This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

Views

View-Specific Frontend Usage Information.

Usage descriptions for each view of the ClusterCockpit web interface.

1 - My Jobs

All Jobs as Table of the Active User
User Job View

Personal User Job View. Similar to the general job list view, this view expands it by user-specific meta data, as well as distributions histograms.

The “My Jobs” View is available to all users regardless of authority and displays the users personal jobs, i.e. jobs started by this users username on the cluster systems.

The view is a personal variant of the user job view and therefore also consists of three components: Basic Information about the users jobs, selectable statistic histograms of the jobs, and a generalized job list.

Users are able to change the sorting, select and reorder the rendered metrics, filter, and activate a periodic reload of the data.

User Information and Basic Distributions

The top row always displays personal usage information, independent of the selected filters.

Additional histograms depicting the distribution of job duration and number of nodes occupied by the returned jobs are affected by the selected filters.

Information displayed:

  • Username
  • Total Jobs
  • Short Jobs (as defined by the configuration, default: less than 300 second runtime)
  • Total Walltime
  • Total Core Hours

Selectable Histograms

Histograms depicting the distribution of the selected jobs’ statistics can be selected from the top navbar “Select Histograms” button. The displayed data is based on the jobs returned from active filters, and will be pulled from the database, or in case of running jobs, calculated from the available metric data directly.

Available Metrics for Histograms: cpu_load, flops_any, mem_used, mem_bw, net_bw, file_bw

Job List

The job list displays all jobs started by your username on the systems. Additional filters will always respect this limitation. For a detailed description of the job list component, see the related documentation.

2 - User Jobs

All Jobs as Table of a Selected User
User Job View

User Job View. Similar to the general job list view, this view expands it by user-specific meta data, as well as distribution histograms.

The “User Jobs” View is only available to management and supporting staff and displays jobs of the selected user, i.e. jobs started by this users username on the cluster systems.

The view consists of three components: Basic Information about the users jobs, selectable statistic histograms of the jobs, and a generalized job list.

Users are able to change the sorting, select and reorder the rendered metrics, filter, and activate a periodic reload of the data.

User Information and Basic Distributions

The top row always displays information about the user, independent of the selected filters.

Additional histograms depicting the distribution of job duration and number of nodes occupied by the returned jobs are affected by the selected filters.

Information displayed:

  • Username
  • Total Jobs
  • Short Jobs (as defined by the configuration, default: less than 300 second runtime)
  • Total Walltime
  • Total Core Hours

Selectable Histograms

Histograms depicting the distribution of the selected jobs’ statistics can be selected from the top navbar “Select Histograms” button. The displayed data is based on the jobs returned from active filters, and will be pulled from the database, or in case of running jobs, calculated from the available metric data directly.

Available Metrics for Histograms: cpu_load, flops_any, mem_used, mem_bw, net_bw, file_bw

Job List

The job list displays all jobs started by this users username on the systems. Additional filters will always respect this limitation. For a detailed description of the job list component, see the related documentation.

3 - Job List

A Configurable Table Displaying Jobs According to Filters
Job View

Job List. In this example, the optional footprint is displayed, two filters are active, and the table is refreshed every minute. The first job has a high node count, therefore the plots are rendered in the statistics variant. The ‘mem_bw’ metric likely has artifacts as shown by the grey footprint. The second job has tags and displays less than optimal performance in the ‘flops_any’ metric, coloring the respective plot background in orange.

The primary view of ClusterCockpits webinterface is the tabular listing of jobs, which displays various information about the jobs returned by the selected filters. This information includes the jobs’ full meta data, such as runtime or job state, as well as an optional footprint, allowing quick assessment of the jobs performance.

Most importantly, the list displays a selectable array of metrics as time dependent metric plots, which allows detailed insight into the jobs performance at a glance.

Job List Toolbar

Job View

Several options allow configuration of the displayed data, which are also persisted for each user individually, either for general usage or by cluster.

