2017 Fiscal Year Annual Research Report
Structural Recursion on Bulk Synchronous Parallelism for Efficient Large-Graph Querying
| Project/Area Number |
15K12011
|
|---|
| Research Institution | National Institute of Informatics |
|---|
Principal Investigator |
胡 振江 国立情報学研究所, 大学共同利用機関等の部局等, 教授 (50292769)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
LI CHONG 国立情報学研究所, 大学共同利用機関等の部局等, 特任研究員 (50745312) [Withdrawn]
|
|---|
| Project Period (FY) |
2015-04-01 – 2018-03-31
|
| Keywords | Graph Processing / Structural Recursion / Parallelization / Pregel |
|---|
| Outline of Annual Research Achievements |
The objective of this research is to combine the expressive power of structural recursion with Pregel, a popular vertex-centric graph processing system based on the BSP model.
In this year, we extended our system with a more powerful domain specific language that supports remote data access for the vertex-centric graph processing. More concretely, (1) we proposed a new high-level model for vertex-centric computation, where the concept of algorithmic supersteps is introduced as the basic computation unit for constructing vertex-centric computation in such a way that remote reads and writes are ordered in a safe way; (2) based on the new model, we designed and implemented Palgol, a more declarative and powerful DSL, which supports both remote reads and writes, and allows programmers to use a more declarative syntax called chain access to directly read data on remote vertices, with a new logic system to compile chain access to efficient message passing where the number of supersteps is reduced whenever possible; (3) we demonstrated the power of Palgol by working on a set of representative examples, including the Shiloach-Vishkin connected component algorithm and the list ranking algorithm, which use communication over dynamic data structures to achieve fast convergence.
The experimental results are encouraging. The efficiency of Palgol is comparable with hand-written code for many representative graph algorithms on practical big graphs, where execution time varies from a 2.53% speedup to a 6.42% slowdown in ordinary cases, while the worst case is less than a 30% slowdown.
|
Research Products
(3 results)
