ABSTRACT
This paper introduces a framework to derive lower bounds on boundary overhead for accelerated local graph methods, formalizing the problem through a communication model and establishing theoretical bounds that are validated through extensive experiments.
PAPER · PDF
Loading PDF…
Key findings
Establishes a communication model capturing distributed local graph computation costs.
Proves that any P-processor implementation must incur significant boundary overhead.
Validates theoretical predictions on large-scale graphs, predicting real-world performance limitations.
Limitations & open questions
Theoretical bounds are tight up to polylogarithmic factors, leaving room for further refinement.
Experimental validation is limited to specific types of graphs and system configurations.