How Cloud Gaming Tackles Resource Management: Scheduling, Costs, and AI Solutions

Introduction

Machine learning and artificial intelligence are increasingly applied to resource management. Professors Wang Gang and Li Yusen from Nankai University discuss their research on large‑scale distributed system resource management, covering cloud gaming request scheduling, resource allocation, and load balancing in search‑engine data centers.

Challenges

Key challenges in cloud gaming include high latency due to round‑trip client‑server communication, bandwidth requirements of at least 3–5 Mbps for high‑quality video, and substantial server and operational costs because each server can host only a limited number of concurrent games.

Research Goal

The objective is to minimize the total running time of all servers while satisfying user requirements, effectively reducing overall cost.

Problem Definition

How to schedule game requests so that the cumulative server runtime is minimized, considering online arrival, unknown request completion times, and the inability to migrate running games.

Related Problem

The problem resembles the dynamic bin‑packing problem, where items (requests) arrive and depart over time, and the goal is to keep the maximum number of simultaneously used servers as low as possible.

Research Work

The study is divided into three parts: (1) classic scheduling algorithms, (2) new algorithms based on request‑end‑time prediction, and (3) extensions addressing game deployment overhead.

Classic Algorithms

Any Fit – launch a new server only when all running servers are fully loaded. First Fit – assign a request to the earliest‑available server with free capacity. Best Fit – place a request on the server with the smallest remaining free space, aiming to keep servers as full as possible.

Analysis of Classic Algorithms

Worst‑Case Ratio : First Fit has a bounded worst‑case performance, while Best Fit can be unboundedly bad. Stochastic Analysis : Average performance is evaluated using Markov and other stochastic models.

New Algorithms

Observations of daily request patterns show a peak in the morning and a decline after noon. By predicting request end times, a new scheduling algorithm groups requests with similar expected completion times onto the same server, reducing idle server time.

Request End‑Time Prediction

Experiments on games such as DotAlicious and WOT demonstrate low prediction error rates, indicating that game session lengths are highly predictable, which enables the proposed scheduling approach.

Deployment Cost Extension

Installing games on each server incurs storage overhead. Strategies include shared network storage (often impractical) or assigning distinct game sets to different servers, aiming to minimize total installation cost while ensuring request fulfillment.

Summary

Cloud gaming concept and challenges

Research goal: minimize total server runtime

Problem definition and similarity to dynamic bin packing

Evaluation of classic algorithms (Any Fit, First Fit, Best Fit)

Development of prediction‑based scheduling algorithms

Analysis of game deployment overhead

Next Topics

Future articles will cover resource management in search‑engine data centers and AI applications in resource management.