arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2602.22434cs.DCcs.AIcs.DBcs.LG
收藏

GetBatch: 分布式多对象检索用于机器学习数据加载

GetBatch: Distributed Multi-Object Retrieval for ML Data Loading

NVIDIA, Santa Clara, USA(NVIDIA公司)

Alex Aizman, Abhishek Gaikwad, Piotr Żelasko

AI总结:

GetBatch通过批量检索优化机器学习数据加载,提升吞吐量并降低延迟

摘要:

机器学习训练流水线以批次方式消耗数据。单个训练步骤可能需要从分布在存储集群中的分片中抽取数千个样本。发出数千个单独的GET请求会带来每个请求的开销,这种开销通常主导数据传输时间。为了解决这个问题,我们引入了GetBatch——一个新型的对象存储API,将批量检索提升为第一类存储操作,用单个确定性、容错的流式执行取代独立的GET操作。GetBatch在小型对象上实现了高达15倍的吞吐量提升,并在生产训练工作负载中,将P95批次检索延迟减少2倍,将P99单个对象尾部延迟减少3.7倍,相比单独的GET请求。

英文摘要:

Machine learning training pipelines consume data in batches. A single training step may require thousands of samples drawn from shards distributed across a storage cluster. Issuing thousands of individual GET requests incurs per-request overhead that often dominates data transfer time. To solve this problem, we introduce GetBatch - a new object store API that elevates batch retrieval to a first-class storage operation, replacing independent GET operations with a single deterministic, fault-tolerant streaming execution. GetBatch achieves up to 15x throughput improvement for small objects and, in a production training workload, reduces P95 batch retrieval latency by 2x and P99 per-object tail latency by 3.7x compared to individual GET requests.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