datacentertracesdatasets 0.3.3

Data center traces for machine learning tasks

DataCenter-Traces-Datasets

Pip package that makes available the datasets published in: https://github.com/alejandrofdez-us/DataCenter-Traces-Datasets. Please, check the mentioned repository for a deeper understanding.

Alibaba 2018 machine usage

Processed from the original files found at: https://github.com/alibaba/clusterdata/tree/master/cluster-trace-v2018

This repository dataset of machine usage includes the following columns:

+--------------------------------------------------------------------------------------------+
| Field            | Type       | Label | Comment                                            |
+--------------------------------------------------------------------------------------------+
| cpu_util_percent | bigint     |       | [0, 100]                                           |
| mem_util_percent | bigint     |       | [0, 100]                                           |
| net_in           | double     |       | normarlized in coming network traffic, [0, 100]    |
| net_out          | double     |       | normarlized out going network traffic, [0, 100]    |
| disk_io_percent  | double     |       | [0, 100], abnormal values are of -1 or 101         |
+--------------------------------------------------------------------------------------------+

Three sampled datasets are found: average value of each column grouped every 10 seconds as original, and downsampled to 30 seconds and 300 seconds as well. Every column includes the average utilization of the whole data center.

Figures

Some figures were generated from these datasets

Figure: CPU utilization sampled every 10 seconds

Figure: Memory utilization sampled every 300 seconds

Figure: Net in sampled every 300 seconds

Figure: Net out sampled every 300 seconds

Figure: Disk io sampled every 300 seconds

Google 2019 instance usage

Processed from the original dataset and queried using Big Query. More information available at: https://research.google/tools/datasets/google-cluster-workload-traces-2019/

This repository dataset of instance usage includes the following columns:

+--------------------------------------------------------------------------------------------+
| Field                         | Type       | Label | Comment                               |
+--------------------------------------------------------------------------------------------+
| cpu_util_percent              | double     |       | [0, 100]                              |
| mem_util_percent              | double     |       | [0, 100]                              |
| assigned_mem_percent          | double     |       | [0, 100]                              |
| avg_cycles_per_instruction    | double     |       | [0, _]                                |
+--------------------------------------------------------------------------------------------+

One sampled dataset is found: average value of each column grouped every 300 seconds as original. Every column includes the average utilization of the whole data center.

Figures

Some figures were generated from these datasets

Figure: CPU usage day 26 sampled every 300 seconds

Figure: Mem usage day 26 sampled every 300 seconds

Figure: Assigned mem day 26 sampled every 300 seconds

Figure: Cycles per instruction day 26 sampled every 300 seconds

Azure v2 virtual machine workload

Processed from the original dataset. More information available at: https://github.com/Azure/AzurePublicDataset/blob/master/AzurePublicDatasetV2.md

This repository dataset of instance usage includes the following columns:

+--------------------------------------------------------------------------------------------+
| Field                         | Type       | Label | Comment                               |
+--------------------------------------------------------------------------------------------+
| cpu_usage                     | double     |       | [0, _]                                |
| assigned_mem                  | double     |       | [0, _]                                |
+--------------------------------------------------------------------------------------------+

One sampled dataset is found: sum value of each column grouped every 300 seconds as original. Every column includes the total consumption of the whole data center virtual machines.

Figures

Some figures were generated from these datasets

Figure: CPU total usage by virtual machines sampled every 300 seconds.

Figure: Total assigned memory for virtual machines sampled every 300 seconds.

Installation

pip install datacentertracesdatasets

Usage

To load the original Alibaba's 2018 machine usage, with the mean usage of all machines for each timestamp (8 days, 10 seconds timestep) as a Pandas DataFrame:

from datacentertracesdatasets import loadtraces
alibaba_2018_original_machine_usage_df = loadtraces.get_trace(trace_name='alibaba2018', trace_type='machine_usage', stride_seconds=10)

If, instead of a Pandas DataFrame, a numpy NDArray is needed, the format parameter can be used:

azure_v2_machine_usage_ndarray = loadtraces.get_trace(trace_name='azure_v2', trace_type='machine_usage', stride_seconds=300, format='ndarray')

Or, for Google 2019 machine usage:

azure_v2_machine_usage_ndarray = loadtraces.get_trace(trace_name='google2019', trace_type='machine_usage', stride_seconds=300, format='ndarray')

In addition to the original Alibaba 2018 machine usage dataset, which has a 10-seconds timestep, two additional downsampled versions of 30 and 300 seconds timesteps are provided, which can be retrieved by using the stride_seconds argument:

alibaba_2018_machine_usage_300_timestep_df = loadtraces.get_trace(trace_name='alibaba2018', trace_type='machine_usage', stride_seconds=300, format='ndarray')

Dataset metadata

The dataset structure and metadata can be retrieved with the get_dataset_info method:

dataset_info = get_dataset_info(trace_name='alibaba2018', trace_type='machine_usage', stride_seconds=300)

Which returns:

dataset_info = {
                "timestamp_frequency_secs": 300,
                "column_config": {
                    "cpu_util_percent": {
                        "column_index": 0,
                        "y_axis_min": 0,
                        "y_axis_max": 100
                    },
                    "mem_util_percent": {
                        "column_index": 1,
                        "y_axis_min": 0,
                        "y_axis_max": 100
                    },
                    "net_in": {
                        "column_index": 2,
                        "y_axis_min": 0,
                        "y_axis_max": 100
                    },
                    "net_out": {
                        "column_index": 3,
                        "y_axis_min": 0,
                        "y_axis_max": 100
                    },
                    "disk_io_percent": {
                        "column_index": 4,
                        "y_axis_min": 0,
                        "y_axis_max": 100
                    }

                },
                "metadata": {
                    "fields": {
                        "cpu_util_percent": {
                            "type": "numerical",
                            "subtype": "float"
                        },
                        "mem_util_percent": {
                            "type": "numerical",
                            "subtype": "float"
                        },
                        "net_in": {
                            "type": "numerical",
                            "subtype": "float"
                        },
                        "net_out": {
                            "type": "numerical",
                            "subtype": "float"
                        },
                        "disk_io_percent": {
                            "type": "numerical",
                            "subtype": "float"
                        }
                    }
                }
            }

Currently only Alibaba 2018's, Google 2019's and Azure_V2 (virtual) machine usage traces are available. In the future, we plan to add the following:

• Alibaba's 2018 batch_task workload trace.
• Google's 2019 batch_task workload trace.