Computation on Natural Graphs Presenter: Mengxiao Wang Problem: - - PowerPoint PPT Presentation

CSE 6350 File and Storage System Infrastructure in Data centers Supporting Internet-wide Services PowerGraph: Distributed Graph-Parallel Computation on Natural Graphs

Presenter: Mengxiao Wang

Problem:

Existing distributed graph computation systems perform poorly on natural graphs.

Properties of Natural Graphs: Power-Law Degree Distribution

Properties of Natural Graphs: Low-Quality Partition

• Power-Law graphs do not have low cost

balanced cut

• Traditional graph-partitioning algorithms

perform poorly on Power-Law graphs

Q1: Use Figure 1 to illustrate highly skewed power-law degree distribution in a graph and explain how this presents challenges to graph-parallel execution engines like Pregel (in terms of work balance, partitioning, communication, storage, and computation).

• Work balance: Due to the property of power-low distribution, the

runtime of vertices varies widely with graph-parallel execution.

• Partitioning: Pregel uses random partitioning of vertices, which

results in poor locality (only a small part of machines will have most edge cuts).

• Communication: High-degree vertices will have too much

communication other vertices resulting in bottleneck, like traffic problem and too many same messages.

• Storage: High-degree vertices will have too many edge metadata on

a single machine.

• Computation: Since the vertex-programs are executed in parallel

but abstractions within them do not parallelize, the high-degree vertices will have more computation than other vertices.

• Gather all the information of in-neighbors and in-edges of the
• vertex. Find out the minimum value of the sum of in-

neighbors and in-edges.

• Apply the value to the master vertex and update it to other

mirrors on other machines.

• If the value is changed, scatter the new value to all out-

neighbors and activate them to start GAS vertex programs.

Q2: Use Algorithm 1 and Figure 3 about SSSP to explain how an SSSP problem is solved in the execution of GAS vertex programs.

• Evenly assign edges to

machines

• Minimize machines spanned

by each vertex

• Assign each edge as it is

loaded

• Touch each edge only once
• Propose three distributed

approaches

• Random Edge Placement
• Coordinated Greedy Edge

Placement

• Oblivious Greedy Edge

Placement

Q3: Explain how the load balancing issue with a graph

• f highly skewed power-law degree distribution in

Pregel can be addressed in PowerGraph.

For greedy vertex-cuts:

• De-randomization: greedily

minimizes the expected number of machines spanned

• Coordinated

– Requires coordination to place each edge – Slower: higher quality cuts

• Oblivious

– Approx. greedy objective without coordination – Faster: lower quality cuts

• Rather than edge-cut
• Prefer vertex-cut

Must synchronize edges Must synchronize vertices

Distributed Execution of a PowerGraph Vertex-Program

Thank You!

Questions?