Unreachable hosts

[pierlauro]

Hi,
when using singularity-compose, the local hostname is overridden and to each containers is assigned a different local ip address. E.g.: on a machine with hostname localhost and address 10.1.2.345, singularity runs a container xyz with local ip 10.0.0.1 and hostname xyz .

When executing applications using sockets in containers, however, I noticed it is impossible to reach the endpoints both from outside (from other containers or local machine) and inside (from the same container). Sockets are opened, but they are unreachable.

I also tried running singularity-compose with the --no-resolv option, but the issue persists. Any clue why this is happening?

When running containers without singularity-compose, hostname and ip are not overridden and everything works fine.

Thank you

[vsoch]

This is unfortunately likely an issue with Singularity - there was a lot of trouble with this first issue when I was developing this library. Could you please provide a dummy case to reproduce your issue? I can try it and if it works for me you’ll need to take this up with Singularity developers.

[pierlauro]

I could isolate the problem while using some complex containers, give me some time to write down a more simple example.

[vsoch]

Ah found it! This issue has the commands that were useful to debugging core instances: sylabs/singularity#3766

[pierlauro]

Ah found it! This issue has the commands that were useful to debugging core instances: sylabs/singularity#3766

Having the same issue! If I start one alpine container with --network-args, then its ip is unreachable (even from within the container itself). If I start the container without network options, I am able to ping it from outside. [Obviously I deactivated ICMP on my local machine to test the container one].

Given the issue, would it be possible to configure singularity-compose in order to don't use singularity's network options?

[pierlauro]

By enabling the debug flag (singularity-compose -d up), I can see that the command executed to start a container named xyz is:

singularity instance start --bind ./etc.hosts:/etc/hosts --net --network none --network-args "IP=10.22.0.2" --hostname xyz --writable-tmpfs PATH_TO_SIF xyz

Trying manually directly with singularity, if the command is executed without --hostname xyz everything is working because the network is not virtualized. It definitely seems that --network none is ignored if a hostname is speficied.

[pierlauro]

Solved: #27

[vsoch]

Fixed by #27, thanks again @pierlauro !