[Paper Review] Pied Piper: Rethinking Internet Data Delivery

Aran Bergman, Israel Cidon+5 more|arXiv (Cornell University)|Dec 13, 2018

Network Traffic and Congestion Control31 references4 citations

TL;DR

Pied Piper proposes a cloudified data delivery architecture that rethinks Internet transport by offloading routing and congestion control to cloud-based relays, using TCP splitting and kernel-level optimizations to achieve up to 30x faster file downloads compared to traditional Internet delivery, outperforming advanced end-to-end protocols like BBR and PCC by an order of magnitude.

ABSTRACT

We contend that, analogously to the transition from resource-limited on-prem computing to resource-abundant cloud computing, Internet data delivery should also be adapted to a reality in which the cloud offers a virtually unlimited resource, i.e., network capacity, and virtualization enables delegating local tasks, such as routing and congestion control, to the cloud. This necessitates rethinking the traditional roles of inter- and intra-domain routing and conventional end-to-end congestion control. We introduce Optimized Cloudified Delivery (OCD), a holistic approach for optimizing joint Internet/cloud data delivery, and evaluate OCD through hundreds of thousands of file downloads from multiple locations. We start by examining an OCD baseline approach: traffic from a source A to a destination B successively passes through two cloud virtual machines operating as relays - nearest to A and B; and the two cloud relays employ TCP split. We show that even this naive strategy can outperform recently proposed improved end-to-end congestion control paradigms (BBR and PCC) by an order of magnitude. Next, we present a protocol-free, ideal pipe model of data transmission, and identify where today's Internet data delivery mechanisms diverge from this model. We then design and implement OCD Pied Piper. Pied Piper leverages various techniques, including novel kernel-based transport-layer accelerations, to improve the Internet-Cloud interface so as to approximately match the ideal network pipe model.

Motivation & Objective

• Address the persistent inefficiencies in traditional Internet routing (BGP) and congestion control (TCP), which remain suboptimal despite decades of research.
• Explore the performance potential of leveraging cloud networks as a resource-abundant infrastructure for data delivery, analogous to the shift from on-prem to cloud computing.
• Investigate whether delegating transport-layer functions like routing and congestion control to virtualized cloud relays can significantly improve end-to-end performance.
• Design and evaluate a holistic system—OCD Pied Piper—that optimizes joint Internet and cloud data delivery by approximating an ideal network pipe model.

Proposed method

• Implement an OCD baseline using two cloud relays (near source and destination) and TCP splitting to divide the end-to-end connection into three segments, minimizing time spent in the traditional Internet.
• Develop a protocol-free, ideal pipe model of data transmission to identify deviations in current Internet mechanisms, particularly the seriality introduced by BGP and end-to-end congestion control.
• Design and implement OCD Pied Piper using kernel-based transport-layer accelerations to reduce latency and improve fairness, especially in competitive flow scenarios.
• Use a hybrid architecture where traffic is routed through cloud relays that handle transport-layer functions, enabling better control over congestion and path selection.
• Leverage virtualization to deploy relays with high bandwidth (up to 2Gbps per VM) and low-latency connectivity, effectively bypassing traditional Internet bottlenecks.
• Apply novel techniques such as split TCP with standard Berkeley sockets and packet-level optimizations to improve performance without requiring protocol changes.

Experimental results

Research questions

• RQ1 Can offloading routing and congestion control to cloud-based relays significantly improve end-to-end data delivery performance compared to traditional end-to-end TCP?
• RQ2 How does a simple OCD baseline with two cloud relays and TCP splitting compare to advanced end-to-end congestion control protocols like BBR and PCC in real-world conditions?
• RQ3 To what extent do current Internet mechanisms deviate from an idealized, fixed-capacity network pipe model, and how can these deviations be mitigated?
• RQ4 Can kernel-level transport optimizations in the cloud approximate the behavior of an ideal network pipe and improve fairness between competing flows with different RTTs?
• RQ5 What is the performance impact of using cloud relays for data delivery across diverse topologies, including inter-continental and intra-continental paths?

Key findings

• The OCD baseline—using two cloud relays and TCP splitting—achieved up to 30× faster file downloads for large files compared to the wild Internet, significantly outperforming BBR and PCC, which improved performance by less than 30%.
• For 4MB files, the OCD baseline improved download times by a factor of 5×, demonstrating substantial gains even with a minimal architectural change.
• In competitive flow scenarios, the long-RTT flow was previously starved due to unfairness; with Pied Piper enabled, both short- and long-RTT flows achieved roughly equal throughput, demonstrating improved fairness.
• The ideal pipe model revealed that today’s Internet introduces unnecessary seriality in transport, which cloudification can eliminate by decoupling path selection and congestion control from end-hosts.
• Kernel-based transport optimizations in OCD Pied Piper reduced latency and improved performance, particularly in high-RTT, multi-hop scenarios, validating the benefits of cloud-based transport control.
• Empirical evaluations across three major clouds and multiple geographic locations confirmed consistent performance gains, especially for large files and long-haul paths.