For all details and references of the project one can visit darcs-GSoC-2014-History-Reordering-Performance-and-Features.

martes, 19 de agosto de 2014

Last Few Weeks

In the last three weeks I was working in a couple of things, unfortunately I couldn't complete any to 100% :

- Finishing the command minimize-context.
- Implement the "show dependencies" for darcsden.
- Solve the issue2405.

Finish the command minimize-context

I can start telling that the command itself is implemented, but making tests I find out a case when making minimize-context fails to update correctly the files of repository. I comment the problem in here, however here goes a little summary.

Some "preconditions" to take into account:
1. If exist a tag, I search for dependencies in the patchs after the tag.
2. If happends that not exist dependencies after the tag, the only patch in the context of the bundle to send is the tag. This helps the command darcs apply to merge the patches and seems good to have at least a "tag patch" dependency.
3. (In relation with 1) If not exist a tag, the search is made in all the repository. One problem here is that with repositories that have huge amount of patches (darcs for example with 11000~ patches) the command maybe not finish of calculate the dependencies. Hopefully, having so many patches without tags seems a little odd. Nevertheless, I suppose that even the search in 700~ patches of 10~ patches to send should have a decent performance.

Well, passing to comment the problem. In some older post I mention the way we calculate the dependencies, so with that in mind, suppose we have to repositories $r_1$, $r_2$ with the following patches, where $m_x$ and $a_x$ represents the $x$-th modification and adding respectively.

$r_1$,$r_2$ = $m_2$ $tag$ $m_1$ $a_1$

suppose now that $r_2$ makes a third modification $m_3$ that adds lines without touch any existing line in the only file adding for $a_1$ and sequentially modified for $m_x$. With the idea of send a bundle to $r_1$. Now if we compute the dependencies to find out what is the context to send. For (1) we need to try to commute $m_2$ with $m_3$, which success but of course the end result is a $m'_3$ and a $m'_2$ because the lines which modifies every one now are different. But beyond that for our definition of dependency we can throw $m_2$, so the final context of the bundle with minimize context is the patch $tag$.

The problem now is that if we make darcs apply of the minimized bundle the merge is made not considering $m_2$ and the final state of the file isn't right. For example, if we are talking about that $m_3$ adds a function in some empty place, this could end in $m_3$ adding the function in the middle of an existing function.

My, more or less, mature thought is something like have Direct Dependency and  Indirect Dependency (the names could be differents $\smile$). The Direct Dependency is a dependency as already we know it and the Indirect Dependency is that if we have,

$p_1$ and $p_2$ patches, and $p'_1$ and $p'_2$  are the results of commute the patches, then if $p_1 \ne p'_1$ we conclude that $p_1$ indirectly depends of $p_2$.

I have almost implemented this idea, but I change task and still missing a couple of details. In particular, I'm not so sure of a couple of things,

- How to correctly compare the patches ($p_1 \ne p'_1$)
- In my unclean implementation I have to carry on $MyEq$ for all over the code and it's very ugly I think.

Finishing with this topic, here is a script (improved by Guillaume) that shows the problem:

Implement the "show dependencies" for darcsden

An idea that quickly comes up to my mind is that, considering the last part of the previous topic, since we have two types of dependencies it would be nice generate the graph of dependencies with two types of arrows to differentiate the type of dependency between patches. Talking about the implementation itself, I advanced in the drawing of the graph and some different presentations, but very little about the integration with darcsden, which I think is more or less direct having implemented the $js$ that draw the graph. Would have been nice have more time for finish this but I swap to the task of the following topic.

I hope be able to work in the last two topics in the next few months outside of the GSoC.

Solve the issue2405

I would like to start saying, "what a shame...". I can't solve the issue, I improve a little bit the memory but nothing too much significant. Maybe the good news is that, more o less, I found the "place" in the code where the memory triggers "to the sky" :)

Now, the reality is that I tried so many things that I'm a little bit confused about everything. So, I gonna comment the last feelings and tests that I make and, any question(or idea) for the future person (maybe me) that would try to solve the issue, it would well received.

So things that I tried to change:

- Strict version of Map, Monad, etc.
- More strict version of mapM_, gets and modify.
- Some partials changes in the data type PatchIndex.
- Using darcs as library; differents uses of the functions...

The "place" that I suspect the code has the memory glitch is in $applyPatchMods \rightarrow go (pid, PTouch fn)$I think that my lack of experience dealing with this kind of problem comes to light :) the
bright side is that I learn many things on the way of understand the problem and try different solutions.

Closing, maybe this could be my last post in a long time. I would like to thanks Guillaume and the darcs community in general. It was great to contribute to the development of darcs!

Again, I hope be able to work a little outside the GSoC, in particular in the "dependency problems and tasks".


No hay comentarios.:

Publicar un comentario