Presentation is loading. Please wait.

Presentation is loading. Please wait.

Mayank Bhatt, Jayasi Mehar

Similar presentations


Presentation on theme: "Mayank Bhatt, Jayasi Mehar"— Presentation transcript:

1 Mayank Bhatt, Jayasi Mehar
Topology-Aware Distributed Graph Processing for Tightly-Coupled Clusters Mayank Bhatt, Jayasi Mehar DPRG:

2 Our work explores the problem of graph partitioning, focused on reducing the communication cost on tightly coupled clusters

3 Why? Experimenting with cloud frameworks on HPC systems
Interest in supercomputing as a service More big data jobs running on supercomputers

4 Tightly-Coupled Clusters
Supercomputers Compute nodes embedded inside the network topology Messages routed via compute nodes Communication patterns can influence performance “Hop count” is an approximate measure of cost of communication

5 Blue Waters Interconnect
3D Torus Subset of nodes returned for running job Static routing - number of hops between two nodes will remain constant

6 Graph Processing Systems
Lot of real world data is expressed in the form of graphs Billion of vertices, trillions of edges, need to distribute Algorithms - ex. Shortest path, PageRank 2 stages - Ingress and Processing

7 Types of Partitioning System of choice: PowerGraph Masters and Mirrors
Masters communicate with all mirrors Our hypothesis: placing masters and mirrors close by should reduce communication cost Vertex Cuts Edge Cuts

8 Master mirror placement
Place replicas of a vertex first and then decide where to place the master Place the master of each vertex first and then decide where to place the replica - Hashing M R M R M

9 Random Partitioning Fast ingress
Communication cost between master and mirrors can be high Replication factor could be high M R R

10 Oblivious Partitioning
Slower ingress Heuristic based partitioning Leads to smaller replication factor than random Starting point to optimize Master mirror communication M R

11 Grid Partitioning Intersecting constraint sets
Leads to a controlled replication factor Master mirror communication not optimized M R

12 Topology Aware Variants
Make the partitioning step aware of the underlying network topology Place masters and mirrors such that communication cost is minimized

13 Choosing a master Pick master such that total number of hops are minimum Geometric centroid Edge degrees of each replica can be different Weighted Centroid

14 Grid Centroid Edges are placed using the Grid partitioning Strategy first Load: number of masters on candidate Number of edges on mirror Number of hops between mirror and candidate

15 Restricted Oblivious

16 Restricted Oblivious Number of edges on candidate
Maximum number of edges on a node Minimum number of edges on a node Number of hops between candidate and master

17 Experiments Cluster size: 36 nodes Algorithm: Approximate diameter
Graph: Power-law, 20 million vertices

18 Tradeoff between runtime and ingress

19 Data intensive algorithms benefit more
Graph Algorithms Data intensive algorithms benefit more

20 Improvements depend on type of graph
Graph Type Improvements depend on type of graph

21 Network Data Transfer

22 Other System Optimizations
Controlling the frequency of data injection into network impacts runtime in certain algorithms Smaller network buffers => flushed more frequently

23 Small computation and network data benefit from frequent flushing
Buffer Sizes PageRank Approximate Diameter Small computation and network data benefit from frequent flushing

24 Decisions, decisions

25 DPRG: http://dprg.cs.uiuc.edu
Conclusions Two new topology-aware algorithms for graph partitioning No ‘one size fits all’ approach to graph partitioning We propose a decision tree that can help decide which partitioning algorithm is best System optimizations complement performance DPRG:

26 Questions and Feedback?
DPRG:


Download ppt "Mayank Bhatt, Jayasi Mehar"

Similar presentations


Ads by Google