Connecting notebook to remote kernel

Question

I'm trying to connect my FrontEnd session to remote kernel. Manual connection works by creating LinkObject like so:

link = LinkCreate["[email protected],[email protected]", LinkMode -> Listen, LinkProtocol -> "TCPIP"]

Connect to remote machine forwarding ports 8000 and 8001:

ssh -R8000:127.0.0.1:8000 -R8001:127.0.0.1:8001 user@host

Run:

wolfram -wstp -LinkMode Connect -LinkProtocol TCPIP -LinkName [email protected],[email protected]

Then locally this works as expected:

LinkRead[link]
(*InputNamePacket["In[1]:= "]*)

LinkWrite[link, Unevaluated@EvaluatePacket[$MachineName]]

LinkRead[link]
(*ReturnPacket["remote_machine_name"]*)

But any attempt to connect my FrontEnd results in a frozen evaluating cell that doesn't return anything. Even though the connection is clearly there, because it drops remotely after closing it locally.

I've tried many things: Evaluation -> Kernel Configuration Options, the tunnel.sh script suggested in other answers, assigning to $ParentLink, nothing works!

Remote machine has linux with Wolfram Engine For Developers installed. And I'm running WolframDesktop 12.0 locally on Mac.

EDIT: I've managed to get it working with Raw MathLink connection option ("RawMathLinkProgram" -> True). But it's pretty limited, no graphics, just text output.

EDIT: Suprisingly it works if I connect to my other laptop within my home network. It also has Ubuntu and docker with Wolfram Engine. But if I try basically the same steps with a cloud server, it refuses to work (unless it's raw mathlink connection).

Reverse connection:

ssh -L8000:127.0.0.1:8000 -L8001:127.0.0.1:8001 user@host

Replace Connect with Listen

wolfram -wstp -LinkMode Listen -LinkProtocol TCPIP -LinkName [email protected],[email protected]

And local Kernel Configuration:

-LinkMode Connect -LinkProtocol TCPIP -LinkName [email protected],[email protected]

Same result, local FrontEnd just crashes if I connect to remote cloud machine, but works with my laptop (shrug)

Very certain that it's a bug at this point, some low level protocol discrepancies maybe.

Answer 1

Sorry this is not an answer but the text goes too long for the comment.

There'a whole bunch of things that could go wrong with remote kernels in my experience.

Basically FE starts remote kernel and asks it to connect to the FE, passing FE addresses and ports in the link parameters. In my case the remote command line looks like:

WolframKernel -wstp -LinkMode Connect -LinkProtocol TCPIP -LinkName [email protected],61518
@10.0.0.3

The ports (61517 & 61518) are randomly chosen by the FE so it's hard to really use reverse forwarding via ssh (tunnel.sh tries to do this by parsing command line options. I was unable to make it work though).

Please note the reverse address (10.0.0.3 in my case) that the kernel should be able to connect to. It's not always the case so check if it is accessible from the machine where the remote kernel runs.

Even if you have not an ssh link but the full blown VPN connection to the remote system there can be another interesting problem:

The reverse address is chosen by FE using some heuristic and is based on the address on a remote machine:

FE`Evaluate@FEPrivate`FindBestInterfaceForDestination[hostname_or_ipaddr_as_string]

But please note the following:

Even if you're using "Advanced" group of dialog options the magic that choses reverse address uses "Remote host" field to select the base interfaces to use for kernel reverse connection.

And last but not least: 12.0 had a "feature" that changes to "Kernel Configuration" dialog were not always saved even if you reopen the dialog and see intended configuration. So you may want to double check if the settings are in your Kernel/init.m file.

Hope some of the above will help.