Skip to content

README.md

Latest commit

 

History

History
66 lines (50 loc) · 3.3 KB

README.md

File metadata and controls

66 lines (50 loc) · 3.3 KB
Raw

Alibaba Cluster Trace GPU v2026

Introduction

This release provides an anonymized GPU cluster trace from Alibaba Serverless Infrastructure (ASI), a hyperscale production AI cluster that serves heterogeneous workloads including ad-hoc development, training, online inference, and offline inference. The trace spans about six months and, at the hourly peak, covers 155,410 GPUs across 37,707 GPU servers.

The trace accompanies the paper Heterogeneity at Hyperscale: Characterization and Scheduling of Large Production AI Clusters at Alibaba. The paper studies six months of production behavior from a large heterogeneous GPU fleet, including workload mix, GPU and CPU requests, priority classes, resource utilization, network traffic, standby capacity, and scheduler-facing topology. The released dataset is intended to support research on AI cluster characterization, heterogeneous GPU scheduling, resource fragmentation, priority-aware resource sharing, and trace-driven simulation.

If you use this trace, please cite the paper:

@inproceedings{asi_trace_2026,
  title = {Heterogeneity at Hyperscale: Characterization and Scheduling of Large Production AI Clusters at Alibaba},
  author = {Suyi Li and Lingyun Yang and Haoxuan Yu and Sheng Yao and Tianyuan Wu and Xiaoxiao Jiang and Hanfeng Lu and Kangjin Wang and Chenhao Wang and Shenglin Xu and Lun Wang and Qingyang Duan and Shenghao Liang and Xiu Lin and Wenchao Wu and Yinghao Yu and Guodong Yang and Liping Zhang and Wei Wang},
  booktitle = {20th USENIX Symposium on Operating Systems Design and Implementation (OSDI 26)},
  year = {2026}
}

Data Download

The fact tables are hosted separately from GitHub at:

https://tre-clusterdata.oss-cn-hangzhou.aliyuncs.com/cluster-trace-gpu-v2026/data/

Download the four ZIP archives, place them under data/, and unzip them before running the reproduction workflow. The archive list, sizes, and expected directory layout are documented in docs/data_download.md.

Dataset Contents

The public release contains four fact tables:

Table Coverage Purpose
asi_opensource_pod_hourly day=0..184, all 24 hours Hourly pod workload, request, utilization, priority, state, and public job/model buckets.
asi_opensource_server_hourly day=0..184, all 24 hours Hourly server inventory and anonymized ASW/rack topology.
asi_opensource_network_hourly day=109..115, all 24 hours Hourly node-level receive/transmit traffic samples for the network figure.
asi_opensource_job_execution_summary Non-partitioned Pod/workload execution-span summary for execution-time CDFs.

Time is relative. day=0 is the trace start day. The released fact tables do not contain calendar dates, raw pod names, raw server serial numbers, raw cluster names, raw owner names, raw pod YAML, or internal organization fields.

Detailed field definitions are in docs/schema.md.

Reproducing Figures

The release includes scripts for rebuilding the aggregate inputs and trace-data figures used in the paper from the public fact tables. Detailed commands, figure-to-script mapping, and derived input locations are documented in docs/figure_reproduction.md.