Re: [abc] PLDI paper

One thing that's strange is that on lod-sim, we are much worse than we
were in abc 1.0.0, in all the configurations. I just checked to make
sure those numbers were correct, and they are. I wonder what we messed
up since 1.0.0...

Ondrej

On Sat, Apr 09, 2005 at 04:01:37PM -0400, Prof. Laurie HENDREN wrote:
> Ondrej,
>
> Those numbers look good. I think those are the categories we want.
> Since threadlocal is always better, I think we should just go with
> those numbers (it is also our default).
>
> I don't remember if ajc1.2 uses ThreadLocal or not. Do you have
> a copy of the ajc 1.2 runtime lib to check? I'm pretty sure that
> ajc1.2.1 does use ThreadLocal (but not with my caching trick).
>
> Can you do ajc 1.2 and 1.2.1 (with lib
> runtime library for each) so we can compare those too. Looking
> at the numbers from the paper, I think we are doing well on
> everything but Cona-sim ... I wonder why we are slower there?
>
> Cheers, Laurie
>
> +-------------------------------------------------------------+
> | Laurie Hendren, Professor, School of Computer Science |
> | McGill University |
> | 318 McConnell Engineering Building tel: (514) 398-7391 |
> | 3480 University Street fax: (514) 398-3883 |
> | Montreal, Quebec H3A 2A7 hendren@cs.mcgill.ca |
> | CANADA http://www.sable.mcgill.ca/~hendren |
> | http://wwww.sable.mcgill.ca http://aspectbench.org |
> +-------------------------------------------------------------+
>
> On Sat, 9 Apr 2005, Ondrej Lhotak wrote:
>
> > Here are the intraprocedural numbers.
> >
> > noopt means no intraproc optimizations
> > sharing means sharing only
> > shar+cnt means sharing and using counters
> > all3 means sharing, using counters, and -cflow-share-thread-locals
> >
> > "new" (not threadlocal)
> > noopt sharing shar+cnt all3
> > figure 1219.82 285.22 139.29 24.56
> > quicksort 122.02 75.58 28.39 27.51
> > sablecc 30.52 30.11 23.83 22.72
> > ants 21.25 21.31 21.21 18.73
> > lod-sim 1741.00 50.35 34.75 26.58
> > lod-weka 1495.65 155.02 102.33 74.65
> > conastack 689.40 91.99 52.60 29.07
> > conasim 76.11 75.56 73.49 72.11
> >
> > "threadlocal"
> > noopt sharing shar+cnt all3
> > figure 1074.36 238.68 90.26 20.35
> > quicksort 122.06 75.08 27.86 27.41
> > sablecc 29.65 29.75 23.31 23.05
> > ants 19.03 19.27 18.96 17.95
> > lod-sim 1723.44 46.47 32.92 25.99
> > lod-weka 1351.21 143.32 92.20 74.84
> > conastack 594.46 80.00 41.13 27.24
> > conasim 75.54 75.34 73.54 71.95
> >
> > The times are just from one run each, for now, not average-of-5.
> >
> > Ondrej
> >
> > On Sat, Apr 09, 2005 at 02:12:07PM -0400, Prof. Laurie HENDREN wrote:
> > > I have been working on the format (ACM has a new conference proceedings
> > > style) and the list of things we need to look at for the final version.
> > > Ondrej and I discussed the experiments to do for intraproc cflow opts -
> > > now that there are more of them, and he is working on those.
> > >
> > > In the pldi05abc/ svn directory, the final version we will submit is
> > > final.tex. I think most things with the new format have been resolved and
> > > I have added the general terms, keywords and so on. We can add more/different
> > > keywords if we want.
> > >
> > > Also, in that directory is TODO.txt, which I include here:
> > >
> > > ----------- TODO.txt
> > >
> > > 1) Run around benchmarks -
> > > are there some other variations we can do on the inlining?
> > >
> > > 2) Run intraproc cflow benchmarks -
> > > decide if we should show both versions of handling threadlocals?
> > >
> > > 3) Run interproc cflow benchmarks
> > >
> > > 4) Check if performance difference between ajc 1.2 and 1.2.1 is due to
> > > cflow counters only
> > >
> > > In reality, there could be several effects:
> > >
> > > - using counters
> > > - better sharing of stacks/counters
> > > - not introducing extra copies of counters
> > > (may have had a bug in 1.2 - seemed to introduce too many
> > > counters)
> > > - improvements in runtime lib (i.e. ThreadLocal ... did that come
> > > between 1.2 and 1.2.1?)
> > > - some other optimization introduced in ajc between 1.2 and 1.2.½
> > >
> > > 5) Fix bib entries - make sure all titles have caps in right places and
> > > page numbers.
> > >
> > > 6) Change figure for reweaving and perhaps expand upon that section (one
> > > reviewer wanted more description of reweaving)
> > >
> > > 7) Decide if we should mention "other optimizations" or not. If so, fix them
> > > up.
> > >
> > > 8) Damien mentioned a variation of our inter-proc algorithm ... Damien ...
> > > what was it - something about not needed something on "all paths"?
> > >
> > > 9) Formatting of figures and tables. Does everything look ok. Should we put
> > > some of the code segments into figures?
> > >
> > > -----------
> > >
> > > Cheers, Laurie
> > >
> > >
> > >
> > >
> > >
> > > +-------------------------------------------------------------+
> > > | Laurie Hendren, Professor, School of Computer Science |
> > > | McGill University |
> > > | 318 McConnell Engineering Building tel: (514) 398-7391 |
> > > | 3480 University Street fax: (514) 398-3883 |
> > > | Montreal, Quebec H3A 2A7 hendren@cs.mcgill.ca |
> > > | CANADA http://www.sable.mcgill.ca/~hendren |
> > > | http://wwww.sable.mcgill.ca http://aspectbench.org |
> > > +-------------------------------------------------------------+
> > >
> > >
> >
>
>
Received on Sat Apr 9 23:41:01 2005