Skip to content

Latest commit

 

History

History
44 lines (31 loc) · 1.72 KB

GraphP-Reducing Communication for PIM-based Graph Processing with Efficient Data Partition-LX.md

File metadata and controls

44 lines (31 loc) · 1.72 KB

Paper title:

GraphP: Reducing Communication for PIM-based Graph Processing with Efficient Data Partition

Publication:

HPCA’18

Problem to solve:

In graph application, there are significant data movements. This posed great challenges to conventional computer architecture and memory systems. It is well-known for the poor locality in traversing the neighborhood vertices, and high memory bandwidth requirement, because the computations on data accesses from memory are typically simple.

Major contribution:

  1. This paper proposed GraphP, a novel HMC-based software/hardware co-designed graph processing system that drastically reduces communication and energy consumption compared to state-of-the-art system, TESSERACT.

  2. The first feature is “Source-cut” partitioning, which fundamentally changes the cross-cube communication from one remote put per cross-cube edge to one update per replica.

  3. The second feature is “Two-phase vertex program”, a programming model designed for “source-cut” partitioning with two operations: GenUpdate and ApplyUpdate.

  4. The third feature is “Hierarchical communication and overlapping”, which further improves performance with unique opportunities offered by the proposed partitioning and programming model.

Lessons learnt:

  1. Designing an architecture may start out data organization, not based on programming model. This is a heuristic-based design idea to improve performance.

  2. PIM is an effective technique that reduces data movements by integrating processing units within memory. Therefore, PIM can be employed into the NN fields and improve the speedup and energy saving.