WANify: Gauging and Balancing Runtime WAN Bandwidth for Geo-distributed Data Analytics

Presentation by Anshuman Das Mohapatra

Advisor: Dr. Kwangsung Oh

TABLE OF CONTENT

1. CONTEXT and MOTIVATION
2. PROBLEM STATEMENT
3. OUR APPROACH and DESIGN
4. EVALUATIONS
5. CONCLUSION

WANify: Gauging and Balancing Runtime WAN Bandwidth

CONTEXT

1. Incurs high latency and cost
2. Data Protection laws prohibit raw data movement

WANify: Gauging and Balancing Runtime WAN Bandwidth

GEO-DISTRIBUTED DATA ANALYTICS (GDA)

WANify: Gauging and Balancing Runtime WAN Bandwidth

Input Data

Intermediate Data

BW-AWARE GDA

WANify: Gauging and Balancing Runtime WAN Bandwidth

Intermediate Data

Identify

BW-rich �Data centers

MOTIVATION (1)

WANify: Gauging and Balancing Runtime WAN Bandwidth

Prior works measure BWs in a static-independent way to make optimal task placement decisions.

Statically measured BW-rich data centers

MOTIVATION (1)

WANify: Gauging and Balancing Runtime WAN Bandwidth

Prior works measure BWs in a static-independent way to make optimal task placement decisions.

Simultaneously measured dynamic BWs are significantly different from static-independent measurements

Dynamically measured BW-rich data centers

Statically measured BW-rich data centers

MOTIVATION (2)

1. Single Connection is slow
2. Distant Data centers still bottlenecked by network

WANify: Gauging and Balancing Runtime WAN Bandwidth

PROBLEM STATEMENT

How to use simultaneous runtime BWs as inputs to GDA applications to enable optimal decision-making?

How to improve the achievable minimum BW in a GDA cluster to improve overall latency?

WANify: Gauging and Balancing Runtime WAN Bandwidth

WHY NOT MEASURE DYNAMIC BWs SIMULTANEOUSLY?

WANify: Gauging and Balancing Runtime WAN Bandwidth

FEATURES FOR PREDICTING RUNTIME BWs?

WANify: Gauging and Balancing Runtime WAN Bandwidth

DETERMINING RUNTIME BWs

• iPerf3: 20 seconds to stable runtime BWs
• 1-second snapshot BWs – High Person Correlation Coefficient with 20-second BWs
• Train Random Forest (RF) regressor for determining runtime BWs

DATASET

…

AVERAGING

WANify: Gauging and Balancing Runtime WAN Bandwidth

UNIFORM PARALLELISM GIVES UNFAIR NETWORK CONSUMPTION

WANify: Gauging and Balancing Runtime WAN Bandwidth

BW-rich data centers

BW-deficit data centers

Min. BW

Max. BW

HETEROGENEOUS PARALLELISM GIVES BETTER MIN. BW

WANify: Gauging and Balancing Runtime WAN Bandwidth

HETEROGENEOUS PARALLELISM GIVES BETTER MIN. BW

WANify: Gauging and Balancing Runtime WAN Bandwidth

DESIGN

• WANify prototype - abstraction for GDA applications
• Snapshots for runtime bandwidths
• Random Forest for multi-variate regression using snapshots
• Global and local optimization for heterogenous parallel connections

WANify: Gauging and Balancing Runtime WAN Bandwidth

WANIFY ARCHITECTURE

WANify: Gauging and Balancing Runtime WAN Bandwidth

WANIFY-ENABLED GDA ENVIRONMENT

Spark 2.2.1

Prototype implementation and testing

Amazon Web Services

8 Geo-distributed data centers

Compute Resources

t2.large for Master (2 vCPUs and 8 GB RAM)

t2.medium for worker (2 vCPUs and 4 GB RAM)

WANify: Gauging and Balancing Runtime WAN Bandwidth

EVALUATIONS: GEO-DISTRIBUTED ML

WANify: Gauging and Balancing Runtime WAN Bandwidth

EVALUATIONS: TPC-DS (STATE-OF-THE-ART)

WANify: Gauging and Balancing Runtime WAN Bandwidth

Query Latency Improvement

EVALUATIONS: TPC-DS (STATE-OF-THE-ART)

WANify: Gauging and Balancing Runtime WAN Bandwidth

Cost Improvement

Query Latency Improvement

CONCLUSION

• WANify helps GDA applications with accurate runtime bandwidths.�
• WANify enhances achievable bandwidths (up to 3.3x) in WAN by using optimal and parallel heterogeneous connections.�
• Comparisons with baselines show significant performance-cost improvements.

WANify: Gauging and Balancing Runtime WAN Bandwidth