Search papers, labs, and topics across Lattice.
This paper introduces ShardMeter, an analytical performance model designed to optimize the training of large-scale AI models across sharded and geo-distributed environments. By predicting runtime metrics such as throughput and training costs based on model characteristics and hardware configurations, ShardMeter helps researchers navigate the complex landscape of resource allocation without exhaustive benchmarking. Key findings include the identification of diminishing returns in island size and the quantification of transitions between compute- and communication-bound scaling, enabling more efficient deployment strategies.
ShardMeter reveals that larger training islands can lead to diminishing returns, fundamentally changing how we approach resource allocation in distributed AI training.
Training large-scale AI models often outgrows a single data center, demanding sharded, multi-cluster, and decentralized training. However, the huge space of resource allocations makes exhaustive benchmarking and manual tuning impractical, while performance depends on tightly coupled factors like model size, GPU memory, batch size, bandwidth, and sharding strategy. We introduce ShardMeter, a lightweight analytical performance model that predicts the end-to-end runtime of transformer-based workloads across arbitrary sharded, distributed, and even decentralized training. Given a model's characteristics and a target hardware topology, ShardMeter estimates per-GPU and per-island throughput, training cost, total wall-clock time, and identifies performance bottlenecks. Our analysis reveals diminishing-return regimes as island size increases, quantifies transitions between compute- and communication-bound scaling, evaluates hyperparameter trade-offs, and models cost-throughput for large-scale decentralized training. ShardMeter exposes these insights to quickly explore the configuration space, choose near-optimal deployment plans, and avoid costly trial and error.
