JRuby

break performance is slower than MRI, sometimes slower than interpreted

Details

  • Type: Bug Bug
  • Status: Closed Closed
  • Priority: Major Major
  • Resolution: Fixed
  • Affects Version/s: JRuby 1.1RC2
  • Fix Version/s: JRuby 1.2
  • Component/s: Compiler
  • Labels:
    None
  • Number of attachments :
    0

Description

Ruby:
control, five short while loops  0.140000   0.000000   0.140000 (  0.145039)
control, five whiles with blocks  0.320000   0.000000   0.320000 (  0.318974)
five whiles that break          0.210000   0.000000   0.210000 (  0.239146)
control, five whiles with blocks  0.430000   0.000000   0.430000 (  0.431001)

JRuby interpreted:
control, five short while loops  0.156000   0.000000   0.156000 (  0.155000)
control, five whiles with blocks  0.331000   0.000000   0.331000 (  0.332000)
five whiles that break          0.457000   0.000000   0.457000 (  0.457000)
control, five whiles with blocks  1.027000   0.000000   1.027000 (  1.027000)

JRuby compiled:
control, five short while loops  0.050000   0.000000   0.050000 (  0.049000)
control, five whiles with blocks  0.269000   0.000000   0.269000 (  0.269000)
five whiles that break          0.047000   0.000000   0.047000 (  0.046000)
control, five whiles with blocks  0.782000   0.000000   0.782000 (  0.782000)

Activity

Ascending order - Click to sort in descending order
Hide
Permalink
Charles Oliver Nutter added a comment -

I read these numbers wrong; the compiled local break performance is 0.046, around 5 times faster than MRI. The non-local break performance is still slower, however, so I'll look into that.

Show
Charles Oliver Nutter added a comment - I read these numbers wrong; the compiled local break performance is 0.046, around 5 times faster than MRI. The non-local break performance is still slower, however, so I'll look into that.
Hide
Permalink
Charles Oliver Nutter added a comment -

We'll look to fix this after 1.1, since it's not really a huge issue.

Show
Charles Oliver Nutter added a comment - We'll look to fix this after 1.1, since it's not really a huge issue.
Charles Oliver Nutter made changes -
Field Original Value New Value
Fix Version/s JRuby 1.1+ [ 13421 ]
Fix Version/s JRuby 1.1 [ 13990 ]
Hide
Permalink
Charles Oliver Nutter added a comment -

Here's updated numbers; still slower than MRI for non-local breaks, but it has improved in the compiled case. Interpreted still seems to have some performance issues:

jruby compiled:
                                              user     system      total        real
control, five short while loops           0.023000   0.000000   0.023000 (  0.022774)
control, five whiles with blocks          0.276000   0.000000   0.276000 (  0.276131)
five whiles that break                    0.025000   0.000000   0.025000 (  0.024213)
five whiles with blocks that break        0.593000   0.000000   0.593000 (  0.593634)
jruby interpreted:
                                              user     system      total        real
control, five short while loops           0.078000   0.000000   0.078000 (  0.077959)
control, five whiles with blocks          0.256000   0.000000   0.256000 (  0.255260)
five whiles that break                    0.299000   0.000000   0.299000 (  0.299437)
five whiles with blocks that break        1.112000   0.000000   1.112000 (  1.112932)
MRI:
                                              user     system      total        real
control, five short while loops           0.150000   0.000000   0.150000 (  0.162975)
control, five whiles with blocks          0.350000   0.010000   0.360000 (  0.353873)
five whiles that break                    0.230000   0.000000   0.230000 (  0.234755)
five whiles with blocks that break        0.460000   0.000000   0.460000 (  0.481575)

The bad case at this point is really just the while loop with a breaking block. Still an issue, but better.

Show
Charles Oliver Nutter added a comment - Here's updated numbers; still slower than MRI for non-local breaks, but it has improved in the compiled case. Interpreted still seems to have some performance issues: jruby compiled: user system total real control, five short while loops 0.023000 0.000000 0.023000 ( 0.022774) control, five whiles with blocks 0.276000 0.000000 0.276000 ( 0.276131) five whiles that break 0.025000 0.000000 0.025000 ( 0.024213) five whiles with blocks that break 0.593000 0.000000 0.593000 ( 0.593634) jruby interpreted: user system total real control, five short while loops 0.078000 0.000000 0.078000 ( 0.077959) control, five whiles with blocks 0.256000 0.000000 0.256000 ( 0.255260) five whiles that break 0.299000 0.000000 0.299000 ( 0.299437) five whiles with blocks that break 1.112000 0.000000 1.112000 ( 1.112932) MRI: user system total real control, five short while loops 0.150000 0.000000 0.150000 ( 0.162975) control, five whiles with blocks 0.350000 0.010000 0.360000 ( 0.353873) five whiles that break 0.230000 0.000000 0.230000 ( 0.234755) five whiles with blocks that break 0.460000 0.000000 0.460000 ( 0.481575) The bad case at this point is really just the while loop with a breaking block. Still an issue, but better.
Charles Oliver Nutter made changes -
Comment [ I made the following changes, which appear to help quite a bit:

