Key points are not available for this paper at this time.
With widespread advances in machine learning, a number of large enterprises beginning to incorporate machine learning models across a number of. These models are typically trained on shared, multi-tenant GPU. Similar to existing cluster computing workloads, scheduling aim to provide features like high efficiency, resource isolation, sharing across users, etc. However Deep Neural Network (DNN) based, predominantly trained on GPUs, differ in two significant ways from big data analytics workloads. First, from a cluster utilization, GPUs represent a monolithic resource that cannot be shared at a granularity across users. Second, from a workload perspective, deep frameworks require gang scheduling reducing the flexibility of and making the jobs themselves inelastic to failures at runtime. In paper we present a detailed workload characterization of a two-month long from a multi-tenant GPU cluster in a large enterprise. By correlating logs with logs from individual jobs, we study three distinct issues affect cluster utilization for DNN training workloads on multi-tenant: (1) the effect of gang scheduling and locality constraints on, (2) the effect of locality on GPU utilization, and (3) failures during. Based on our experience running a large-scale operation, we provide guidelines pertaining to next-generation cluster schedulers for DNN workloads.
Jeon et al. (Thu,) studied this question.