提出GetBatch，用批量请求替代单个文件获取，显著提升机器学习数据加载速度。
GetBatch: Distributed Multi-Object Retrieval for ML Data Loading
- 设计新API，将多个文件获取合并为一次批量操作，减少请求开销。
- 小文件场景下吞吐提升最高达15倍，生产负载中95%延迟减半。
- 适合大规模分布式训练的数据加载场景，尤其关注低延迟需求。
机器学习训练流程以批为单位读取数据，单次训练步骤可能需要从存储集群中多个分片获取数千个样本。若对每个样本单独发起GET请求，请求开销常成为数据传输时间的主要瓶颈。为此，我们提出GetBatch——一种新的对象存储API，将批量获取提升为第一类存储操作，用一次确定性、容错的流式执行取代独立的GET操作。GetBatch在小文件场景下实现最高15倍的吞吐提升；在真实训练负载中，相比独立GET请求，批次检索的P95延迟降低2倍，每对象尾部延迟（P99）降低3.7倍。
原文摘要 · Abstract (English)
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.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。