Scale Graph A HighPerformance Library for BillionScale Graph

Scale. Graph: A High-Performance Library for Billion-Scale Graph Analytics Miraj Kheni Authors: Toyotaro Suzumura, Koji Ueno

Scale. Graph: Motivations • Growing interest in the field of large-scale graph mining • Achieving high performance and low software complexity at the same time • Harnessing performance and productivity of X 10

Scale. Graph: Related Work • Parallel Boost Graph Library (PBGL) • Google’s Pregel • Apache Giraph • Graph. Lab

Scale. Graph: X 10 Overview • A type-safe, imperative, concurrent, object-oriented language • Asynchronous partitioned global address space (APGAS) programming model • Key concepts: • • • place: analogous to process activity: analogous to lightweight thread activity can access remote data using at keyword or X 10. util. Team class Seamless migration of an activity to different place Supports Clik-style fork-join programming Async, Finish

Scale. Graph: Improved X 10 Team Class • X 10 Team: contains routines for collective communication • Scale. Graph proposes a new X 10 Team class that optimizes collective communication by not emulating X 10 RT • Implements both blocking and non-blocking collective communication routines that directly uses MPI collective communication routines.

Scale. Graph: New Memory Management System • X 10 back-end uses Boehm-Demers-Weiser conservative garbage collector (BDWGC) • Does not differentiate small and large memory allocation requests • Authors proposes EMM (Explicitly Managed Memory) • For small memory allocation- use BDWGC and for large memory allocation- explicitly use malloc and free

Scale. Graph 2: Overview • Implemented new readers and writes for parallel IO • Implemented new X 10 Team • Implemented XPregel framework • Three main components: • XPregel framework • Basic linear Algebra Subprograms (BLAS) • File IO • Provides a number of graph algorithms

Scale. Graph: Graph Representation • Distributed Graph Object • Distributed mutable lists of edges, vertex attributes and edge attributes • Does not support query operations on graph • Distributed Sparse Matrix • Efficient way for storing immutable, sparse, adjacency matrix • Supports querying vertex neighbors, vertex attributes and edge attributes

Scale. Graph: Graph Representation Distributions of distributed sparse Matrix on four places: 1) 1 D Row wise (R=4, C=1) 2) 1 D Column wise (R=1, C=4) 3) 2 D Block (R=2, C=2)

Scale. Graph: XPregel Framework • Pregel Computation Model • Supersteps- during each superstep, a user-defined function is invoked for each vertex • Three main interfaces- Message Passing, Combiner and Aggregator • XPregel interface:

Scale. Graph: More Optimizations • Send. All Optimization Technique • Aims at reducing number of messages when a vertex sends the same massage to all of its neighbors • If enabled, the vertex will send only one message to destination place, which in turn would duplicate the message before passing it to respective vertices • BLAS for Sparse Matrix • For graph problems that can be expressed using repeated matrix-vector multiplication

Scale. Graph: Evaluation • Updated X 10 Team:

Scale. Graph: Evaluation • Weak Scaling: • Strong Scaling:

Scale. Graph: Evaluation • Strong Scaling: Scale. Graph vs Giraph vs PBGL

Conclusion • Use of X 10 – helps in productivity and performance • Use of optimized X 10 Team – optimizes collective communications • Introduces EMM – better memory management • XPregel’s Send. All – Reduces duplication of messages • These optimizations helps Scale. Graph in achieving high productivity, scalability and productivity.

Thank You