TY - JOUR
T1 - Learning While Scheduling in Multi-Server Systems With Unknown Statistics
T2 - 26th International Conference on Artificial Intelligence and Statistics, AISTATS 2023
AU - Yang, Zixian
AU - Srikant, R.
AU - Ying, Lei
N1 - The work of Zixian Yang and Lei Ying is supported in part by NSF under grants 2001687, 2112471, 2207548, and 2228974. The work of R. Srikant is supported in part by NSF CCF 22-07547, NSF CNS 21-06801, NSF CCF 1934986, ONR N00014-19-1-2566, and ARO W911NF-19-1-0379.
PY - 2023
Y1 - 2023
N2 - Multi-server queueing systems are widely used models for job scheduling in machine learning, wireless networks, and crowdsourcing. This paper considers a multi-server system with multiple servers and multiple types of jobs, where different job types require different amounts of processing time at different servers. The goal is to schedule jobs on servers without knowing the statistics of the processing times. To fully utilize the processing power of the servers, it is known that one has to at least learn the service rates of different job types on different servers. Prior works on this topic decouple the learning and scheduling phases which leads to either excessive exploration or extremely large job delays. We propose a new algorithm, which combines the MaxWeight scheduling policy with discounted upper confidence bound (UCB), to simultaneously learn the statistics and schedule jobs to servers. We obtain performance bounds for our algorithm that hold for both stationary and nonstationary service rates. Simulations confirm that the delay performance of our algorithm is several orders of magnitude better than previously proposed algorithms. Our algorithm also has the added benefit that it can handle non-stationarity in the service processes.
AB - Multi-server queueing systems are widely used models for job scheduling in machine learning, wireless networks, and crowdsourcing. This paper considers a multi-server system with multiple servers and multiple types of jobs, where different job types require different amounts of processing time at different servers. The goal is to schedule jobs on servers without knowing the statistics of the processing times. To fully utilize the processing power of the servers, it is known that one has to at least learn the service rates of different job types on different servers. Prior works on this topic decouple the learning and scheduling phases which leads to either excessive exploration or extremely large job delays. We propose a new algorithm, which combines the MaxWeight scheduling policy with discounted upper confidence bound (UCB), to simultaneously learn the statistics and schedule jobs to servers. We obtain performance bounds for our algorithm that hold for both stationary and nonstationary service rates. Simulations confirm that the delay performance of our algorithm is several orders of magnitude better than previously proposed algorithms. Our algorithm also has the added benefit that it can handle non-stationarity in the service processes.
UR - http://www.scopus.com/inward/record.url?scp=85164013125&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85164013125&partnerID=8YFLogxK
M3 - Conference article
AN - SCOPUS:85164013125
SN - 2640-3498
VL - 206
SP - 4275
EP - 4312
JO - Proceedings of Machine Learning Research
JF - Proceedings of Machine Learning Research
Y2 - 25 April 2023 through 27 April 2023
ER -