The Tail at Scale.¶
Dean, J., & Barroso, L. A. (2013). The Tail at Scale. Communications of the ACM, 56(2), 74-80.
Cited by¶
4 citations across 4 artifacts.
Each citation links to the sentence it supports in the citing article.
Primes¶
- Aggregate-Marginal Divergence
- Software performance. Average latency improves while tail latency worsens, so aggregate experience trends favourably while the worst-served users abandon the product.
This sourceEstablishes that tail latency can worsen while average latency improves, and why worst-served users dominate aggregate service experience.
- Software performance. Average latency improves while tail latency worsens, so aggregate experience trends favourably while the worst-served users abandon the product.
- Interference and Contention
- Understanding the contention tolerance of a given application is as important as measuring the contention itself, as Dean and Barroso (2013) argue in their analysis of the tail at scale, where eliminating contention for one cohort routinely amplifies it for others.
This sourceCanonical analysis of latency variability in large-scale services: shows how contention-mitigation strategies for some consumers redistribute interference cost to others, amplifying tail latency at scale.
- Understanding the contention tolerance of a given application is as important as measuring the contention itself, as Dean and Barroso (2013) argue in their analysis of the tail at scale, where eliminating contention for one cohort routinely amplifies it for others.
- Unevenness Waste
- In software it is the bursty request stream against a fixed thread pool that produces tail latency far worse than average latency, so capacity provisioned to mean throughput is unusable.
This sourceDocuments how request burstiness against shared service capacity produces tail latency far worse than mean latency, so provisioning to the mean is inadequate.
- In software it is the bursty request stream against a fixed thread pool that produces tail latency far worse than average latency, so capacity provisioned to mean throughput is unusable.
Domain-specific¶
- Fallacy of Zero Latency
- With tail latency effects — where the slowest of n calls determines the total — the share of requests that breach the latency budget rises steeply with n, even when no single call is usually slow
This sourceIntroduces hedged requests and tied requests by name as tail-tolerance techniques in Google's large fan-out serving systems, with measured reductions in tail latency.
- With tail latency effects — where the slowest of n calls determines the total — the share of requests that breach the latency budget rises steeply with n, even when no single call is usually slow
Verification¶
This reference passed the adversarial substantiation pipeline: it was checked to exist and to support the claim it is attached to. See how references were verified.
Registry ID ref:35e809461ab5 · see in the full table