Locality Aware Dynamic Load Management for Massively Multiplayer Games Jin Chen, Baohua Wu, Margaret Delap, Bjorn Knutson, Honghui Lu and Cristina Amza presented by Sagnik Nandy
Basic Idea How to schedule game regions across multiple servers in a massively parallel multiplayer game environment?
Overview Problem Description Existing Techniques Suggested Solution Experimental Results Conclusion
Overview Problem Description Existing Techniques Suggested Solution Experimental Results Conclusion
Problem Description How do you map various regions of a multiplayer game across different servers?
Issue 1 - Locality 1 1
1 1
Issue 2 – Load balancing
Problem Statement Balance server load by replicating existing game world partitions across several servers Decrease inter-server communication by maintaining locality of adjacent regions
Overview Problem Description Existing Techniques Suggested Solution Experimental Results Conclusion
Existing Solutions Built-in load balancing in the game concept (e.g. countries, airports etc.) Static Partitioning – row based, column based, cyclic, etc. Dynamic Uniform Load Spread (Spread) Tries to minimize the difference between most and least loaded nodes Doesn’t consider locality
Existing Solutions (contd.) Dynamic Load Shedding to Lightest Loaded Node (Lightest) Choose loaded server and shed load to system-wide lightest loaded node Locality is not an objective (but can get maintained)
Suggested Solution (Locality Aware Dynamic Load Management) SLA violation 90% users exceed update interval Overload threshold load (# users) for which violation happens Safe load threshold max load for which all users meet SLA Light load 2*safe_load – over_load
Objectives Meet SLA (= load balancing) Happy users Maintain locality of game regions Reduce transition time Minimize # of region migrations Reduce inter-server communication
Overview Problem Description Existing Techniques Suggested Solution Experimental Results Conclusion
Suggested Approach Load shedding algorithm How to distributed load and meet SLA requirements Load aggregation algorithm Help restore locality Help in future load shedding
Load Shedding Algorithm If load > over_load While load > over_load Find lightest (neighbor < safety_load) and shed load If no neighbor exists then do this globally across system
Shed Load How to choose a component to shed? Given a neighbor S j Choose a boundary node for S j With node as root Find strongly connected cluster using BFS as long cluster weight within bounds
Load Aggregation Reasons Load can be shed to remote server Load can be shed across multiple neighbors Tries to reduce number of boundaries For each neighbor of S i Find partition such that new_load < safe_load Transfer cluster if boundaries reduce
Overview Problem Description Existing Techniques Suggested Solution Experimental Results Conclusion
Experiments First did single server and a smaller cluster based experiment Used results to simulate more comprehensive system Simulated for CPU and network usage Simulated for a LAN and WAN setting
Real Experiments (single server)
Real Experiments (multiple server)
Simulation results (LAN)
Simulation results (WAN)
Conclusions The paper introduces the issue of locality into scheduling Dynamic scheduling is better than static scheduling Locality is more important as the network spreads out (curious to know effect on Internet scale games) Aggregation didn’t help much