Network working group call notes: Sep. 23, 2024 #4178
Conversation
Co-authored-by: Tyler Potyondy <77175673+tyler-potyondy@users.noreply.github.com>
| * Branden: Could have Docker images just to compile, and leave loading to local host. But maybe loading is actually the hard part... | ||
| * Leon: Docker setup isn't exactly easy either. Could have an RDP / remote environment to some network server to build code, then download from there and flash it. | ||
| * Tyler: That feels like quite a bit of work | ||
| * Leon: I saw someone do this with a thing called "Guacamole", which was basically a VNC viewer in the browser to an AWS VM. |
There was a problem hiding this comment.
fwiw, this is very mature technology; UCSD licenses it at scale and many of our classes use it
however, accessing local resources is pretty cumbersome (e.g. for file transfer, a weird javascript window overlay thing); I'm not sure USB passthrough is an option/stable/etc, so unlikely to solve any of the local programming / etc needs
| * Leon: One thing that's missing right now: don't have a way to guarantee that when you launch a jobset of multiple jobs on multiple boards, you don't have an assurance that they'll run in parallel. So if we have three RPis, each of the three can have jobs, but the current form doesn't guarantee that they _will_ run in parallel. | ||
| * Leon: The way to solve this now is that you can reserve all three of these, and only launch jobs on the reservation itself. It's possible that another job sneaks in though. | ||
| * Leon: So, an integration test across distributed boards on distributed hosts is tough right now | ||
| * Tyler: So reservations: we could have three RPis here at UCSD which are only for networks? |
There was a problem hiding this comment.
this might make sense if this is viewed not as 'three RPis' being reserved, but as 'the networking testbed' being reserved — the runner for which happens to control N RPis. That probably generalizes better, and keeps the synchronization out of scope for treadmill, which has enough going on...
that aligns with how larger academic testbeds operate
[n.b., if we're feeling fancy, could probably actually use existing large-scale 15.4 testbeds (e.g. FlockLab, Indria, Twist, whatever the TU Graz one is called, etc) as runners ... gets you 100+ nodes managed by others ... lots of glue logic though]
Pull Request Overview
This pull request adds notes from the Network working group call on September 23, 2024.
Rendered
Testing Strategy
Documentation
TODO or Help Wanted
Good to go
Documentation Updated
/docs, or no updates are required.Formatting
make prepush.