Use bulk IntMap operations

[treeowl]

Use Data.IntMap.differenceWith to implement (&) and match
for Data.Graph.Inductive.PatriciaTree. Instead of modifying
the graph manually, one key at a time, differenceWith will
efficiently partition the set of keys to be modified along the
structure of the graph. This should be considerably more efficient
when inserting or matching on well-connected nodes.

Fixes #39

[treeowl]

We may want to change this a bit, actually. I started to make the change, then got confused and stopped. The differenceWith approach is not great when the node is not well-connected. We can actually detect this by building the connection IntMaps manually with foldl', counting as we go, rather than using the (entirely equivalent) fromListWith.

[treeowl]

Hmm.... Actually, I'm not sure if it's worth trying to specialize for the poorly-connected case. Do you have any benchmarks that this PR could be run against?

[treeowl]

Unfortunately, it seems we do have to worry about the poorly-connected case; the benchmarks show a regression. I'm working on it. It's not so clear to me just what the benchmarked graphs tend to look like; that might be nice to document.

[treeowl]

Let me take that back again! The benchmark change must have been some insignificant artifact, because the benchmarks don't actually touch any of that code. We definitely need more benchmarks!

[ivan-m]

Yes, benchmarks are solely lacking.

Alas, I have little excuses to do graph Haskell usage at work (as the few times I thought it may have been relevant for some data processing it ended up running for so long that I killed it as it looked untenable) so I don't have any ready made large-scale graph examples to use :/

[treeowl]

Another little issue is that the benchmark framework is ancient and weird.
microbench has made exactly one release, in 2008. Looking at the
documentation, I just realized I've been using it wrong. But that's partly
explained by the fact that it's a bit hard to see how one can use it right
for graph building. We should almost certainly be using Criterion instead.

On Tue, Aug 30, 2016 at 7:35 PM, Ivan Lazar Miljenovic <
[email protected]> wrote:

Yes, benchmarks are solely lacking.

Alas, I have little excuses to do graph Haskell usage at work (as the few
times I thought it may have been relevant for some data processing it ended
up running for so long that I killed it as it looked untenable) so I don't
have any ready made large-scale graph examples to use :/

—
You are receiving this because you modified the open/close state.
Reply to this email directly, view it on GitHub
#40 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/ABzi_ct7FQM6IOXZj51abHWkLwH-wcF8ks5qlL49gaJpZM4JwKVB
.

[ivan-m]

Definitely; I just hadn't gotten around to doing new benchmarks yet.