Search for: All records

Today's data centers often need to run various machine learning (ML) applications with stringent SLO (Service-Level Objective) requirements, such as inference latency. To that end, data centers prefer to 1) over-provision the number of servers used for inference processing and 2) isolate them from other servers that run ML training, despite both use GPUs extensively, to minimize possible competition of computing resources. Those practices result in a low GPU utilization and thus a high capital expense. Hence, if training and inference jobs can be safely co-located on the same GPUs with explicit SLO guarantees, data centers could flexibly run fewer training jobs when an inference burst arrives and run more afterwards to increase GPU utilization, reducing their capital expenses. In this paper, we propose GPUColo, a two-tier co-location solution that provides explicit ML inference SLO guarantees for co-located GPUs. In the outer tier, we exploit GPU spatial sharing to dynamically adjust the percentage of active GPU threads allocated to spatially co-located inference and training processes, so that the inference latency can be guaranteed. Because spatial sharing can introduce considerable overheads and thus cannot be conducted at a fine time granularity, we design an inner tier that puts training jobs into periodic sleep, so that the inference jobs can quickly get more GPU resources for more prompt latency control. Our hardware testbed results show that GPUColo can precisely control the inference latency to the desired SLO, while maximizing the throughput of the training jobs co-located on the same GPUs. Our large-scale simulation with a 57-day real-world data center trace (6500 GPUs) also demonstrates that GPUColo enables latency-guaranteed inference and training co-location. Consequently, it allows 74.9% of GPUs to be saved for a much lower capital expense. 

Power capping is an important technique for high-density servers to safely oversubscribe the power infrastructure in a data center. However, power capping is commonly accomplished by dynamically lowering the server processors’ frequency levels, which can result in degraded application performance. For servers that run important machine learning (ML) applications with Service-Level Objective (SLO) requirements, inference performance such as recognition accuracy must be optimized within a certain latency constraint, which demands high server performance. In order to achieve the best inference accuracy under the desired latency and server power constraints, this paper proposes OptimML, a multi-input-multi-output (MIMO) control framework that jointly controls both inference latency and server power consumption, by flexibly adjusting the machine learning model size (and so its required computing resources) when server frequency needs to be lowered for power capping. Our results on a hardware testbed with widely adopted ML framework (including PyTorch, TensorFlow, and MXNet) show that OptimML achieves higher inference accuracy compared with several well-designed baselines, while respecting both latency and power constraints. Furthermore, an adaptive control scheme with online model switching and estimation is designed to achieve analytic assurance of control accuracy and system stability, even in the face of significant workload/hardware variations.