--- Begin Message ---
Subject: |
poor performance since grep 2.19 when comparing files with grep |
Date: |
Mon, 26 Oct 2015 12:54:01 +0000 |
Apologies in advance if this is more of a "discuss" question, but it looks like
a particular use-case shows a marked change in performance between recent
versions of grep.
A colleague mentioned a performance issue with grep to me, and its puzzling me
a bit.
It turns out that he was using "grep -Fvif" to find lines in one file that are
not present in another.
Up until grep 2.18 this seems to work with linear performance and it takes less
than 50ms to compare files up to about 20,000 lines.
With grep 2.19 and later, ever relatively small files are quite slow, runtime
(and memory use) increases exponentially (e.g. 300ms to compare 200 lines, 1.5s
to compare 400 lines, 5s to compare 600 lines).
I've shown my colleague how to use sort and diff (and "comm", which I think is
vastly underrated), but it made me wonder if this is a reasonable thing to
expect grep to be able to do, and whether such a performance drop should be
seen as a bug.
The way he was using it, he had two (unsorted) data sets (about 6000 rows in
each), with most lines being common, and he was just using:
grep -Fvif FILE1 FILE2
In his case, the older version of grep took way less than a second to run, but
after he had upgraded his machine it took 20 minutes before running out of swap
and seg faulting.
In terms of comparing performance, I've found that the following works to
compare performance (vary N to try different sized data files):
N=600; F=/tmp/zz.$$; seq -f '%g bottles of beer on the wall' 1 $N > $F;
time grep -Fvif $F $F; rm $F
Steve.
--- End Message ---
--- Begin Message ---
Subject: |
Re: bug#22357: grep -f not only huge memory usage, but also huge time cost |
Date: |
Tue, 20 Dec 2016 21:17:01 -0800 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.5.1 |
I installed the attached patches into grep master. These fix the performance
regressions noted at the start of Bug#22357. I see that the related performance
problems noted in Bug#21763 seem to be fixed too, I expect because of Norihiro
Tanaka's recent changes, so I'll boldly close both bug reports.
To some extent the attached patches restore the old behavior for grep -F, when
grep is given two or more patterns. The patch doesn't change the underlying
algorithms; it merely uses a different heuristic to decide whether to use the -F
matcher. Although I wouldn't be surprised if the attached patches hurt
performance in some cases, I didn't uncover any such cases in my performance
testing, which I admit mostly consisted of running the examples in the
abovementioned bug reports.
I'll leave Bug#22239 open, as I get the following performance figures
(user+system CPU time) for the Bug#22239 benchmark, where list.txt is created by
"aspell dump master | head -n 100000 >list.txt", and the grep commands all use
the operands "-F -f list.txt /etc/passwd" in the en_US.utf8 locale on Fedora 24
x86-64.
no -i -i grep version
0.25 0.33 2.16
0.26 10.95 2.21
0.11 2.90* current master (including attached patches)
In the C locale, the current grep master is always significantly faster than
grep 2.16 or 2.21 on the benchmark, so the only significant problem is the
number marked "*". I ran the benchmarks on an AMD Phenom II X4 910e.
0001-grep-simplify-line-counting-in-patterns.patch
Description: Text Data
0002-grep-simplify-matcher-configuration.patch
Description: Text Data
0003-grep-fix-performance-with-multiple-patterns.patch
Description: Text Data
--- End Message ---