500K/bits=64-12     3.77MB/s ± 3%  164.95MB/s ± 4%  +4277.60%  (p=0.008 n=5+5)
New/text=zero/size=1M/bits=32-12       3.46MB/s ± 3%  158.42MB/s ± 1%  +4476.08%  (p=0.008 n=5+5)
New/text=zero/size=1M/bits=64-12       3.41MB/s ± 4%  163.70MB/s ± 2%  +4700.65%  (p=0.008 n=5+5)
New/text=zero/size=5M/bits=32-12       3.12MB/s ± 2%  151.92MB/s ± 4%  +4775.48%  (p=0.008 n=5+5)
New/text=zero/size=5M/bits=64-12       3.09MB/s ± 2%  166.19MB/s ± 2%  +5274.84%  (p=0.008 n=5+5)
New/text=zero/size=10M/bits=32-12      2.97MB/s ± 1%  157.75MB/s ± 1%  +5211.38%  (p=0.008 n=5+5)
New/text=zero/size=10M/bits=64-12      2.92MB/s ± 1%  162.75MB/s ± 2%  +5473.77%  (p=0.008 n=5+5)
New/text=zero/size=50M/bits=32-12      2.67MB/s ± 1%  144.43MB/s ± 5%  +5305.39%  (p=0.008 n=5+5)
New/text=zero/size=50M/bits=64-12      2.61MB/s ± 1%  125.19MB/s ± 2%  +4700.33%  (p=0.016 n=5+4)
New/text=rand/size=100K/bits=32-12     8.69MB/s ± 6%   27.60MB/s ± 1%   +217.73%  (p=0.008 n=5+5)
New/text=rand/size=100K/bits=64-12     8.92MB/s ± 1%   26.37MB/s ± 4%   +195.50%  (p=0.008 n=5+5)
New/text=rand/size=500K/bits=32-12     7.11MB/s ± 2%   25.23MB/s ± 2%   +254.78%  (p=0.008 n=5+5)
New/text=rand/size=500K/bits=64-12     7.08MB/s ± 1%   25.45MB/s ± 2%   +259.56%  (p=0.008 n=5+5)
New/text=rand/size=1M/bits=32-12       6.45MB/s ± 2%   24.47MB/s ± 3%   +279.11%  (p=0.008 n=5+5)
New/text=rand/size=1M/bits=64-12       6.09MB/s ± 4%   23.00MB/s ± 4%   +277.85%  (p=0.008 n=5+5)
New/text=rand/size=5M/bits=32-12       3.68MB/s ± 3%   10.34MB/s ± 5%   +181.08%  (p=0.008 n=5+5)
New/text=rand/size=5M/bits=64-12       3.25MB/s ± 1%    6.23MB/s ± 1%    +91.93%  (p=0.008 n=5+5)
New/text=rand/size=10M/bits=32-12      3.03MB/s ± 1%    5.61MB/s ± 2%    +85.28%  (p=0.008 n=5+5)
New/text=rand/size=10M/bits=64-12      2.80MB/s ± 1%    4.29MB/s ± 2%    +53.40%  (p=0.008 n=5+5)
New/text=rand/size=50M/bits=32-12      2.11MB/s ± 0%    2.45MB/s ± 1%    +16.23%  (p=0.029 n=4+4)
New/text=rand/size=50M/bits=64-12      2.04MB/s ± 1%    2.24MB/s ± 1%    +10.03%  (p=0.016 n=5+4)
SaveRestore/bits=32-12                  327MB/s ± 5%     319MB/s ± 2%       ~     (p=0.310 n=5+5)
SaveRestore/bits=64-12                  306MB/s ± 3%     306MB/s ± 2%       ~     (p=0.841 n=5+5)