Sorting

Basic selection of sorting parameter and direction. By default, jobs are sorted by starting timestamp in descending order (latest jobs first). Other selections to sort by are

  • Duration
  • Number of Nodes
  • Maximum Memory Used
  • Average FLOPs
  • Average Memory Bandwidth
  • Average Network Bandwidth

Switching of the sort direction is achieved by clicking on the arrow icon next to the desired sorting parameter.

Metrics

Selection of metrics shown in the tabular view for each job. The list is compiled from all available configured metrics of the ClusterCockpit instance, and the tabular view will be updated upon applying the changes.

Job View

In addition to the metric names themselves, the availability by cluster is indicated as comma seperated list next to the metric identifier. This information will change to the availablility by partition if the cluster filer is active.

It is furthermore possible to edit the order of the selected metrics. This can be achieved by dragging and dropping the metric selectors to the desired order, where the topmost metric will be displayed next to the “Job Info” column, and additional metrics will be added on the right side.

Lastly, the optional “Footprint” Column can be activated (and deactivated) here. It will always be rendered next to the “Job Info” column, while metrics start right of the “Footprint” column, if activated.

Job Count

The total number of jobs returned by the backend for the given set of filters.

Filters

Selection of filters applied to the queried jobs. By default, no filters are activated if the view was opened via the navigation bar. At multiple location throughout the web-interface, direct links will lead to this view with one or more preset filters active, e.g. selecting a clusters’ “running jobs” from the home page will open this view displaying only running jobs of that cluster.

Possible options are:

  • Cluster/Partition: Filter by configured cluster (and partitions thereof)
  • Job State: Filter by defined job state(s)
  • Start Time: Filter by start timestamp
  • Duration: Filter by job duration
  • Tags: Filter by tags assigned to jobs
  • Resources: Filter by allocated resources or named node
  • Statistics: Filter by average usage of defined metrics

Each filter and its default value is described in detail here.

Search and Reload

Search for specific username or project using the searchbox, force a complete reload of the table data, or set a timed periodic reload (30, 60, 120, 300 Seconds).

Job List Table

The main component of the job list view renders data pulled from the database, the job archive (completed jobs) and the configured metric data source (running jobs).

Job Info

The meta data containing general information about the job is represented in the “Job Info” column, which is always the first column to be rendered. From here, users can navigate to the detailed view of one specific job as well as the user or project specific job lists.

FieldExampleDescriptionDestination
Job Id123456The JobId of the job assigned by the scheduling daemonJob View
Job NamemyJobNameThe name of the job as supplied by the user-
Usernameabcd10The username of the submitting userUser Jobs
ProjectabcdThe name of the usergroup the submitting user belongs toJoblist with preset Filter
Resourcesn100Indicator 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.-
PartitionmainThe cluster partition this job was startet at-
Start Timestamp10.1.2024, 10:00:00The epoch timestamp the job was started at, formatted for human readability-
Duration0:21:10The runtime of the job, will be updated for running jobs on reload. Additionally indicates the state of the job as colored pill-
Walltime24:00:00The allocated walltime for the job as per job submission script-

Footprint

The optional footprint column will show base metrics for job performance at a glance, and will hint to performance (and performance problems) in regard to configurable metric thresholds.

FieldDescriptionNote
cpu_loadAverage CPU utilization-
flops_anyFloprate calculated as f_any = (f_double x 2) + f_single-
mem_bwAverage memory bandwidth usedNon-GPU Cluster only
mem_usedMaximum memory usedNon-GPU Cluster only
acc_utilizationAverage accelerator utilizationGPU Cluster Only

Colors and icons differentiate between the different warning states based on the configured threshold 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.

ColorLevelDescriptionNote
BlueInfoMetric value below maximum configured peak thresholdJob performance above expected parameters - Inspection recommended
GreenOKMetric value below normal configured thresholdJob performance within expected parameters
YellowCautionMetric value below configured caution thresholdJob performance might be impacted
RedWarningMetric value below configured warning thresholdJob performance impacted with high probability - Inscpection recommended
Dark GreyErrorMetric value extremely above maximum configured thresholdInspection required - Metric spikes in affected metrics can lead to errorneous average values

Metric Row

Selected metrics are rendered here in the selected order as metric lineplots. Aspects of the rendering can be configured at the settings page.

