RADIUSS

cnpetra pushed to LLNL/hiop

finished renaming to VectorSpace

View Commit

cyrush pushed to LLNL/conduit

changelog

View Commit

tgamblin pushed to spack/spack

Spec: use short-circuiting, stable comparison

Background

Spec comparison on develop used a somewhat questionable optimization to get decent spec comparison performance – instead of comparing entire spec DAGs, it put a hash() call in _cmp_iter() and compared specs by their runtime hashes. This gets us good performance abstract specs, which don’t have complex dependencies and for which hashing is cheap. But it makes the order of specs unstable and hard to reproduce.

We really need to do a full, consistent traversal over specs to compare and to get a stable ordering. Simply taking the hash out and yielding dependencies recursively (i.e. yielding dep.spec._cmp_iter() instead of a hash) goes exponential for concrete specs because it explores all paths. Traversal tracks visited nodes, but it’s expensive to set up the data structures for that, and it can slow down comparison of simple abstract specs. Abstract spec comparison performance is important for concretization (specifically setup), so we don’t want to do that.

New comparison algorithm

We can have (mostly) the best of both worlds – it’s just a bit more complicated.

This changes Spec comparison to do a proper, stable graph comparison:

Spec comparison will now short-circuit whenever possible for concrete specs, when DAG hashes are known to be equal or not equal. This means that concrete spec == and != comparisons will no longer have to traverse the whole DAG.
Spec comparison now traverses the graph consistently, comparing nodes and edges in breadth-first order. This means Spec sort order is stable, and it won’t vary arbitrarily from run to run.
Traversal can be expensive, so we avoid it for simple specs. Specifically, if a spec has no dependencies, or if its dependencies have no dependencies, we avoid calling traverse_edges() by doing some special casing.

The _cmp_iter method for Spec now iterates over the DAG and yields nodes in BFS order. While it does that, it generates consistent ids for each node, based on traversal order. It then outputs edges in terms of these ids, along with their depflags and virtuals, so that all parts of the Spec DAG are included. The resulting total ordering of specs keys on node attributes first, then dependency nodes, then any edge differences between graphs.

Optimized cases skip the id generation and traversal, since we know the order and therefore the ids in advance.

Performance ramifications

Abstract specs

This seems to add around 7-8% overhead to concretization setup time. It’s worth the cost, because this enables concretization caching (as input to concretization was previously not stable) and setup will eventually be parallelized, at which point it will no longer be a bottleneck for solving. Together those two optimizations will cut well over 50% of the time (likely closer to 90+%) off of most solves.

Concrete specs

Comparison for concrete specs is faster than before, sometimes way faster because comparison is now guaranteed to be linear time w.r.t. DAG size. Times for comparing concrete Specs:

def compare(json):
    a = spack.spec.Spec(json)
    b = spack.spec.Spec(json)
    print(a == b)
    print(timeit.timeit(lambda: a == b, number=1))

compare("./py-black.json")
compare("./hdf5.json")

develop (uses our prior hash optimization):
- py-black: 7.013e-05s
- py-hdf5: 6.445e-05s
develop with full traversal and no hash:
- py-black: 3.955s
- py-hdf5: 0.0122s
This branch (full traversal, stable, short-circuiting, no hash)
- py-black: 2.208e-06s
- py-hdf5: 3.416e-06s