Tcl Source Code

View Ticket
Login
Ticket UUID: 1675258
Title: mem alloc performance on Linux
Type: Bug Created on: 2007-03-06 20:47:45
Submitter: dgp Assigned to: msofer
Subsystem: 41. Memory Allocation Severity:
Priority: 6 Last modified: 2009-05-01 08:25:13
Status: Open Closed by:
Resolution: None Closed on:
Version: None
Description:
Some tclbench benchmarks suggest
loss of [lappend] performance
in Tcl 8.5.  Normalized form:

000 VERSIONS:                             1:8.5a6 2:8.4.15  3:8.3.5
010 LIST concat EVAL/LAPPEND 2x10            2.34    1.00    0.93
011 LIST concat EVAL/LAPPEND 2x100           2.80    1.00    0.87
012 LIST concat EVAL/LAPPEND 2x1000          3.66    1.00    0.78
013 LIST concat EVAL/LAPPEND 2x10000         1.61    1.00    0.80
014 LIST concat FOREACH/LAPPEND 2x10         2.47    1.00    2.06
015 LIST concat FOREACH/LAPPEND 2x100        2.45    1.00    3.02
016 LIST concat FOREACH/LAPPEND 2x1000       2.64    1.00    3.17
017 LIST concat FOREACH/LAPPEND 2x10000      2.39    1.00    3.12

Original numbers:

010 LIST concat EVAL/LAPPEND 2x10           10.12     4.32     4.00
011 LIST concat EVAL/LAPPEND 2x100          19.43     6.93     6.00
012 LIST concat EVAL/LAPPEND 2x1000        107.59    29.39    23.00
013 LIST concat EVAL/LAPPEND 2x10000       620.62   386.44   311.00
014 LIST concat FOREACH/LAPPEND 2x10        10.79     4.37     9.00
015 LIST concat FOREACH/LAPPEND 2x100       55.14    22.49    68.00
016 LIST concat FOREACH/LAPPEND 2x1000     540.40   204.56   649.00
017 LIST concat FOREACH/LAPPEND 2x10000   4992.72  2088.81  6508.00


These slowdowns are not seen on Solaris.
User Comments:
nobody added on 2007-03-07 04:09:01:
Logged In: NO 

tclguyperhaps it's a LE vs. BE opt issue?

dgp added on 2007-03-07 23:06:16:
Logged In: YES 
user_id=80530
Originator: YES


another strange data point...

I see this slowdown only when I
run the tcl/list.bench file alone.

When I test the entire tclbench
benchmark suite, and then look at
these tests, these large performance
losses are not seen.

dgp added on 2007-03-07 23:13:12:
Logged In: YES 
user_id=80530
Originator: YES


ok, when running the whole
tclbench suite, it is the
ARRAY benchmarks that are slow.

I'm guessing that it's the first
tests that allocate significant
memory that see the slowdown, and
it's memory allocation that has
actually lost performance.

dgp added on 2007-03-22 03:53:50:
Logged In: YES 
user_id=80530
Originator: YES


There seems to be a common thread
in these benchmarks of

  for .... {
    lappend l ....
  }

Have we somehow lost the efficiency
of our growth algorithm?  Is it
related to the new List intrep?

dgp added on 2007-03-22 03:58:23:
Logged In: YES 
user_id=80530
Originator: YES


I see that [lappend] is now
implemented via a call to
Tcl_ListObjReplace() while it
used to be done via direct
intrep surgery.

msofer added on 2007-03-22 04:40:29:
Logged In: YES 
user_id=148712
Originator: NO

EVAL/LAPPEND performance might be due to the Tcl_ListObjReplace scenic tour, but FOREACH/LAPPEND not: the lappend bcc implementation did not change and does not go through T_LOR: it goes through Tcl_ListObjAppendElement (via TclPtrSetVar) just like in 8.4. I do not immediately see that T_LOAE could have gotten slower because of the new list intrep.

(Sidenote: EVAL/LAPPEND is not idiomatic for 8.5: one would use {*} instead of [eval]. A new benchmark EXPAND/LAPPEND might be useful).

msofer added on 2007-03-22 05:25:04:
Logged In: YES 
user_id=148712
Originator: NO

There seems to be a problem where the first interp to run pays some kind of penalty. Benchmarking two copies of the same executable I get huge diffs! 

mig@ice:/home/CVS/tclbench$ ./runbench.tcl -verbose -paths /home/CVS/tcl-core-8-4-branch/unix/ tcl/array.bench TCL_INTERP: 1:8.4.15 2:8.4.15
STARTED 2007-03-21 19:22:49 (runbench.tcl v1.20)
Benchmark 1:8.4.15 /home/CVS/tcl-core-8-4-branch/unix/tclsh
a 00:00:03 elapsed
3161 milliseconds
Benchmark 2:8.4.15 /home/CVS/tcl-core-8-4-branch/unix/tclsh1
a 00:00:03 elapsed
2520 milliseconds
000 VERSIONS:             1:8.4.15 2:8.4.15
001 ARRAY genKeys 50       1391.59  1153.22
002 ARRAY genKeys 500      2880.64  1418.72
003 ARRAY makeHash 500 50 57144.28 47765.00
003 BENCHMARKS            1:8.4.15 2:8.4.15
FINISHED 2007-03-21 19:22:55

msofer added on 2007-03-22 06:04:40:
Logged In: YES 
user_id=148712
Originator: NO

More data points: seems to really be a linux issue?

* http://paste.tclers.tk/180 shows 8.5a6 under linux (with strange timings, consistent accross reruns, same thing for tclkit)

* http://paste.tclers.tk/178 (tclkitsh) and http://paste.tclers.tk/179 (tclsh) show that there is no such weirdness under WinXP.

dgp added on 2007-03-22 22:51:24:
Logged In: YES 
user_id=80530
Originator: YES


I confirm that on 

Linux <REDACTED> 2.6.19-1.2911.6.5.fc6 #1 SMP <DATE> i686 athlon i386 GNU/Linux

the first identical copy runs slower
than the second.

dgp added on 2007-11-28 01:02:18:
Logged In: YES 
user_id=80530
Originator: YES


did we ever puzzle this out?

is it still true on the HEAD?

hobbs added on 2007-11-28 01:25:32:
Logged In: YES 
user_id=72656
Originator: NO

It seems true for FOREACH/LAPPEND still, but not as severe as the original report.