dJay: Enabling High-density Multi-tenancy for Cloud Gaming Servers with Dynamic Cost-Beneο¬t GPU Load Balancing Sergey Grizan, David Chu, Alec Wolman, Roger Wattenhofer Challenge Density Optimization Problem Perceived Quality max π(ππ ) π Challenges β’ Predicting πΊππ_ππΌππΈ ππ β’ Predicting π(ππ ) Constraint: πΊππ_ππΌππΈ(ππ ) < 33ππ π β’ β¦ for arbitrary scenes β’ β¦ for all graphics settings Impacted by: β’ Player churn: join/leave β’ Scene change: outdoors/small room Much lower CapEx and OpEx Christmas Morning: GPU oversubscription!! System Architecture Key Insight: Tunable Visual QoS HQ Texture Detail LQ HI LO Ant-Aliasing Lights Shadows Environment details Fog Tesselation GPU time: 3 ms Perceived quality: 1 GPU time: 1 ms Perceived quality: 0.7 PSNR, SSIM Utility Estimating: Cost and Benefit Results on Fable 3 29 fps 29 fps 40 fps 40 fps 75 fps 0.9999 75 fps Increasing multi-tenancy 51 fps 51 fps 3 players 16 fps 78 fps 2 players 18 fps 76 fps 78 fps 80 fps 0.9 83 fps 0.99 82 fps SSIM score 0.999 0 1 player All HI All ME All LO 4 players dJay 30 fps 30 fps Increasing scene complexity 0.9999 0.9 0 Simple scene 40 fps 51 fps 20 fps 17 fps 0.99 83 fps 49 fps SSIM score 0.999 Complex scene All HI All ME All LO Auto Responsiveness 45 40 critical level 35 25 20 15 10 5 0 0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 64 68 72 76 80 84 88 92 96 100 104 108 112 116 120 124 128 132 136 140 144 148 Frame time, ms 30 Frame number Player 1 Player 2
© Copyright 2026 Paperzz