[SPARK-3936] Add aggregateMessages, which supersedes mapReduceTriplets

aggregateMessages enables neighborhood computation similarly to mapReduceTriplets, but it introduces two API improvements:

1. Messages are sent using an imperative interface based on EdgeContext rather than by returning an iterator of messages.

2. Rather than attempting bytecode inspection, the required triplet fields must be explicitly specified by the user by passing a TripletFields object. This fixes SPARK-3936.

Additionally, this PR includes the following optimizations for aggregateMessages and EdgePartition:

1. EdgePartition now stores local vertex ids instead of global ids. This avoids hash lookups when looking up vertex attributes and aggregating messages.

2. Internal iterators in aggregateMessages are inlined into a while loop.

In total, these optimizations were tested to provide a 37% speedup on PageRank (uk-2007-05 graph, 10 iterations, 16 r3.2xlarge machines, sped up from 513 s to 322 s).

Subsumes apache/spark#2815. Also fixes SPARK-4173.

Author: Ankur Dave <ankurdave@gmail.com>

Closes #3100 from ankurdave/aggregateMessages and squashes the following commits:

f5b65d0 [Ankur Dave] Address @rxin comments on apache/spark#3054 and apache/spark#3100
1e80aca [Ankur Dave] Add aggregateMessages, which supersedes mapReduceTriplets
194a2df [Ankur Dave] Test triplet iterator in EdgePartition serialization test
e0f8ecc [Ankur Dave] Take activeSet in ExistingEdgePartitionBuilder
c85076d [Ankur Dave] Readability improvements
b567be2 [Ankur Dave] iter.foreach -> while loop
4a566dc [Ankur Dave] Optimizations for mapReduceTriplets and EdgePartition

0 files changed, 0 insertions, 0 deletions