Use Zygote process #775
Use Zygote process #775
Conversation
jprendes
force-pushed
the
zygote
branch
2 times, most recently
from
61cad63 to
db471b3
Compare
jprendes
force-pushed
the
zygote
branch
2 times, most recently
from
a495cb8 to
eff8a9d
Compare
jprendes
force-pushed
the
zygote
branch
3 times, most recently
from
75c928e to
1654ca5
Compare
|
This build on the changes from #774 The benchmark results show no change in performance w.r.t. the scheduled nightly runs: |
Signed-off-by: Jorge Prendes <[email protected]>
Signed-off-by: Jorge Prendes <[email protected]>
Signed-off-by: Jorge Prendes <[email protected]>
| } | ||
|
|
||
| let container = builder.as_init(&bundle).with_systemd(false).build()?; | ||
| let (root, state) = Zygote::global() |
There was a problem hiding this comment.
I would like to see the memory usage of this zygote process vs. the cloned shim process running containers
There was a problem hiding this comment.
Memory usage as reported by VmPeak in /proc/xxx/status
Without zygote
Shim peak memory usage was: 3341772 kB
With zygote:
Shim peak memory usage was: 3342252 kB
Zygote peak memory usage was: 31056 kB
I am a bit surprised by the memory usage of the shim, I expected it to consume less memory.
There was a problem hiding this comment.
Revisiting the numbers, for wasmtime the shim peaks at ~19000 kB of resident memory with or without the zygote.
The zygote itself peaks at 5500kB of resident memory.
Signed-off-by: Jorge Prendes <[email protected]>
There was a problem hiding this comment.
This is really great work! Looking forward to seeing less sporadic CI failures 🚀
Do I understand correctly that the container creation would look like:
->shim
->zygote (cloned from shim)
-> container 1 (message sent from shim to zygote; cloned from zygote)
-> container 2 (message sent from shim to zygote; cloned from zygote)
|
Yes, that's correct. We could still see issues if we ever "pollute" the zygote. We do create an engine inside the zygote before cloning the process, and if that has global state, starts threads, etc, it can be a problem. Currently the engine struct for all our shims are very simple, so that's not an issue. I'm thinking on addressing that as part of the split engine-precompiler. |
This PR introduces a zygote process (a very small process cloned early on during shim startup), from which new containers are created. This effectively decouples the global state of the shim from the state of the container process.
Fixes #755 and #357