[Paper Review] A cost-benefit analysis of low latency via added utilization

Ashish Vulimiri, Brighten Godfrey+3 more|arXiv (Cornell University)|Jun 15, 2013

Caching and Content Delivery4 references3 citations

TL;DR

This paper proposes an economic cost-benefit framework to evaluate latency-reduction techniques that increase bandwidth usage in wide-area client-server systems. By quantifying the trade-off between latency savings and added utilization costs, it establishes a benchmark: any technique saving more than 10 ms of latency per kilobyte of extra traffic is economically beneficial, even under pessimistic cost estimates, and validates this with a case study on redundant DNS requests.

ABSTRACT

Several recently proposed techniques achieve latency reduction by trading it off for some amount of additional bandwidth usage. But how would one quantify whether the tradeoff is actually beneficial in a given system? We develop an economic cost vs. benefit analysis for answering this question. We use the analysis to derive a benchmark for wide-area client-server applications, and demonstrate how it can be applied to reason about a particular latency saving technique --- redundant DNS requests.

Motivation & Objective

• To develop a systematic framework for evaluating the economic trade-off between latency reduction and increased bandwidth usage in wide-area client-server applications.
• To quantify the cost and benefit of latency-saving techniques from the perspectives of both clients and servers, accounting for real-world pricing and performance metrics.
• To derive a practical benchmark that determines when added utilization for latency reduction is economically justified.
• To validate the framework through a case study on redundant DNS requests, assessing its real-world impact on web page load times.

Proposed method

• The framework models latency savings per kilobyte of added traffic ($\ell$) as a ratio of benefit to cost, normalized across client and server sides.
• It introduces normalized cost parameters ($p_c$, $p_s$) for client and server-side bandwidth usage, and value parameters ($v_c$, $v_s$) for latency improvement at each end.
• The core criterion is $\ell \geq \max\left\{ \frac{p_s}{v_s}, \frac{p_c}{v_c} \right\}$, ensuring both clients and servers see net benefit from the technique.
• The analysis uses publicly available pricing data from cloud providers (e.g., AWS, Azure, CloudFront) and consumer plans (e.g., AT&T, O2) to estimate $p_c$ and $p_s$.
• The benchmark is derived by computing the break-even latency savings threshold ($\ell$) across various service plans, using conservative cost estimates.
• The framework is applied to redundant DNS requests by measuring mean and 95th percentile page load time improvements across real deployments (Chrome on Emulab, Firefox on DSL).

Experimental results

Research questions

• RQ1 What is the minimum latency improvement per kilobyte of added traffic that makes a latency-saving technique economically viable for both clients and servers?
• RQ2 How do varying cost structures (e.g., cellular vs. DSL, cloud vs. self-hosted) affect the economic threshold for adopting latency-reduction techniques?
• RQ3 To what extent do redundant DNS requests improve real-world web page load times, and at what point do they become economically inefficient?
• RQ4 How does the incremental benefit of adding more DNS servers scale, and where does it cross the economic threshold?

Key findings

• The benchmark for economic viability is 10 ms of latency saved per kilobyte of added traffic, even under pessimistic cost estimates for both clients and servers.
• In realistic scenarios—such as clients using DSL and servers using external hosting—the required latency savings threshold drops to as low as 0.25 ms/KB.
• Redundant DNS requests to 10 or more servers yield a 6–15% improvement in mean and 95th percentile page load times, translating to 200–700 ms and 500–2300 ms absolute improvements, respectively.
• The incremental benefit of adding each additional DNS server diminishes, with 4-way or higher replication being at best economically neutral due to diminishing returns.
• Replicating DNS requests to 2–3 servers provides a net benefit for mean latency, but higher redundancy levels may only be justified in tail latency scenarios, where data remain too noisy for conclusive analysis.
• The framework confirms that redundant DNS requests to up to 10 servers are cost-effective across all tested configurations, as they exceed the 10 ms/KB benchmark in both mean and tail latency metrics.