3 Dec
2009
3 Dec
'09
10:33 p.m.
On Dec 3, 2009, at 10:27 PM, Nicolas Cellier wrote:
I experimented a quick hack for StandardFileStream
BufferedFileStream
-> 5x to 7x on micro-benchmarks
I stopped developing it because working on current Stream hierarchy
just is horrorfull !
Yes this is why damien worked on Nile.
Like for example an inst var change of semantic in a
subclass, or is
just not used (try upToAnyOf: in various streams to see what I
mean)...
Disgusting...
Argh.
We saw some cool stuff too like passing block to simulate exception (I can understand why
because
exception were not there but this was not that clean)
I prefer a full rewrite, and it is under development
at
http://www.squeaksource.com/XTream
It is read/write to anyone and MIT.
Yet i have Xtreams on collections and readOnly files passing some
tests (only 40% coverage).
It tries to gather lots of ideas popping regularly in various flavours
- squeak/vw/gst (lazy streams, Nile, VW Xtreams, generators, etc...).
Excellent.
I was not really convinced by the composition of streams in Nile.
Even if it was ok. I thin kthat this is not about Nile in particular more about the filter
approach.
I do not know about stream composition - I like it when I was programming in functional
languages but in OOP I often
feel it a bit overkill. Anyway we will learn. May be because we do not program a lot with
stream
anymore.
The real bottleneck in Squeak/Pharo is
MultiByteFileStream, and it is
a piece of code I recommend for teaching how to not program in
Smalltalk.
:)
I won't spend much time to change it any more,
just replace and throw
away will now be my policy.
Excellent!
I love that policy. :)
The Buffered trick should also be ported to Nile, it
would be then
interesting to compare complexity and efficiency of both libraries.
Our main problem right now is man power not wish and dreams.
Cheers
Nicolas
2009/12/3 Stéphane Ducasse <stephane.ducasse(a)inria.fr>fr>:
> hi guys
>
> Apparently nicolas cellier is implementing faster stream so this would probably help
moose
> loading huge models.
> Nicolas could you let us know when you have a version that we can try. We could come
up
> with real bench on really large files.
>
> Cyrille it could be really worth that you have a look. There is a thread on
squeak-dev
>
> Stef