Overview

This repository contains public releases of SenseTime Helios traces for the benefit of the deep learning system research community.

If you do use the Helios traces in your research, please make sure to cite our SC '21 paper "Characterization and Prediction of Deep Learning Workloads in Large-Scale GPU Datacenters", which includes a comprehensive analysis of the deep learning workloads in Helios from April 2020 to September 2020.

We encourage anyone to use the traces for academic purposes, and if you had any questions, feel free to send an email to us, or file an issue on Github.

Note that only the Venus trace is public available now. Other traces are being censored. We will release them as soon as possible.

Helios Description

Helios is a private datacenter dedicated to developing DL models for research and production in SenseTime. It contains 8 independent GPU clusters and over 12,000 GPUs in total.

In this repository, we publicly release the workload trace in 4 representative GPU clusters: Earth, Saturn, Uranus, and Venus. You can find a detailed description of the Helios datacenter in the SC '21 paper mentioned above.

Besides, we also release the analysis scripts for Helios traces in HeliosArtifact.

Helios Dataset

The main trace characteristics, dataset structure and schema are:

Main Characteristics:

• Dataset size: 343MB
• Compressed dataset size: 36.4MB
• Number of files: 8
• Duration: 6 months
• Number of independent GPU clusters: 4
• Total number of jobs: 3,362,981
• Total number of GPU jobs: 1,580,464

Dataset Structure

Each cluster provides a job trace file (cluster_log.csv) and a VC configuration file (cluster_gpu_number.csv).

📦data
 ┣ 📂Earth
 ┃ ┣ 📜cluster_gpu_number.csv
 ┃ ┗ 📜cluster_log.csv
 ┣ 📂Saturn
 ┃ ┣ 📜cluster_gpu_number.csv
 ┃ ┗ 📜cluster_log.csv
 ┣ 📂Uranus
 ┃ ┣ 📜cluster_gpu_number.csv
 ┃ ┗ 📜cluster_log.csv
 ┗ 📂Venus
 ┃ ┣ 📜cluster_gpu_number.csv
 ┃ ┗ 📜cluster_log.csv

Schema and Description

cluster_log.csv

Description

Provides rich information on all jobs submitted to Slurm in each cluster.

Example

job_id	user	vc	gpu_num	cpu_num	node_num	state	submit_time	start_time	end_time	duration	queue
1425511	uXBbc	vcJkd	1	1	1	COMPLETED	2020-06-09 18:41:01	2020-06-09 18:41:01	2020-06-10 04:55:09	36848	0
1425512	uVMrF	vchbv	4	16	1	FAILED	2020-06-09 18:41:27	2020-06-09 18:41:27	2020-06-09 18:45:36	249	0
1425513	uzqls	vcpDC	1	1	1	CANCELLED	2020-06-09 18:41:28	2020-06-09 18:41:28	2020-06-17 14:15:21	675233	0

Schema

Field	Description
job_id	unique id of the job 1
user	hashed id for the user, prefix is 'u'
vc	hashed id for the virtual cluster, prefix is 'vc'
gpu_num	number of GPUs required for the job
cpu_num	number of CPUs required for the job
node_num	number of nodes in the job
state	the job's status upon termination 2
submit_time	the job's submission time
start_time	the job's start execution time
end_time	the job's termination time
duration	total job execution time of the job 3
queue	total job queue time of the job 4

Notes

1. job_id is generated by Slurm and reflects the job submission order in each cluster.
2. A job can end up with one of five statuses: (1) COMPLETED: it is finished successfully; (2) CANCELLED: it is terminated by the user; (3) FAILED: it is terminated due to internal or external errors; (4) TIMEOUT: the execution time is out of limit; (5) NODE_FAIL: it is terminated due to the node crash. TIMEOUT and NODE_FAIL are very rare in our traces, and are regarded as failed in our analysis. (Another status SUSPENDED happens only once in cluster Uranus, so we ignore it.)
3. Calculated from the difference between end_time and start_time. (Unit: seconds)
4. Calculated from the difference between start_time and submit_time. (Unit: seconds)

cluster_gpu_number.csv

Description

Lists the number of GPUs per day in each VC.

Example

date	vchbv	vc4om	vcVP5	vc6YE	vchA3	vccaA	vcTJs	vcvlY	vcSoL	vcMod	vcpDC	vc3sl	vc8Sj	vcJLV	vcLJZ	vcIya	vcJkd	vcdI0	vcira	vcgkz	vcxS0	vc7hD	vcXrB	vcvcM	vcp4O	total
2020-09-01	64	96	176	216	40	40	48	96	32	64	56	64	0	32	64	0	16	16	16	8	0	0	0	0	0	1144

Schema

Field	Description
date	record granularity is daily
vc***	the number of GPUs of the VC
total	the total number of GPUs of the cluster