4 - Job

Detailed Single Job Information View
Job View

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.

FieldExampleDescriptionDestination
Job Id123456The JobId of the job assigned by the scheduling daemonJob View
Job NamemyJobNameThe name of the job as supplied by the user-
Usernameabcd10The username of the submitting userUser Jobs
ProjectabcdThe name of the usergroup the submitting user belongs toJoblist with preset Filter
Resourcesn100Indicator 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.-
PartitionmainThe cluster partition this job was startet at-
Start Timestamp10.1.2024, 10:00:00The epoch timestamp the job was started at, formatted for human readability-
Duration0:21:10The runtime of the job, will be updated for running jobs on reload. Additionally indicates the state of the job as colored pill-
Walltime24:00:00The 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.

FieldDescriptionNote
cpu_loadAverage CPU utilization-
flops_anyFloprate calculated as f_any = (f_double x 2) + f_single-
mem_bwAverage memory bandwidth used-
mem_usedMaximum memory usedNon-GPU Cluster only
acc_utilizationAverage accelerator utilizationGPU 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.

ColorLevelDescriptionNote
BlueInfoMetric value below maximum configured peak thresholdJob performance above expected parameters - Inspection recommended
GreenOKMetric value below normal configured thresholdJob performance within expected parameters
YellowCautionMetric value below configured caution thresholdJob performance might be impacted
RedWarningMetric value below configured warning thresholdJob performance impacted with high probability - Inspection recommended
Dark GreyErrorMetric value extremely above maximum configured thresholdInspection required - Metric spikes in affected metrics can lead to errorneous average values

Examples

Footprint with good Performance

Footprint of a job with performance well within expected parameters, ‘mem_bw’ even overperforms.

Footprint with mixed Performance

Footprint of an accelerated job with mixed performance parameters.

Footprint with Errors

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.

Tagging

Create Tag Window

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

Job View Statistics Table

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.

5 - Users

Table of All Users Running Jobs on the Clusters
User Table

User Table, sorted by ‘Total Jobs’ in descending order. In addition, active filters reduce the underlying data to jobs with more than one hour runtime, started on the GPU accelerated cluster.

This view lists all users which are, and were, active on the configured clusters. Information about the total number of jobs, walltimes and calculation usages are shown.

It is possible to filter the list by username using the equally named prompt, which also accepts partial queries.

The filter component allows limitation of the returned users based on job parameters like start timestamp or memory usage.

The table can be sorted by clicking the respective icon next to the column headers.

Details

ColumnDescriptionNote
User NameThe user jobs are associated withLinks to the users’ job list with preset filter returning only jobs of this user and additional histograms
Total JobsUsers’ total of all started jobs
Total WalltimeUsers’ total requested walltime
Total Core HoursUsers’ total of all used core hours
Total Accelerator HoursUsers’ total of all used accelerator hoursPlease Note: This column is always shown, and will return 0 for clusters without installed accelerators

6 - Projects

Table of All Projects Running Jobs on the Clusters
User Table

Project Table, sorted by ‘Total Jobs’ in descending order. In addition, active filters reduce the underlying data to jobs with less than six hours runtime, started on the CPU exclusive cluster.

This view lists all projects (usergroups) which are, and were, active on the configured clusters. Information about the total number of jobs, walltimes and calculation usages are shown.

It is possible to filter the list by project name using the equally named prompt, which also accepts partial queries.

The filter component allows limitation of the returned projects based on job parameters like start timestamp or memory usage.

The table can be sorted by clicking the respective icon next to the column headers.

Details

ColumnDescriptionNote
Project NameThe project (usergoup) jobs are associated withLinks to a job list with preset filter returning only jobs of this project
Total JobsProject total of all started Jobs
Total WalltimeProject total requested walltime
Total Core HoursProject total of all used core hours used
Total Accelerator HoursProject total of all used accelerator hoursPlease Note: This column is always shown, and will return 0 for clusters without installed accelerators

7 - Tags

Lists Active Tags Used in the Frontend
Tag List View

This view lists all tags currently used within the ClusterCockpit instance:

  • The type of the tag(s) is displayed as dark grey header, collecting all tags which share it.
  • The names of all tags sharing one type are rendered as yellow pills below the header.
  • How often a tag was applied to a job is shown in the number following the tags name

Each tags’ pill is clickable, and leads to a job list with a preset filter matching only jobs tagged with this specific label.

