A Fast and Scalable Feedback-Driven Scheduler for Datacenter Applications

Microsecond-scale datacenter applications demand strict latency guarantees while operating under high load and variable service times. This environment often involves a mix of extremely short and long requests, where short requests — lasting just a few microseconds — are frequently delayed by longer...

Full description

Bibliographic Details
Author: MAYCO SOUZA BERGHETTI
Format: master thesis
Status:Published version
Publication Date:2025
Country:Brasil
Institution:Universidade Federal de Mato Grosso do Sul (UFMS)
Repository:Repositório Institucional da UFMS
Language:Portuguese
OAI Identifier:oai:repositorio.ufms.br:123456789/12732
Online Access:https://repositorio.ufms.br/handle/123456789/12732
Access Level:Open access
Keyword:Datacenter
Head-of-Line Blocking
User-Level Scheduler
Description
Summary:Microsecond-scale datacenter applications demand strict latency guarantees while operating under high load and variable service times. This environment often involves a mix of extremely short and long requests, where short requests — lasting just a few microseconds — are frequently delayed by longer ones due to Head-of-Line (HOL) blocking, leading to higher latencies, especially at the tail. However, existing approaches to mitigate HOL blocking, such as centralized dispatching, fine-grained preemption, and resource reservation, face fundamental scalability limitations. This work introduces Synergy, a cooperative, application-aware scheduling system that uses direct feedback from applications to prioritize short requests, dynamically adapts scheduling parameters, and avoids unnecessary preemptions. Synergy adopts a decentralized architecture with distributed queues, job-aware preemption, and dynamic quantum sizing. By eliminating centralized classification and using real-time application measurements, Synergy effectively mitigates HOL blocking without compromising throughput. Synergy outperforms state-of-the-art systems, achieving up to 43% higher throughput while meeting microsecond-scale service-level objectives.