On Fri, Mar 8, 2013 at 11:44 AM, Sven Van Caekenberghe <sven@stfx.eu> wrote:
On 08 Mar 2013, at 13:09, Camillo Bruni <camillobruni@gmail.com> wrote:
On 2013-03-08, at 11:53, Sven Van Caekenberghe <sven@stfx.eu> wrote:
On 08 Mar 2013, at 11:16, Sven Van Caekenberghe <sven@stfx.eu> wrote:
But I have said this before: the wall clock time of loading a lot of code is actually close to unacceptable - I don't think it is the download or the compilation, but more all the dynamic stuff that happens after that. There should be a way to not do all those updates for each method and move the updates to one big batch update after the load - if that is possible.
To continue my rant (sorry ;-) about the problem with slow code loading.
These are some benchmarks on the same machine:
$ ./vm.sh experimental.image eval '[Smalltalk allClassesAndTraits do: #compileAll] timeToRun' 106532
$ ./stack/vm.sh experimental.image eval '[Smalltalk allClassesAndTraits do: #compileAll] timeToRun' 221708
So it takes like 3 minutes to recompile every method in the system.
How in the hell can it take 40 minutes to load some code (with all packages already present in the package-cache (but then again the package-cache is only 3.5 Mb, which could be downloaded in seconds)) ?
I with you on this one! :D I always wondered what Metacello and Compiler do there! :D
Well, open the Time Profiler and load your favourite Config. You'll have a hard time finding the actual time slice of the compiler. As far as I could see in my run it was less than 1% ! All the rest MC, Gofer, RPackage and SystemAnnouncement stuff, even completion stuff.
Indeed, we noticed this as well when benchmarking Tanker to replace source compilation in Monticello. We got to a similar conclusion. Quite a challenge to optimize if you ask me.
- Batch notifications: do not do a notification per method/class added. Instead do one notification with all the methods/classes together. -- Mariano http://marianopeck.wordpress.com