8 - Nodes

Node Based Metric Information of one Cluster
Nodes View

Nodes View. This example shows the last two hours of the ‘clock’ metric of eight nodes. Node ‘f0147’ of the ‘main’ partition has an average below the configured ‘alert’ threshold, and is colored in red.

The nodes view, or systems view, is always called in respect to one specified cluster. It displays the current state of all nodes in that cluster in respect to one selected metric, rendered in form of metric plots, and independent of job meta data, i.e. without consideration for job start and end timestamps.

Selection Bar

Nodes View

Selections regarding the display, and update, of the plots rendered in the node table can be performed here:

  • (Periodic) Reload: Force reload of fresh data from the backend or set a periodic reload in specified intervals
    • 30 Seconds, 60 Seconds, 120 Seconds, 5 Minutes
  • Displayed Time: Select the timeframe to be rendered in the node table
    • Custom: Select timestamp from and to in which the data should be fetched. It is possible to select date and time.
    • 15 Minutes, 30 Minutes, 1 Hour, 2 Hours, 4 Hours, 12 Hours, 24 Hours
  • Metric:: Select the metric to be fetched for all nodes. If no data can be fetched, messages are displayed per node.
  • Find Node:: Filter the node table by hostname. Partial queries are possible.

Node Table

Nodes (hosts) are ordered alphanumerically in this table, rendering the selected metric in the selected timeframe.

Each heading links to the singular node view of the respective host.

9 - Node

All Metrics of One Selected Node
Node View

Node View. This example shows the last twelve hours of all metrics of the specified node ‘a0122’. The metric ‘acc_mem_used’ has an average below the configured ‘alert’ threshold, and is colored in red.

The node view is always called in respect to one specified cluster and one specified node (host). It displays the current state of all metrics for that node, rendered in form of metric plots, and independent of job meta data, i.e. without consideration for job start and end timestamps.

Selection Bar

Information and selections regarding the data of the plots rendered in the node table can be performed here:

  • Name: The hostname of the inspected node
  • Concurrent Jobs: Number of jobs currently allocated to this node. Exclusively used nodes will always display 1 if a job is running at the moment, or 0 if not.
    • A link is provided which leads to the joblist with preset filter fetching only currently allocated jobs.
  • (Periodic) Reload: Force reload of fresh data from the backend or set a periodic reload in specified intervals
    • 30 Seconds, 60 Seconds, 120 Seconds, 5 Minutes
  • Displayed Time: Select the timeframe to be rendered in the node table
    • Custom: Select timestamp from and to in which the data should be fetched. It is possible to select date and time.
    • 15 Minutes, 30 Minutes, 1 Hour, 2 Hours, 4 Hours, 12 Hours, 24 Hours

Node Table

Metrics are ordered alphanumerically in this table, rendering each metric in the selected timeframe.

10 - Analysis

Metric Data Analysis View
Analysis View

Analysis View General Information Section. Two filters are active, the pie chart displays top user node hour utilization fractions.

The analysis view is always called in respect to one specified cluster. It collects and renders data based on the jobs returned by the active filters, which can be specified to a high detail, allowing analysis of specific aspects.

General Information

The general information section of the analysis view is always rendered and consists of the following elements

Totals

Total counts of collected data based on the returned jobs matching the requested filters:

  • Total Jobs
  • Total Short Jobs (By default defined as jobs shorter than 5 minutes)
  • Total Walltime
  • Total Node Hours
  • Total Core Hours
  • Total Accelerator Hours

Top Users and Projects

The ten most active users or projects are rendered in a combination of pie chart and tabular legend with values displayed. By default, the top ten users with the most jobs matching the selected filters will be shown.

Hovering over one of the pie chart fractions will display a legend featuring the identifier and value of the selected parameter.