- Removed synchronization in searchMethod. Because it's only doing a get, and the methods hash is now a ConcurrentHashMap, this should be safe. Confirmation from others would be appreciated.
- Modified the methods map load factor to 0.75 from 2.0, so the hash is more likely to have only one entry in a given bucket.

These combined brought the numbers much closer but not quite down to MRI's times (best times shown):

{noformat}
MRI:
control, 1 and :to_ary 0.180000 0.010000 0.190000 ( 0.178157)
1m 1.respond_to?(:to_ary) 0.380000 0.000000 0.380000 ( 0.481805)

JRuby:
control, 1 and :to_ary 0.150000 0.000000 0.150000 ( 0.150000)
1m 1.respond_to?(:to_ary) 0.500000 0.000000 0.500000 ( 0.500000)
{noformat}

I'll leave this bug open, since we know MRI and IronRuby have a separate cache for respond_to?, and we're still about 2x slower than MRI. ]
Charles Oliver Nutter made changes -
Comment [ Oops, that last comment was for bug JRUBY-2181. ]
Hide
Permalink
Charles Oliver Nutter added a comment -

As of 1.2, this seems to be a non-issue in the compiler:

                                              user     system      total        real
control, five short while loops           0.010000   0.000000   0.010000 (  0.009846)
control, five whiles with blocks          0.124000   0.000000   0.124000 (  0.124394)
five whiles that break                    0.011000   0.000000   0.011000 (  0.010870)
five whiles with blocks that break        0.328000   0.000000   0.328000 (  0.328391)

And pretty good in the interpreter:

                                              user     system      total        real
control, five short while loops           0.043000   0.000000   0.043000 (  0.042733)
control, five whiles with blocks          0.157000   0.000000   0.157000 (  0.156774)
five whiles that break                    0.171000   0.000000   0.171000 (  0.171438)
five whiles with blocks that break        0.627000   0.000000   0.627000 (  0.627163)

When compared to MRI:

                                              user     system      total        real
control, five short while loops           0.120000   0.000000   0.120000 (  0.115604)
control, five whiles with blocks          0.250000   0.000000   0.250000 (  0.253488)
five whiles that break                    0.170000   0.000000   0.170000 (  0.212932)
five whiles with blocks that break        0.360000   0.000000   0.360000 (  0.365677)

We'll call this resolved and look for future performance reports on specific real-world cases.

Show
Charles Oliver Nutter added a comment - As of 1.2, this seems to be a non-issue in the compiler: user system total real control, five short while loops 0.010000 0.000000 0.010000 ( 0.009846) control, five whiles with blocks 0.124000 0.000000 0.124000 ( 0.124394) five whiles that break 0.011000 0.000000 0.011000 ( 0.010870) five whiles with blocks that break 0.328000 0.000000 0.328000 ( 0.328391) And pretty good in the interpreter: user system total real control, five short while loops 0.043000 0.000000 0.043000 ( 0.042733) control, five whiles with blocks 0.157000 0.000000 0.157000 ( 0.156774) five whiles that break 0.171000 0.000000 0.171000 ( 0.171438) five whiles with blocks that break 0.627000 0.000000 0.627000 ( 0.627163) When compared to MRI: user system total real control, five short while loops 0.120000 0.000000 0.120000 ( 0.115604) control, five whiles with blocks 0.250000 0.000000 0.250000 ( 0.253488) five whiles that break 0.170000 0.000000 0.170000 ( 0.212932) five whiles with blocks that break 0.360000 0.000000 0.360000 ( 0.365677) We'll call this resolved and look for future performance reports on specific real-world cases.
Charles Oliver Nutter made changes -
Status Open [ 1 ] Resolved [ 5 ]
Resolution Fixed [ 1 ]
Fix Version/s JRuby 1.2 [ 14757 ]
Fix Version/s JRuby 1.x+ [ 13421 ]
Charles Oliver Nutter made changes -
Status Resolved [ 5 ] Closed [ 6 ]

People

  • Assignee:
    Charles Oliver Nutter
    Reporter:
    Charles Oliver Nutter
Vote (0)
Watch (1)

Dates

  • Created:
    Updated:
    Resolved: