Salesforce/cloudops_tsf

Pushing the Limits of Pre-training for Time Series Forecasting in the CloudOps Domain

Paper | Code

Datasets accompanying the paper "Pushing the Limits of Pre-training for Time Series Forecasting in the CloudOps Domain".

Quick Start

azure_vm_traces_2017

from datasets import load_dataset

dataset = load_dataset('Salesforce/cloudops_tsf', 'azure_vm_traces_2017')
print(dataset)

DatasetDict({
    train_test: Dataset({
        features: ['start', 'target', 'item_id', 'feat_static_cat', 'feat_static_real', 'past_feat_dynamic_real'],
        num_rows: 17568
    })
    pretrain: Dataset({
        features: ['start', 'target', 'item_id', 'feat_static_cat', 'feat_static_real', 'past_feat_dynamic_real'],
        num_rows: 159472
    })
})

borg_cluster_data_2011

dataset = load_dataset('Salesforce/cloudops_tsf', 'borg_cluster_data_2011')
print(dataset)

DatasetDict({
    train_test: Dataset({
        features: ['start', 'target', 'item_id', 'feat_static_cat', 'past_feat_dynamic_real'],
        num_rows: 11117
    })
    pretrain: Dataset({
        features: ['start', 'target', 'item_id', 'feat_static_cat', 'past_feat_dynamic_real'],
        num_rows: 143386
    })
})

alibaba_cluster_trace_2018

dataset = load_dataset('Salesforce/cloudops_tsf', 'alibaba_cluster_trace_2018')
print(dataset)

DatasetDict({
    train_test: Dataset({
        features: ['start', 'target', 'item_id', 'feat_static_cat', 'past_feat_dynamic_real'],
        num_rows: 6048
    })
    pretrain: Dataset({
        features: ['start', 'target', 'item_id', 'feat_static_cat', 'past_feat_dynamic_real'],
        num_rows: 58409
    })
})

Dataset Config

from datasets import load_dataset_builder

config = load_dataset_builder('Salesforce/cloudops_tsf', 'azure_vm_traces_2017').config
print(config)

CloudOpsTSFConfig(
    name='azure_vm_traces_2017',
    version=1.0.0,
    data_dir=None,
    data_files=None,
    description='',
    prediction_length=48,
    freq='5T',
    stride=48,
    univariate=True,
    multivariate=False,
    optional_fields=(
        'feat_static_cat',
        'feat_static_real',
        'past_feat_dynamic_real'
    ),
    rolling_evaluations=12,
    test_split_date=Period('2016-12-13 15:55', '5T'),
    _feat_static_cat_cardinalities={
        'pretrain': (
            ('vm_id', 177040),
            ('subscription_id', 5514),
            ('deployment_id', 15208),
            ('vm_category', 3)
        ),
        'train_test': (
            ('vm_id', 17568),
            ('subscription_id', 2713),
            ('deployment_id', 3255),
            ('vm_category', 3)
        )
    },
    target_dim=1,
    feat_static_real_dim=3,
    past_feat_dynamic_real_dim=2
)

test_split_date is provided to achieve the same train-test split as given in the paper. This is essentially the date/time of rolling_evaluations * prediction_length time steps before the last time step in the dataset. Note that the pre-training dataset includes the test region, and thus should also be filtered before usage.

Acknowledgements

The datasets were processed from the following original sources. Please cite the original sources if you use the datasets.

• Azure VM Traces 2017

  • Bianchini. Resource central: Understanding and predicting workloads for improved resource management in large cloud platforms. In Proceedings of the 26th Symposium on Operating Systems Principles, pp. 153–167, 2017.
  • https://github.com/Azure/AzurePublicDataset

• Borg Cluster Data 2011

  • John Wilkes. More Google cluster data. Google research blog, November 2011. Posted at http://googleresearch.blogspot.com/2011/11/more-google-cluster-data.html.
  • https://github.com/google/cluster-data

• Alibaba Cluster Trace 2018

  • Jing Guo, Zihao Chang, Sa Wang, Haiyang Ding, Yihui Feng, Liang Mao, and Yungang Bao. Who limits the resource efficiency of my datacenter: An analysis of alibaba datacenter traces. In Proceedings of the International Symposium on Quality of Service, pp. 1–10, 2019.
  • https://github.com/alibaba/clusterdata

Citation

@article{woo2023pushing,
  title={Pushing the Limits of Pre-training for Time Series Forecasting in the CloudOps Domain},
  author={Woo, Gerald and Liu, Chenghao and Kumar, Akshat and Sahoo, Doyen},
  journal={arXiv preprint arXiv:2310.05063},
  year={2023}
}