馃惛 #36
馃惛 #36
Conversation
This switches from using @frankmcsherry鈥檚 differential dataflow to using @frankmcsherry鈥檚 bespoke datalog engine: datafrog. 99% of this commit was authored by the one and only @frankmcsherry.
5% perf for free
src/output/datafrog_opt.rs
Outdated
| let requires_1 = iteration.variable("requires_1"); | ||
| let requires_2 = iteration.variable("requires_2"); | ||
| let requires_bp = iteration.variable("requires_bp"); | ||
| let requires_rp = iteration.variable("requires_rp"); |
There was a problem hiding this comment.
I apologize for not noticing this before, but each of the requires_ variables can be variable_indistinct(). We just need requires itself (without a _) to be a variable() without _indistinct(). This has a solid perf benefit for me, taking the time from 9.43s to 7.99s.
There was a problem hiding this comment.
Solid benefits for me as well, 10-15% 馃憤
Another 10-15% for free
The old code dropped `p` and renamed `q` to `p`, which confused me.
| @@ -12,8 +12,8 @@ arg_enum! { | |||
| #[derive(Debug, Clone, Copy)] | |||
| pub enum Algorithm { | |||
| Naive, | |||
| TimelyOpt, | |||
| LocationInsensitive, | |||
| DatafrogOpt, | |||
src/output/naive.rs
Outdated
| subset.from_join(&subset_2, ®ion_live_at, |&(r2,q),&r1,&()| (r1,r2,q)); | ||
| } | ||
|
|
||
| subset_r1p.complete() |
There was a problem hiding this comment.
Just to check my understanding here: it seems to me that the subset variable is not needed. We could just use subset_r1p instead? (That is, nothing directly joins from subset; the only use of it is to be mapped into different arrangements, and we could do those maps from subset_r1p just as easily, right?)
(cc @frankmcsherry)
There was a problem hiding this comment.
Answer: yes, I checked elsewhere, and even got a slight win from doing so (in datafrog-opt).
|
|
||
| // since we're using subset mapped ((r, p), r) we can use it directly out of iteration 1 | ||
| let subset_r1p = iteration2.variable::<((Region, Point), Region)>("subset_r1p"); | ||
| subset_r1p.insert(subset); |
There was a problem hiding this comment.
This "variable" is invariant, right? Do we really need to 'copy' the values from subset in here, or could we use subset directly?
(cc @frankmcsherry)
There was a problem hiding this comment.
Well, let me answer my own question: it won't build if we don't, because from_join wants a Variable and not a Relation: presumably that could be modified in datafrog? (Or is this insert very cheap anyway?)
There was a problem hiding this comment.
(I realize in this specific case we ought to collapse the iteration anyway.)
There was a problem hiding this comment.
(this is also relevant, it seems to the "redundantly computed index" entries we see below)
There was a problem hiding this comment.
Inserting a sorted relation should be pretty cheap. It will still do linear work, confirming that none of subset exists in the empty list of prior tuples, which we could totally optimize. But it shouldn't be too expensive either way.
More generally, there a bunch of cases where the difference between Variable and Relation and between Key and (Key, ()) bite us a little. These can be picked off as appropriate; just some repetition in the library code in the interest of squeaking out some perf. :)
There was a problem hiding this comment.
Oh, my mistake. It will not even do the linear work, I think, on account of self.stable should be empty. I could be wrong about that, and we could work through exactly what happens when you install a relation. Not a great amount of care was used putting the wrappers in place. :)
|
Final measurements:
Amazing! Great work @frankmcsherry and @lqd! |
This switches from using @frankmcsherry鈥檚 differential dataflow to
using @frankmcsherry鈥檚 bespoke datalog engine: datafrog.
99% of this commit was authored by the one and only @frankmcsherry.