sssp comments

cc.cc

@@ -22,8 +22,8 @@ Author: Scott Beamer
 
 Will return comp array labelling each vertex with a connected component ID
 
-This CC implementation makes use of the Shiloach-Vishkin[2] algorithm with
-implementation optimizations from Bader et al.[1].
+This CC implementation makes use of the Shiloach-Vishkin [2] algorithm with
+implementation optimizations from Bader et al. [1].
 
 [1] David A Bader, Guojing Cong, and John Feo. "On the architectural
     requirements for efficient execution of graph algorithms." International

sssp.cc

@@ -15,12 +15,43 @@
 #include "pvector.h"
 #include "timer.h"
 
+
+/*
+GAP Benchmark Suite
+Kernel: Single-source Shortest Paths (SSSP)
+Author: Scott Beamer
+
+Returns array of distances for all vertices from given source vertex
+
+This SSSP implementation makes use of the ∆-stepping algorithm [1]. The type
+used for weights and distances (WeightT) is typedefined in benchmark.h. The
+delta parameter (-d) should be set for each input graph.
+
+The bins of width delta are actually all thread-local and of type std::vector
+so they can grow but are otherwise capacity-proportional. The currently
+processed bin makes use of the Bucket object and is generated right before it
+is used. Each iteration is done in two phases separated by barriers. In the
+first phase, the current shared bin is processed by all threads. As they
+find vertices whose distance they are able to improve, they add them to their
+thread-local bins. During this phase, each thread also votes on what the next
+bin should be (smallest non-empty bin). In the next phase, each thread moves
+their selected thread-local bin into the shared bin.
+
+Once a vertex is added to a bin, it is not removed, even if its distance is
+later updated and it now appears in a lower bin. We find ignoring vertices if
+their current distance is less than the min distance for the bin to remove
+enough redundant work that this is faster than removing the vertex from older
+bins.
+
+[1] Ulrich Meyer and Peter Sanders. "δ-stepping: a parallelizable shortest path
+    algorithm." Journal of Algorithms, 49(1):114–152, 2003.
+*/
+
+
 using namespace std;
 
 const WeightT kDistInf = numeric_limits<WeightT>::max()/2;
 
-
-// reduces barriers down to 2
 pvector<WeightT> DeltaStep(const WGraph &g, NodeID source, WeightT delta) {
   Timer t;
   pvector<WeightT> dist(g.num_nodes(), kDistInf);