The selection can be changed directly in the headers of the pie chart and the table, and can be changed to

ElementOptions
Pie ChartUsers, Projects
TableWalltime, Node Hours, Core Hours, Accelerator Hours

The selection is saved for each user and cluster, and will select the last chosen types of list as default the next time this view is opened.

“User Names” and “Project Codes” are rendered as links, leading to user job lists or project job lists with preset filters for cluster and entity ID.

Heatmap Roofline

A roofline plot representing the utilization of available resources as the relation between computation and memory for all jobs matching the filters. In order to represent the data in a meaningful way, the time information of the raw data is abstracted and represented as a heat map, with increasingly red sections of the roofline plot being the most populated regions of utilization.

Histograms

Two histograms depicting the duration and number of allocated cores distributions for the returned jobs matching the filters.

Selectable Data Representations

Analysis View Plot Selection

The second half of the analysis view consists of areas reserved for rendering user-selected data representations.

  • Select Plots for Histograms: Opens a selector listing all configured metrics of the respective cluster. One or more metrics can be selected, and the data returned will be rendered as average distributions normalized by node hours (core hours, accelerator hours; depending on the metric).
  • Select Plots in Scatter Plots: Opens a selector which allows selection of user chosen combinations of configured metrics for the respective cluster. Selected duplets will be rendered as scatter bubble plots for each selected pair of metrics.
Analysis View Scatter Selection

Three pairs of metrics are already selected for scatter representation. Remove a selected pair by pressing the ‘x’ button, add a new pair by selecting two metric from the dropdown menu, and confirming by pressing ‘Add Plot’.

Average Distribution Histograms

Analysis View Average Distributions

Three selected metrics are represented as normalized, average distributions based on returned jobs.

These histograms show the distribution of the normalized averages of all jobs matching the filters, split into 50 bins for high detail.

Normalization is achieved by weighting the selected metric data job averages by node hours (default), or by either accelerator hours (for native accelerator scope metrics) or core hours (for native core scope metrics).

User Defined Scatterplots

Analysis View Scatter Plots

Three user defined scatter plots.

Bubble scatter plots show the position of the averages of two selected metrics in relation to each other.

Each circle represents one job, while the size of a circle is proportional to its node hours. Darker circles mean multiple jobs have the same averages for the respective metric selection.

11 - Status

Hardware Usage Information

The status view is always called in respect to one specified cluster. It displays the current state of utilization of the respective clusters resources, as well as user and project top lists and distribution histograms of the allocated resources per job.

Utilization Information

Subluster Urilization in Status view

For each subluster, utilization is displayed in two parts rendered in one row.

Gauges

Simple gauge representation of the current utilization of available resources

FieldDescriptionNote
Allocated NodesNumber of nodes currently allocated in respect to maximum available-
Flop Rate (Any)Currently achieved flop rate in respect to theoretical maximumFloprate calculated as f_any = (f_double x 2) + f_single
MemBW RateCurrently achieved memory bandwidth in respect to technical maximum-

Roofline

A roofline plot representing the utilization of available resources as the relation between computation and memory for each currently allocated, running job at the time of the latest data retrieval. Therefore, no time information is represented (all dots in blue, representing one job each).

Top Users and Projects

Subluster Urilization in Status view

The ten most active users or projects are rendered in a combination of pie chart and tabular legend. By default, the top ten users or projects with the most allocated, running jobs are listed.

The selection can be changed directly in the tables header at Number of ..., and can be changed to

  • Jobs (Default)
  • Nodes
  • Cores
  • Accelerators

The selection is saved for each user and cluster, and will select the last chosen type of list as default the next time this view is rendered.

Hovering over one of the pie chart fractions will display a legend featuring the identifier and value of the selected parameter.

“User Names” and “Project Codes” are rendered as links, leading to user job lists or project job lists with preset filters for cluster, entity ID, and state == running.

Statistic Histograms

Several histrograms depicting the utilization of the clusters resources, based on all currently running jobs are rendered here:

  • Duration Distribution
  • Number of Nodes Distribution
  • Number of Cores Distribution
  • Number of Accelerators Distribution