Dashboard Diagnostics

Profiling parallel code can be challenging, but the interactive dashboard provided with Dask’s distributed scheduler makes this easier with live monitoring of your Dask computations. The dashboard is built with Bokeh and will start up automatically, returning a link to the dashboard whenever the scheduler is created.

Locally, this is when you create a Client and connect the scheduler:

from dask.distributed import Client
client = Client()  # start distributed scheduler locally.

In a Jupyter Notebook or JupyterLab session displaying the client object will show the dashboard address:

Client html repr displaying the dashboard link in a JupyterLab session.

You can also query the address from client.dashboard_link (or for older versions of distributed, client.scheduler_info()['services']).

By default, when starting a scheduler on your local machine the dashboard will be served at http://localhost:8787/status. You can type this address into your browser to access the dashboard, but may be directed elsewhere if port 8787 is taken. You can also configure the address using the dashboard_address parameter (see LocalCluster).

There are numerous diagnostic plots available. In this guide you’ll learn about some of the most commonly used plots shown on the entry point for the dashboard:

Main dashboard with five panes arranged into two columns. In the left column there are three bar charts. The top two show total bytes stored and bytes per worker. The bottom has three tabs to toggle between task processing, CPU utilization, and occupancy. In the right column, there are two bar charts with corresponding colors showing task activity over time, referred to as task stream and progress.

Bytes Stored and Bytes per Worker

These two plots show a summary of the overall memory usage on the cluster (Bytes Stored), as well as the individual usage on each worker (Bytes per Worker). The colors on these plots indicate the following.

Memory under target (default 60% of memory available)
Memory is close to the spilling to disk target (default 70% of memory available)
Memory spilled to disk
Two bar charts on memory usage. The top chart shows the total cluster memory in a single bar with mostly under target memory in blue and a small part of spilled to disk in grey. The bottom chart displays the memory usage per worker, with a separate bar for each of the 16 workers. The first four bars are orange as their worker's memory are close to the spilling to disk target, with the first worker standing out with a portion in grey that correspond to the amount spilled to disk. The remaining workers are all under target showing blue bars.

The different levels of transparency on these plot is related to the type of memory (Managed, Unmanaged and Unmanaged recent), and you can find a detailed explanation of them in the Worker Memory management documentation

Task Processing/CPU Utilization/Occupancy

Task Processing

The Processing tab in the figure shows the number of tasks being processed by each worker with the blue bar. The scheduler will try to ensure that the workers are processing the same number of tasks. If one of the bars is completely white it means that worker has no tasks and its waiting for them. This usually happens when the computations are close to finished (nothing to worry about), but it can also mean that the distribution of the task across workers is not optimized.

There are three different colors that can appear in this plot:

Processing tasks.
Saturated: It has enough work to stay busy.
Idle: Does not have enough work to stay busy.
Task Processing bar chart, showing a relatively even number of tasks on each worker.

In this plot on the dashboard we have two extra tabs with the following information:

CPU Utilization

The CPU tab shows the cpu usage per-worker as reported by psutils metrics.

Occupancy

The Occupancy tab shows the occupancy, in time, per worker. The total occupancy for a worker is the total expected runtime for all tasks currently on a worker. For example, an occupancy of 10s means an occupancy of 10s means that the worker estimates it will take 10s to execute all the tasks it has currently been assigned.

Task Stream

The task stream is a view of all the tasks across worker-threads. Each row represents a thread and each rectangle represents an individual tasks. The color for each rectangle corresponds to the task-prefix of the task being performed and it matches the color of the Progress plot (see Progress section). This means that all the individual tasks part of the inc task-prefix for example, will have the same randomly assigned color from the viridis color map.

There are certain colors that are reserved for a specific kinds of tasks:

Transferring data between workers tasks.
Reading from or writing to disk.
Serializing/deserializing data.
Erred tasks.

In some scenarios the dashboard will have white spaces between each rectangle, this means that during that time the worker-thread is idle. Having too much white and red is an indication of not optimal use of resources.

An example of a healthy Task Stream, with little to no red or white space. The stacked bar chart, with one bar per worker-thread, has different shades of blue and green for different tasks, with some red.
An example of an unhealthy Task Stream, the bar chart shows a lot of white space and red rectangles, and also some orange.

Progress

The progress bars plot shows the progress of each individual task-prefix. The color of the of each bar matches the color of the individual tasks on the task stream that correspond to the same task-prefix. Each horizontal bar has three different components:

Tasks that are ready to run.
Tasks that have been completed and are in memory.
Tasks that have been completed, been in memory and have been released.
Progress bar chart with one bar for each task-prefix matching with the names "add", "double", "inc", and "sum". The "double", "inc" and "add" bars have a progress of approximately one third of the total tasks, displayed in their individual color with different transparency levels. The "double" and "inc" bars have a grey background, and the "sum" bar is empty.

Dask JupyterLab Extension

The JupyterLab Dask extension allows you to embed Dask’s dashboard plots directly into JupyterLab panes.

Once the JupyterLab Dask extension is installed you can choose any of the individual plots available and integrated as a pane in your JupyterLab session. For example, in the figure below we selected the Task Stream, Progress, Workers Memory, and Graph plots.

Dask JupyterLab extension showing an arrangement of four panes selected from a display of plot options. The panes displayed are the Task stream, Bytes per worker, Progress and the Task Graph.