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

TL;DR

WANify tackles the instability of WAN bandwidth in geo-distributed data analytics by predicting runtime inter-DC BW with a Random Forest model and by optimally distributing heterogeneous parallel connections. It couples offline BW prediction with static global and dynamic local optimizations, while incorporating heterogeneity in data distribution, DC counts, and cloud providers. The approach yields substantial performance and cost benefits—as much as 26% latency reduction and 16% cost saving in AWS-based experiments—by better exploiting WAN capacity and throttling congested links. The work demonstrates broad applicability across GDA workloads and ML tasks, and provides open-source tooling and datasets for further research.

Abstract

Accurate wide area network (WAN) bandwidth (BW) is essential for geo-distributed data analytics (GDA) systems to make optimal decisions such as data and task placement to improve performance. Existing GDA systems, however, measure WAN BW statically and independently between data centers (DCs), while data transfer occurs dynamically and simultaneously among DCs during workload execution. Also, they use a single connection WAN BW that cannot capture actual WAN capacities between distant DCs. Such inaccurate WAN BWs yield sub-optimal decisions, inflating overall query latency and cost. In this paper, we present WANify, a new framework that precisely and dynamically gauges achievable runtime WAN BW using a machine learning prediction scheme, decision tree-based Random Forest. This helps GDA systems make better decisions yielding reduced latency and costs including WAN BW monitoring costs. Based on predicted runtime WAN BW, WANify determines the optimal number of heterogeneous parallel connections for data transfer among DCs. This approach improves performance without additional, or even at reduced cost, by fully exploiting available WAN capacities. In addition, WANify considers dynamics like network and workloads, and heterogeneity like skewed data, heterogeneous compute resources, and a varying number of DCs while making decisions. The WANify prototype running on state-of-the-art GDA systems is evaluated on AWS with 8 geo-distributed DCs. Results show that WANify enhances WAN throughput by balancing between the strongest and weakest WAN links, enabling GDA systems to reduce latency and cost by up to 26% and 16% respectively with minimal effort, all while handling dynamics and heterogeneity efficiently.

Figures (10)

• Figure 1: Highly geo-distributed data and cloud resources
• Figure 2: BWs and network latency for different approaches
• Figure 3: End-to-end architecture of WANify.
• Figure 4: Impact on ML in GDA.
• Figure 5: Comparing different parallel approaches.