Re: [abc] oopsla outline

Oege de Moor wrote:

> 3. Experiments

I've committed some more stuff into experiments.tex. More numbers for
figure 7, and preliminary graphs for JHotDraw (Oege, let me know if
those weren't the styles of graph you were expecting).

The image quality isn't brilliant at the moment, but I don't want to
play with it until we've finalised the graphs; there's a .ods document
in Figures/ which contains the graphs if you're interested.
Alternatively, looking at the eps files should be more legible than
looking at tex's output.

A word about the PQL variations: PQL1 is the 'obvious' PQL pattern, PQL2
is the 'obvious' pattern, except that it prohibits a call_next between
the update_source and final call_next events, i.e. it's

create_enum
update_source
~call_next
call_next

This variation is by Julian; the point is that with the negated event,
many partial matches could be thrown away that otherwise would have to
be kept around. The numbers of course suggest that PQL doesn't make use
of that fact.

PQLFull is PQL2, except the 'update_source' is specified as a full
disjunction of pattern names (i.e. "ds.remove(_) | ds.removeAll() |
..."), where PQL2 uses name patterns (i.e. "ds.remove*(...) | ..."). I'm
not sure if it's worth including this. The point is that PQL takes a
*lot* more time and memory if you spell out the disjunction in full,
presumably because it ends up creating a bigger automaton. Perhaps a
point in favour of trying to reduce automaton size?

There's another bit of bad PQL-related news. Julian spent yesterday and
today trying to get DBPooling to work; it seems like it isn't going to
happen, due to bugs in PQL which we don't feel we have the time to track
down. If anyone is at a loss for what to do, by all means ask me or
Julian about this. It's of course a shame that there's only three
benchmarks where we can compare to PQL (JHotDraw, AJHotDraw and
Jigsaw/LocalUL), but that can't be helped.

Eric is currently gathering numbers for hashcode/aprove, and I'm hoping
to check in numbers for hashcode/weka and jigsaw/lor by tomorrow
morning. Then it's time to flesh out the paper!

- P
Received on Thu Mar 09 20:06:56 2006