1 Advancing Supercomputer Performance Through Interconnection Topology Synthesis Yi Zhu, Michael Taylor, Scott B. Baden and Chung-Kuan Cheng Department.

1

Advancing Supercomputer Performance Through Interconnection Topology Synthesis

Yi Zhu, Michael Taylor, Scott B. Baden and Chung-Kuan Cheng

Department of Computer Science and EngineeringUniversity of California, San Diego

2

Outline

Introduction Design Flow, Formulation & Algorithms Example: Blue Gene/L Packaging

Overview Models & Constraints

Experiments Benchmark Instances Generated Instances

Conclusion & Future Work

3

Interconnection Networks Interconnection networks become a more critical

factor than computing or memory modules (W. Dally, HPCA 2007 Keynote Speech)

Popular network topologies: Hypercube (SGI Origin2000) 2D torus (Cray X1) 3D torus (Cray T3E and XT3, IBM Blue Gene/L) Crossbar (NEC Earth Simulator) Folded Clos (Cray BlackWidow) Fat tree, flattened butterfly, Etc.

4

Our Work We propose a design methodology to

select the best topology to minimize the average latency Design flow is fully automated Physical constraints can be specified by

users Efficient multi-commodity flow algorithm

to evaluate Demonstrate the efficiency using Blue

Gene/L packaging framework

5

Design Flow

MCF Evaluation Solver

Delay Models

Topology Pool

Communication Patterns

Physical Constraints

Best Topology

6

Multi-Commodity Flow (MCF) Graph G(V,E) K commodities, each has a source and a

sink, and demand amount d(k) Each edge e has a capacity u(e) Each edge e has a weight w(e) Minimum Cost MCF: each commodity k is

routed units under the capacity constraints, minimize , where f(e) is the flow routed on edge e

( )d k( ) ( )

e

f e w e

7

Map Supercomputer Performance Evaluation to MCF Problem

Nodes – processors Edges – interconnection links Commodities – communications Demands – communication bandwidth

(injection rate) Flow amount – wires assignments Capacity constraints – physical constraints

(wires, pins, board dim) Edge weight – unit latency (unit power)

8

An Example on Maximum Concurrent Flow

Two commodities: s1->t1, s2->t2, both have demand d(1)=d(2)=1

Optimal throughput = 1.5

s2

s1t1

t2

2

2

3

2

2

9

Approximation Algorithms The duality theory in LP: for a

maximization, primal feasible , dual feasible D, optimal solution OPT

Increase and decrease D iteratively till the duality gap is small enough

OPT D

10

Blue Gene/L: An Example

Midplane: 8x8x8 Torus

11

Assumptions

We follow the same hierarchical structure: midplane – node card – compute card

The properties of boards (dimensions, # layers, dielectric) keep unchanged

We seek better topologies than the existing 3D torus to implement the networks in the midplane

12

Topology Generation

Generate 8-node 1D topologies and duplicate to each row and column

Topologies are isomorph-free and has maximum degree bound for each node

#isomorph-free topologies

13

Node Card Graph Model

Horizontal: Strongly Connected; Vertical: Generated Topology

14

Midplane Graph Model

a

bd

cf

eg

h

Coteus et al., “Packaging the Blue Gene/L Supercomputer”IBM J of Res & Dev, Vol. 43, pp. 213-248

15

Experiment 1: Benchmark Instances

NAS Parallel Benchmarks (121/128 processes)

Benchmark source code

Compiled with Intel Trace Collector & Analyzer

Executable

Run on multi-processor machines

Output

Simulated annealing placement

Traffic Patterns

Task placement

Our design flow

Best topology

16

Benchmarks

Characteristics Communication Pattern: MG

17

Results

Optimal: each instance has different topology Aggregate: one topology for all instances 3D Torus: 3D torus topology

18

Experiment 2: Generated Instances

Randomly generated communications Scalar values which represent the

demand for bandwidth between each pair of nodes

More general, time independent Control Parameters

# communication demands: O(n) pairs Communication amount: uniform traffic

but vary case by case (different congestion level)

19

Latency & Throughput Tradeoffs

Distribution: 40% / 50% / 10%

20

Topologies with Different Injection Rates

1 2 3 4 5 6 7 8

1 2 3 4 5 6 7 8

Injection rate = 1.5

Injection rate = 1.9

With larger injection rate, more(red) links are needed to go through the cut between 4 and 5, in order to reduce the number of hops

21

Conclusion

An design flow for interconnection network synthesis Fully automated Explore large design space Efficient evaluation algorithm

Future work Power consumption Accurate simulation

22

Q&A

Thank you!