QuADD: QUantifying Accelerator Disaggregated Datacenter Efficiency
Anubhav Guleria, J. Lakshmi, Chakri Padala · 2019
In the current era of data explosion accelerators such as GPUs facilitate data-driven applications with requisite compute boost. Availability of GPUs in Public Cloud offerings has expedited their mass adoption. Consequently, varied customer demands and exclusive allocation of GPUs to VMs can leave stranded GPUs across the datacenter. Hardware disaggregation can alleviate this issue to enable a powerefficient datacenter. However, it is important to first quantify the gains associated with this new deployment paradigm. In the absence of real deployments, simulations can be helpful to evaluate the benefits of disaggregation at scale. In this paper, we evaluate the gains associated primarily with disaggregated GPU deployments. For this, we use QUADD-SIM, a simulator we built to model, quantify, and contrast different facets of these emerging GPU deployments. Using QUADD-SIM we model different VM and resource provisioning aspects of disaggregated GPU deployments. We simulate realistic AI workload requests for a period of 3 months with characteristics derived from recent public datacenter traces. Our results attest that disaggregated GPU deployment strategies outperform traditional GPU deployments in terms of failed VM requests and GPU Watt-hours consumption. Overall, 5.14% and 7.90% additional failed VM requests were serviced by disaggregated GPU deployments consuming 10.92% and 3.30% lesser GPU Watt-hours compared to traditional deployment.