Job
Categories:
Job View. This example shows a completed, shared job with lacking ‘flops_any’ performance.
The job view displays all data related to one specific job in full detail, and allows detailed inspection of all metrics at several scopes, as well as manual tagging of the job.
Top Bar
The top bar of each job view replicates the “Job Info” and “Footprint” seen in the job list, and additionally renders general metric information in specialized plots.
For shared jobs, a list of jobs which run (or ran) concurrently is shown as well.
Job Info
Identical to the job list equivalent, this component displays meta data containing general information about the job. From here, users can navigate to the detailed view of one specific job as well as the user or project specific job lists.
| Field | Example | Description | Destination |
|---|---|---|---|
| Job Id | 123456 | The JobId of the job assigned by the scheduling daemon | Job View |
| Job Name | myJobName | The name of the job as supplied by the user | - |
| Username | abcd10 | The username of the submitting user | User Jobs |
| Project | abcd | The name of the usergroup the submitting user belongs to | Joblist with preset Filter |
| Resources | n100 | Indicator for the allocated resources. Single resources will be displayed by name, i.e. exclusive single-node jobs or shared resources. Multiples of resources will be indicated by icons for nodes, CPU Threads, and accelerators. | - |
| Partition | main | The cluster partition this job was startet at | - |
| Start Timestamp | 10.1.2024, 10:00:00 | The epoch timestamp the job was started at, formatted for human readability | - |
| Duration | 0:21:10 | The runtime of the job, will be updated for running jobs on reload. Additionally indicates the state of the job as colored pill | - |
| Walltime | 24:00:00 | The allocated walltime for the job as per job submission script | - |
Footprint
Identical to the job list equivalent, this component will show base metrics for job performance at a glance, and will hint to job quality and problems in regard to configurable metric thresholds. In contrast to the job list, it is always active and shown in the detailed job view.
| Field | Description | Note |
|---|---|---|
| cpu_load | Average CPU utilization | - |
| flops_any | Floprate calculated as f_any = (f_double x 2) + f_single | - |
| mem_bw | Average memory bandwidth used | - |
| mem_used | Maximum memory used | Non-GPU Cluster only |
| acc_utilization | Average accelerator utilization | GPU Cluster Only |
Colors and icons differentiate between the different warning states based on the configured thresholds of the metrics. Reported metric values below the warning threshold simply report bad performance in one or more metrics, and should therefore be inspected by the user for future performance improvement.
Metric values colored in blue, however, usually report performance above the expected levels - Which is exactly why these metrics should be inspected as well. The “maximum” thresholds are often the theoretically achievable performance by the respective hardware component, but rarely are they actually reached. Inspecting jobs reporting back such levels can lead to averaging errors, unrealistic spikes in the metric data or even bugs in the code of ClusterCockpit.
| Color | Level | Description | Note |
|---|---|---|---|
| Blue | Info | Metric value below maximum configured peak threshold | Job performance above expected parameters - Inspection recommended |
| Green | OK | Metric value below normal configured threshold | Job performance within expected parameters |
| Yellow | Caution | Metric value below configured caution threshold | Job performance might be impacted |
| Red | Warning | Metric value below configured warning threshold | Job performance impacted with high probability - Inspection recommended |
| Dark Grey | Error | Metric value extremely above maximum configured threshold | Inspection required - Metric spikes in affected metrics can lead to errorneous average values |
Examples
Footprint of a job with performance well within expected parameters, ‘mem_bw’ even overperforms.
Footprint of an accelerated job with mixed performance parameters.
Footprint of a job with performance averages way above the expected maxima - Look for artifacts!
Concurrent Jobs
In the case of a shared job, this component will display all jobs, which were run on the same hardware at the same time. “At the same time” is defined as “has a starting or ending time which lies between the starting and ending time of the reference job” for this purpose.
A cautious period of five minutes is applied to both limits, in order to restrict display of jobs which have too little overlap, and would just clutter the resulting list of jobs.
Each overlapping job is listed with its jobId as a link leading to this jobs detailed job view.
Polar Representation
A polar plot representing the utilization of three key metrics: flops_any, mem_used, and mem_bw. Both the maximum and the average are rendered. In principle, this is a graphic representation of data also shown in the footprint component.
Roofline Representation
A roofline plot representing the utilization of available resources as the relation between computation and memory usage over time (color scale blue -> red).
Metric Plot Table
The views’ middle section consists of metric plots for each metric selected in the “Metrics” selector, which defaults to all configured metrics.
The data shown per metric defaults to the smallest available granularity of the metric with data of all nodes, but can be changed at will by using the drop down selectors above each plot.
If available, the statistical representation can be selected as well, by scope (e.g. stats series (node)).
Tagging
Manual tagging of jobs is performed by using the “Manage Tags” option.
Existing tags are listed, and can be added to the jobs’ database entry simply by pressing the respective button.
The list can be filtered for specific tags by using the “Search Tags” prompt.
New tags can be created by entering a new type:name combination in the search prompt, which will display a button for creating this new tag.
Statistics and Meta Data
Statistics Table. ‘cpu_power’ granularity is set to ‘socket’. Tabs above switch the contents to the job script or slurm information, both read from the jobs metadata field.
On the bottom of the job view, additional information about the job is collected. By default, the statistics of selected metrics are shown in tabular form, each in their metrics’ native granularity.
Statistics Table
The statistics table collects all metric statistical values (min, max, avg) for each allocated node and each granularity.
The metrics to be displayed can be selected using the “Metrics” selection pop-up window. In the header, next to the metric name, a second drop down allows the selection of the displayed granularity.
Core and Accelerator metrics default to their respective native granularities automatically.
Job Script
This tab displays the job script with which whis job was started on the systems.
Slurm Info
THis tab displays information returned drom the SLURM batch process management software.
Feedback
Was this page helpful?
Glad to hear it! Please tell us how we can improve.
Sorry to hear that. Please tell us how we can improve.