| 1 |
Am 07. Nov 2014, 19:30 schrieb Ciaran McCreesh <ciaran.mccreesh@××××××××××.com>: |
| 2 |
|
| 3 |
> On Fri, 07 Nov 2014 19:11:04 +0100 |
| 4 |
> Jauhien Piatlicki <jauhien@g.o> wrote: |
| 5 |
>> Then the same question for you: where can one read about the algorithm |
| 6 |
>> Paludis uses? |
| 7 |
> |
| 8 |
> It's basically a two stage process: simple constraint solving using |
| 9 |
> value ordering heuristics to enforce "don't do unnecessary work", then |
| 10 |
> ordering (which is not quite a graph process, and which is not as |
| 11 |
> simple as a topological sort, because the tree is full of circular |
| 12 |
> dependencies). |
| 13 |
> |
| 14 |
> But the interesting question isn't "what's the algorithm?", it's |
| 15 |
> "what's the model?". That's where the complexity lies: figuring out how |
| 16 |
> to turn *DEPEND specifications into constraints is an utter pain, and |
| 17 |
> it isn't clean or easily understandable. The primary reason is || |
| 18 |
> dependencies: developers like to write not-really-correct and utterly |
| 19 |
> unobvious dependency strings rather than asking for new syntax so they |
| 20 |
> can just say what they mean... |
| 21 |
|
| 22 |
|
| 23 |
Currently, for portage just to decide that nothing has to be done on my |
| 24 |
machine takes around 1 minute. |
| 25 |
|
| 26 |
What is in your opinion the main reason for this? And how can we knock |
| 27 |
this down to reasonable speed? |
| 28 |
|
| 29 |
- Is our dependency model that more complex than the problem resolvers |
| 30 |
of other package managers for other distributions solve? |
| 31 |
|
| 32 |
- Is it the algorithm that is implemented for the dependency model? |
| 33 |
|
| 34 |
- Is it its implementation? |
| 35 |
|
| 36 |
|
| 37 |
>> And, again, I have herd (did not try myself) that Paludis is as slow |
| 38 |
>> as Portage. |
| 39 |
> |
| 40 |
> Well, you're not comparing like with like. Paludis with "everything |
| 41 |
> turned off" does more than Portage with "everything turned on". If all |
| 42 |
> you're looking for is the wrong answer as fast as possible, there are |
| 43 |
> easier ways of getting it... |
| 44 |
|
| 45 |
The last time I compared the resolver speed of portage and paludis both |
| 46 |
needed almost the same time. |
| 47 |
|
| 48 |
Do you have a speed comparison with a similar feature set of both? (Or, |
| 49 |
alternatively, the speedup one gains by tuning paludis to be as fast as |
| 50 |
possible). |
| 51 |
|
| 52 |
Best, |
| 53 |
Matthias |
Archives