Interpreted case/when is much slower than MRI
---------------------------------------------
Key: JRUBY-2984
URL: http://jira.codehaus.org/browse/JRUBY-2984
Project: JRuby
Issue Type: Bug
Components: Performance
Affects Versions: JRuby 1.1.4
Reporter: Charles Oliver Nutter
Fix For: JRuby 1.1.5
Ouch...the interpreter seems to have *serious* performance issues when it comes
to case/when statements:
Note this is only numbers for the 10-when case in this benchmark
{noformat}
➔ jruby --server -X-C bench_case.rb
user system total
real
1m x5 cases with 10 whens 16.053000 0.000000 16.053000 (
16.053354)
user system total
real
1m x5 cases with 10 whens 24.339000 0.000000 24.339000 (
24.338982)
user system total
real
1m x5 cases with 10 whens 26.326000 0.000000 26.326000 (
26.326204)
user system total
real
1m x5 cases with 10 whens 28.603000 0.000000 28.603000 (
28.602653)
user system total
real
1m x5 cases with 10 whens 32.623000 0.000000 32.623000 (
32.623010)
~/NetBeansProjects/jruby ➔ /opt/local/bin/ruby bench_case.rb
user system total
real
1m x5 cases with 10 whens 9.070000 0.090000 9.160000 (
9.344732)
user system total
real
1m x5 cases with 10 whens 9.080000 0.090000 9.170000 (
9.595736)
user system total
real
1m x5 cases with 10 whens 10.880000 0.140000 11.020000 (
11.791109)
user system total
real
1m x5 cases with 10 whens 11.310000 0.130000 11.440000 (
12.018376)
user system total
real
1m x5 cases with 10 whens 10.100000 0.120000 10.220000 (
10.626142)
{noformat}
Note that in this case, where all the cases are fixnums, Ruby 1.9 is performing
a direct jump. Experimental code for doing the same in JRuby produces an
extremely large boost in performance:
{noformat}
Ruby 1.9:
~/NetBeansProjects/jruby ➔ ../ruby1.9/ruby -I ../ruby1.9/lib
bench_case.rb
user system total
real
1m x5 cases with 10 whens 0.400000 0.000000 0.400000 (
0.415587)
user system total
real
1m x5 cases with 10 whens 0.400000 0.010000 0.410000 (
0.409474)
user system total
real
1m x5 cases with 10 whens 0.400000 0.000000 0.400000 (
0.410365)
user system total
real
1m x5 cases with 10 whens 0.410000 0.000000 0.410000 (
0.414840)
user system total
real
1m x5 cases with 10 whens 0.400000 0.000000 0.400000 (
0.409613)
JRuby:
~/NetBeansProjects/jruby ➔ jruby --server bench_case.rb
user system total
real
1m x5 cases with 10 whens 0.430000 0.000000 0.430000 (
0.429635)
user system total
real
1m x5 cases with 10 whens 0.301000 0.000000 0.301000 (
0.300480)
user system total
real
1m x5 cases with 10 whens 0.235000 0.000000 0.235000 (
0.235214)
user system total
real
1m x5 cases with 10 whens 0.198000 0.000000 0.198000 (
0.198281)
user system total
real
1m x5 cases with 10 whens 0.200000 0.000000 0.200000 (
0.199393)
{noformat}
Compiled performance, even without the switch optimization, is substantially
better than the interpreter, but nowhere near 1.9 performance:
{noformat}
➔ ../jruby-1_1_3/bin/jruby --server bench_case.rb
user system total
real
1m x5 cases with 10 whens 7.026000 0.000000 7.026000 (
7.026314)
user system total
real
1m x5 cases with 10 whens 4.693000 0.000000 4.693000 (
4.692950)
user system total
real
1m x5 cases with 10 whens 4.647000 0.000000 4.647000 (
4.646249)
user system total
real
1m x5 cases with 10 whens 4.674000 0.000000 4.674000 (
4.674111)
user system total
real
1m x5 cases with 10 whens 4.603000 0.000000 4.603000 (
4.603245)
{noformat}
We should also include benchmarks for all literals, to see where 1.9 is doing
optimizations we can mimic.
--
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators:
http://jira.codehaus.org/secure/Administrators.jspa
-
For more information on JIRA, see: http://www.atlassian.com/software/jira
---------------------------------------------------------------------
To unsubscribe from this list, please visit:
http://xircles.codehaus.org/manage_email
