Extending Delayed Fair Sharing: A Generalizable Framework for Multi-Resource Performance Isolation

James DeLoye, Tyler Griggs, Dev Bali, Wenjie Ma, Audrey Cheng, Jae Hong, Soujanya Ponnapalli, Natacha Crooks, Scott Shenker, Ion Stoica and Matei Zaharia

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2025-97
May 16, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-97.pdf

Modern storage systems, often deployed to support multiple tenants, must provide performance isolation. Traditional approaches, like fair sharing resources, work poorly for these systems as storage resources exhibit high preemption delays. These delays force clients to wait to receive their share of resources, and lead to unacceptable spikes in tail latency. Delayed Fair Sharing is a new fairness definition that explicitly bounds the delay a client incurs to receive its fair share of resources. DFS originally proposed policies that bound the preemption delays of commonly contended storage resources, such as write buffers, and read caches and implemented them in FairDB, an extension of RocksDB. In this work we attempt to extend Delayed Fair Sharing to other candidate resources, specifically the RocksDB write-ahead log (WAL) and flush threadpool, and implement these extensions in FairDB and evaluate their effectiveness. FairDB ensures that clients receive their resource shares within a configurable time window (δ) and bounds the resource acquisition delays end-to-end. Here we show that this logic can be extend to more resources than those in the original work and effectively isolate clients from bursty or aggressive clients in the same database.

Advisor: Joseph Gonzalez

"; ?>

BibTeX citation:

@mastersthesis{DeLoye:EECS-2025-97,
    Author = {DeLoye, James and Griggs, Tyler and Bali, Dev and Ma, Wenjie and Cheng, Audrey and Hong, Jae and Ponnapalli, Soujanya and Crooks, Natacha and Shenker, Scott and Stoica, Ion and Zaharia, Matei},
    Editor = {Gonzalez, Joseph},
    Title = {Extending Delayed Fair Sharing: A Generalizable Framework for Multi-Resource Performance Isolation},
    School = {EECS Department, University of California, Berkeley},
    Year = {2025},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-97.html},
    Number = {UCB/EECS-2025-97},
    Abstract = {Modern storage systems, often deployed to support multiple tenants, must provide performance isolation. Traditional approaches, like fair sharing resources, work poorly for these systems as storage resources exhibit high preemption delays. These delays force clients to wait to receive their share of resources, and lead to unacceptable spikes in tail latency.
Delayed Fair Sharing is a new fairness definition that explicitly bounds the delay a client incurs to receive its fair share of resources. DFS originally proposed policies that bound the preemption delays of commonly contended storage resources, such as write buffers, and read caches and implemented them in FairDB, an extension of RocksDB. In this work we attempt to extend Delayed Fair Sharing to other candidate resources, specifically the RocksDB write-ahead log (WAL) and flush threadpool, and implement these extensions in FairDB and evaluate their effectiveness.
FairDB ensures that clients receive their resource shares within a configurable time window (δ) and bounds the resource acquisition delays end-to-end. Here we show that this logic can be extend to more resources than those in the original work and effectively isolate clients from bursty or aggressive clients in the same database.}
}

EndNote citation:

%0 Thesis
%A DeLoye, James
%A Griggs, Tyler
%A Bali, Dev
%A Ma, Wenjie
%A Cheng, Audrey
%A Hong, Jae
%A Ponnapalli, Soujanya
%A Crooks, Natacha
%A Shenker, Scott
%A Stoica, Ion
%A Zaharia, Matei
%E Gonzalez, Joseph
%T Extending Delayed Fair Sharing: A Generalizable Framework for Multi-Resource Performance Isolation
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 16
%@ UCB/EECS-2025-97
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-97.html
%F DeLoye:EECS-2025-97