GoFFish - A Sub-graph centric framework for large scale graph analytics
0 likes2,386 views
1. The document discusses GoFFish, a sub-graph centric programming model for large scale graph analytics. It partitions graphs into connected sub-graphs that are processed independently in parallel. 2. GoFFish showed significant performance improvements over traditional vertex-centric models for connected components and single source shortest path algorithms, with speedups of up to 81x and 38x respectively. 3. While PageRank was less suited to the sub-graph model, overall GoFFish demonstrated it can efficiently analyze large graphs distributed across a cluster using a simple sub-graph centric programming approach.
GoFFish - A Sub-graph centric framework for large scale graph analytics
- 1. 1
- 2. Twitter Storm 1st Mentioning of Big Data Michael Cox and David Ellsworth publish “Application-controlled demand paging for out-of-core visualization” in the Proceedings of the IEEE 8th conference on Visualization, 1997 http://www.gartner.com/it-glossary/big-data/ Pregel Map Reduce 2
- 3. • Scalable platform for large scale graph analytics in commodity clusters and cloud. • Simple yet scalable programming abstraction for large scale graph processing • More than 10x improvements in some graph algorithms over traditional graph processing systems • Components – GoFS : Sub-graph centric Distributed Graph Storage – Gopher : Distributed BSP-based programming abstraction for Large Scale graph analytics framework 3
- 4. • Iterative vertex centric programming model based on Bulk synchronous parallel model. • In each iteration vertex can 1. Receive messages sent to it in previous iteration 2. Send messages to other vertices 3. Modify its own state • Vertex centric programming mode- Think like a vertex • Very simple, • High communication overhead • Pregel, http://giraph.apache.org/ 4
- 5. • Partition graph in to set of sets of vertices and edges • Sub-graphs – Connected components in a partition. • User logic operates on a sub graph – Independent unit of computation • Resource allocation – Single Partition → Single Machine | Single Sub-graph → Single CPU • Data loading – Entire partition is loaded into memory before computation – Tasks retain sub-graphs in memory within the task scope Sub-graph-Task 1 Sub-graph-Task 2 Sub-graph-Task 3 5
- 6. • Algorithm : Seqn of super-steps separated by global barrier synchronization • Super Step i: 1. 2. 3. 4. Sub-graphs compute in parallel Receive messages sent to it in super step i-1 Execute same user defined function Send messages to other sub graphs (to be received in super step i+1) 5. Can vote to halt : I’m done / de-activate http://www.ibm.com/developerworks/opensource/library/os-giraph/ • Global Vote to Halt check BSP Manager Control Channels – termination → all sub-graphs voted to halt • Active sub graphs participate in every computation • De-activated sub-graphs will not get executed/activated unless it get new messages P1 P1 Data Channnels P1 6
- 7. • Algorithm – Max vertex value • Input – Connected graph with different vertex values {n1,n2…. ni} • Output – Each vertex with the Max ({n1,n2…. ni}) • Compared to Vertex centric model • Less communication • Fast convergence 7
- 8. 8
- 9. • Cluster – 12 Nodes, 8-Core Intel Xeon CPU (each) – 16GB RAM (each), 1TB HDD (each) • Network – Gigabit Ethernet • Apache Giraph latest version from trunk – Includes Performance improvements from Facebook ! Data Set Vertices Edges Diameter RN: California Road Network 1,965,206 2,766,607 849 TR: Internet path graph 19,442,778 from Tracesroutes 22,782,842 25 LJ: LiveJournal Social Network 68,475,391 10 4,847,571 9
- 10. • Connected components – 81x improvement using California Road Network (RN) dataset – 21x improvement using a Trace route path network of a CDN. (TR) dataset Page Rank 4x improvement using RN dataset 1.5x improvement using TR dataset Not an ideal algorithm for sub-graph centric programming model • Single Source Shortest Path – 32x improvement using RN dataset – 10x improvement using TR dataset 10
- 11. • Connected components – ~79x reduction using RN dataset. – ~4x reduction using TR dataset – ~2x reduction using Live Journal dataset (LJ) • Single source shortest path – ~38x reduction using RN data set – 4x reduction using TR dataset – ~1.6x reduction using Live journal dataset 11
- 12. • Introduced a sub-graph centric programming abstraction for large scale graph analytics on distributed systems – Simple – Enable using shared memory algorithms at sub-graph level. • Sub-graph centric algorithms and performance results – Connected Components – Single Source Shortest Path – Page Rank • Issues and Future work – Sub-graph aware partitioning – Sub-graph centric algorithms • Try now – https://github.com/usc-cloud/goffish 12