Dask dashboard over ssh

[rdPatmore]

I use Dask with Xarray and I am trying to profile my code through the Dask dashboard. My aim is to tunnel the dashboard over ssh from a remote virtual machine to my local (laptop) browser. This doesn't seem to work for me. Here's what I do.

First of all I am setting up port forwarding from the remote location to my laptop via

ssh -N -L 8000:localhost:8787 <username>@<remote hostname>

Then as a test in a separate terminal, I ssh to the remote host in the normal way

ssh -X <username>@<remote hostname>

and run the following in IPython from that terminal

from dask.distributed import LocalCluster
cluster = LocalCluster(n_workers=1)

then open a browser on my laptop with the address http://localhost:8000/ in an attempt to open the Dask dashboard. This fails and returns the following in my port forwarding terminal

​channel 2: open failed: connect failed: Connection refused

I have made small progress here by defining the scheduler_port variable. I changed my IPython code to

from dask.distributed import LocalCluster
cluster = LocalCluster(n_workers=1, scheduler_port=8787)

​​Now when I try the address http://localhost:8000/ I no longer get connection refused error above. However, the browser just hangs and does nothing as if it is waiting for data.

The way I see it there are three potential sources of my problem:-

1. I have made a trivial error
2. This is a specific problem with the remote machine that I am trying the tunnel the dashboard from
3. There is a bug with Dask

I presume it is 1) or 2) but I thought I'd post here in case it is 3).

[gforsyth]

@funkMasterRyan -- what's the value of cluster.dashboard_url on the remote host?

if it's an empty string, you might not have bokeh installed, which would stop the dashboard from launching

[rdPatmore]

cluster.dashboard_url

returns

'LocalCluster' object has no attribute 'dashboard_url'

Did I misinterpret what you mean here?

I can import bokeh just fine on the remote host, if that helps. It seems version 0.12.4. is installed.

[gforsyth]

Ugh, sorry, cluster.daskboard_link

[rdPatmore]

Ohh yeah, you are right. cluster.daskboard_link returns an empty string.

Though I can import bokeh. Is Dask not linking up with bokeh properly?

[gforsyth]

Could be you have two separate environments -- you might want to install using conda to avoid any missing dependencies.

In terms of specifying ports, if you set scheduler_port to 8787, that's the default dashboard port so that might be messing with things. You could try instead to run

LocalCluster(n_workers=1, dashboard_address="localhost:8787")

[rdPatmore]

I have got to the bottom of my problem, lack of conda was the issue.

Whilst I did have conda and did use it to install some packages, the word 'some' here was root of my problem. I had installed some packages using conda and some with pip. After reinstalling everything with conda, the dask dashboard worked just fine.

Thanks for your comments @gforsyth