Tiresias -- A GPU Cluster Manager for Distributed Deep Learning Training without complete job information
Tiresias is a GPU cluster resource manager that aims at minimizing distributed deep learning (DDL) jobs’ completion times with partial or no a priori knowledge. It does not rely on any intermediate DL algorithm states (e.g., training loss values) or framework specifics (e.g., tensors-to-parameter server mapping).
DDL training jobs bring some unique challenges to the cluster manager:
- unpredictable training time
- over-aggressive job consolidation
- all-or-nothing resource allocation
- inflexibility in GPU sharing (job preemption and resumption)
Tiresias tackles those challenges with the Discretized-2DAS (two-dimensional age/attained-service based) scheduler and the model profile-based job placement scheme. The 2DAS scheduler, which considers both the spatial (GPU requirements) and temporal (job's executed time) aspects of DDL jobs, has two scheduling algorithms (Discretized 2D-LAS and Discretized 2D-Gittins Index). They can minimize the average JCT with no and partial job knowledge, respectively. The profile-based job placement scheme can appropriately relax the consolidation constraints and maintain the resource (GPU) utilization of cluster without hurting jobs’ performance.
Out testbed experiments and large-scale trace-driven simulations show that Tiresias improves the average JCT by up to 5.5x (2x) over current production solutions (state-of-the-art DDL cluster scheduler), and it performs comparably to the solution using perfect knowledge of all job characteristics.
Detailed design and performance are available in our NSDI'19 paper.
What's in this repository?
- Discrete-time simulator of GPU cluster manager for DL training jobs (with both the job scheduler and placement scheme)
Coming soon ...
Network(RDMA)-level message profiler for DL models
What's LAS (Least-Attained Service) algorithm?
Nuyens, Misja, and Adam Wierman. "The foreground–background queue: a survey." Performance evaluation 65.3-4 (2008): 286-307.
What's Gittins Index policy?
Gittins, John, Kevin Glazebrook, and Richard Weber. Multi-armed bandit allocation indices. John Wiley & Sons, 2011.