name                                   old alloc/op   new alloc/op     delta
New/text=opticks/size=100K/bits=32-12     811kB ± 0%       401kB ± 0%    -50.51%  (p=0.008 n=5+5)
New/text=opticks/size=100K/bits=64-12    1.62MB ± 0%      0.80MB ± 0%    -50.51%  (p=0.008 n=5+5)
New/text=opticks/size=500K/bits=32-12    4.04MB ± 0%      2.01MB ± 0%    -50.37%  (p=0.008 n=5+5)
New/text=opticks/size=500K/bits=64-12    8.07MB ± 0%      4.01MB ± 0%    -50.36%  (p=0.016 n=4+5)
New/text=go/size=100K/bits=32-12          811kB ± 0%       401kB ± 0%       ~     (p=0.079 n=4+5)
New/text=go/size=100K/bits=64-12         1.62MB ± 0%      0.80MB ± 0%    -50.50%  (p=0.008 n=5+5)
New/text=go/size=500K/bits=32-12         4.04MB ± 0%      2.01MB ± 0%       ~     (p=0.079 n=4+5)
New/text=go/size=500K/bits=64-12         8.07MB ± 0%      4.01MB ± 0%    -50.36%  (p=0.000 n=4+5)
New/text=go/size=1M/bits=32-12           8.07MB ± 0%      4.01MB ± 0%    -50.36%  (p=0.008 n=5+5)
New/text=go/size=1M/bits=64-12           16.1MB ± 0%       8.0MB ± 0%    -50.36%  (p=0.008 n=5+5)
New/text=go/size=5M/bits=32-12           40.2MB ± 0%      20.0MB ± 0%    -50.18%  (p=0.008 n=5+5)
New/text=go/size=5M/bits=64-12           80.3MB ± 0%      40.0MB ± 0%    -50.18%  (p=0.008 n=5+5)
New/text=go/size=10M/bits=32-12          80.2MB ± 0%      40.0MB ± 0%    -50.09%  (p=0.000 n=5+4)
New/text=go/size=10M/bits=64-12           160MB ± 0%        80MB ± 0%    -50.09%  (p=0.000 n=5+4)
New/text=go/size=50M/bits=32-12           402MB ± 0%       200MB ± 0%    -50.29%  (p=0.000 n=5+4)
New/text=go/size=50M/bits=64-12           805MB ± 0%       400MB ± 0%    -50.29%  (p=0.000 n=5+4)
New/text=zero/size=100K/bits=32-12       1.46MB ± 0%      0.40MB ± 0%    -72.46%  (p=0.008 n=5+5)
New/text=zero/size=100K/bits=64-12       3.02MB ± 0%      0.80MB ± 0%    -73.45%  (p=0.008 n=5+5)
New/text=zero/size=500K/bits=32-12       8.66MB ± 0%      2.01MB ± 0%       ~     (p=0.079 n=4+5)
New/text=zero/size=500K/bits=64-12       19.7MB ± 0%       4.0MB ± 0%    -79.63%  (p=0.008 n=5+5)
New/text=zero/size=1M/bits=32-12         19.7MB ± 0%       4.0MB ± 0%       ~     (p=0.079 n=4+5)
New/text=zero/size=1M/bits=64-12         39.0MB ± 0%       8.0MB ± 0%    -79.48%  (p=0.000 n=5+4)
New/text=zero/size=5M/bits=32-12         85.2MB ± 0%      20.0MB ± 0%    -76.52%  (p=0.008 n=5+5)
New/text=zero/size=5M/bits=64-12          169MB ± 0%        40MB ± 0%    -76.27%  (p=0.008 n=5+5)
New/text=zero/size=10M/bits=32-12         169MB ± 0%        40MB ± 0%    -76.26%  (p=0.000 n=5+4)
New/text=zero/size=10M/bits=64-12         333MB ± 0%        80MB ± 0%    -75.99%  (p=0.008 n=5+5)
New/text=zero/size=50M/bits=32-12         739MB ± 0%       200MB ± 0%    -72.93%  (p=0.000 n=4+5)
New/text=zero/size=50M/bits=64-12        1.63GB ± 0%      0.40GB ± 0%    -75.42%  (p=0.008 n=5+5)
New/text=rand/size=100K/bits=32-12        807kB ± 0%       401kB ± 0%    -50.25%  (p=0.008 n=5+5)
New/text=rand/size=100K/bits=64-12       1.61MB ± 0%      0.80MB ± 0%    -50.25%  (p=0.008 n=5+5)
New/text=rand/size=500K/bits=32-12       4.04MB ± 0%      2.01MB ± 0%       ~     (p=0.079 n=4+5)
New/text=rand/size=500K/bits=64-12       8.07MB ± 0%      4.01MB ± 0%       ~     (p=0.079 n=4+5)
New/text=rand/size=1M/bits=32-12         8.07MB ± 0%      4.01MB ± 0%    -50.36%  (p=0.000 n=5+4)
New/text=rand/size=1M/bits=64-12         16.1MB ± 0%       8.0MB ± 0%    -50.36%  (p=0.008 n=5+5)
New/text=rand/size=5M/bits=32-12         40.3MB ± 0%      20.0MB ± 0%    -50.35%  (p=0.029 n=4+4)
New/text=rand/size=5M/bits=64-12         80.7MB ± 0%      40.0MB ± 0%       ~     (p=0.079 n=4+5)
New/text=rand/size=10M/bits=32-12        80.7MB ± 0%      40.0MB ± 0%    -50.41%  (p=0.008 n=5+5)
New/text=rand/size=10M/bits=64-12         161MB ± 0%        80MB ± 0%    -50.44%  (p=0.029 n=4+4)
New/text=rand/size=50M/bits=32-12         403MB ± 0%       200MB ± 0%    -50.36%  (p=0.000 n=5+4)
New/text=rand/size=50M/bits=64-12         806MB ± 0%       400MB ± 0%       ~     (p=0.079 n=4+5)
SaveRestore/bits=32-12                   5.28MB ± 0%      5.28MB ± 0%       ~     (p=1.000 n=5+5)
SaveRestore/bits=64-12                   9.47MB ± 0%      9.47MB ± 0%       ~     (p=0.286 n=5+5)

https://perf.golang.org/search?q=upload:20190426.1

Fixes #15480.

Change-Id: I0790f6edf67f5a9c02b4462632b4942e0c37988b
Reviewed-on: https://go-review.googlesource.com/c/go/+/174100
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Eric Roshan-Eisner <edre@google.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
src/index/suffixarray/gen.go [new file with mode: 0644]
src/index/suffixarray/gen64.go [deleted file]
src/index/suffixarray/qsufsort.go [deleted file]
src/index/suffixarray/qsufsort64.go [deleted file]
src/index/suffixarray/sais.go [new file with mode: 0644]
src/index/suffixarray/sais2.go [new file with mode: 0644]
src/index/suffixarray/suffixarray.go
src/index/suffixarray/suffixarray_test.go