https://rancher.com/ logo
Title
f

fancy-monitor-61453

09/22/2022, 3:30 PM
@bland-account-99790 and @clever-analyst-23771 we dont use flannel but are working on a much faster fabric for gpus from cloud to iot. ML use cases are very diff. we are working towards something like the AWS elastic fabric adapter.
👍 1
s

strong-musician-35046

09/22/2022, 3:32 PM
This sounds interesting. 😉
b

bland-account-99790

09/22/2022, 3:33 PM
Cool!!!! Really interesting indeed
c

clever-air-65544

09/22/2022, 3:34 PM
absolutely, would love to hear about that!
f

fancy-monitor-61453

09/22/2022, 3:35 PM
its for ml pipelines
s

strong-musician-35046

09/22/2022, 3:36 PM
it would (likely) be useful to show some info on performance, latency, error handling, packetsize, etc to help understand the motivating factors.... every decision comes with its own side-order of thorns and tiger traps yknow
for some reason the 'smoke signals' and 'carrier pigeon' backends seem to be especially lossy 🙂
f

fancy-monitor-61453

09/22/2022, 3:39 PM
im not sure what u mean, i was just stating my use case for the ask
s

strong-musician-35046

09/22/2022, 3:44 PM
Well, I was wondering why a faster fabric was necessary for GPUs... and along what axis you were measuring to assess 'faster' and what struggles you're / you've run into.... clearly it's important to ya, or you wouldn't be building it 😉 and I was just curious as to what those motivations were ... because there's always a trade off
was just idle curiosity, really.
c

cool-forest-29147

09/22/2022, 3:50 PM
Uh, subscribe 🙂 We're also using GPUs heavily, would be really interested to see what you're up to.
👍 1
f

fancy-monitor-61453

09/22/2022, 6:40 PM
i need 10,000 GPUs
...im not kidding
graph neural network problem
ability to go from 5 to 1000 in a minute based on compute need of graph slice; up to 10K.
then the elastic fabric scales back down once compute is finished for that sub-graph
c

clever-analyst-23771

09/22/2022, 8:29 PM
Thats a hell of a use case dang.
f

fancy-monitor-61453

09/22/2022, 8:33 PM
yea, GNNs are rather beefy