谷歌云宣布针对毫秒级延迟工作负载的快速存储

发布于 5 月 10 日

At Google Cloud Next 2025, Google announced Rapid Storage, a new Cloud Storage zonal bucket for frequently accessed data and latency-sensitive apps, offering under 1ms random read/write latency, 20x faster data access, and 6 TB/s throughput.
- It keeps data within the same zone as GPUs and TPUs, helping workloads achieve higher throughput and sub-millisecond latencies, unlike traditional buckets spanning multiple zones.
Rapid Storage can be mounted as a file system using Cloud Storage FUSE and supports common AI frameworks like TensorFlow and PyTorch.
- In the article "Colossus: the secret ingredient in Rapid Storage’s high performance", Google explained how they added a stateful gRPC-based streaming protocol and the ability to append data while maintaining high aggregate throughput and scale.
In a Reddit thread, the community compared Rapid Storage with Amazon S3 Express. Akshay Shah commented that Google is the only major cloud with low-latency single-zone object storage, standard regional storage, and transparently-replicated dual-region storage.
The same week, AWS announced reduced storage prices for S3 Express One Zone. Agarwal and Khan claimed that Rapid Storage delivers 5x lower latency for random reads and writes compared to other hyperscalers, with up to 6 TB/s per bucket and 20 million QPS.
Rapid Storage was not the only storage announcement at Cloud Next 2025 targeting AI workloads. Anywhere Cache is an SSD cache working with existing regional buckets, and Google Cloud Managed Lustre is a high-performance parallel file system providing PB scale at under 1ms latency, millions of IOPS, and TB/s throughput.
Currently, Rapid Storage is in preview.

阅读 9