Change snapshot file format to JSON?

[ipspace]

@Muddyblack is complaining that "netlab connect" takes several seconds. Unless he has a DNS problem (or some such), the delay could be caused by YAML parsing of the netlab.snapshot.yml file.

Would anyone have a problem if we start using JSON encoding for the snapshot file? We could even create both files for a while just in case someone uses the YAML file.

/cc: @ssasso @jbemmel @DanPartelly @ddutt

[ssasso]

Do we really need a "human-readable" file as snapshot? I think saving the Box structure into a pickle file should save more time.
And, as you said, we could still generate the yaml version.

[ipspace]

Good one. Thank you! We could then try to load the pickle file, falling back to YAML if opening the jar fails.

[DanPartelly]

No problem at all but I agree with @ssasso above.

And second, did anyone actually timed the function parsing the yaml snapshot ?

[ipspace]

And second, did anyone actually timed the function parsing the yaml snapshot ?

While I know YAML parsing is slow (I know someone where parsing the Ansible inventory took minutes), I never measured it.

@Muddyblack has a topology with 600 devices; you can imagine how large the YAML file is.

Alternatively, if unpickling is still too slow (I doubt it; if one can run 600 containers on a server, one probably has enough CPU power), we could parse only the relevant bits of Ansible inventory in "netlab connect"

[Muddyblack]

I am at 1000 devices now 😅 it works but for my machine its the limit with ospf startup flooding (once neighborhood is established it has no problems unless you deactivate a lot of devices instantly).
With other protocols i could imagine with frr you could even go higher in the number.
With OSPF deactivated I was able to go up to 3000 devices with no problems I could probably even do much more ...

[sdargoeuves]

@Muddyblack that's impressive!! What is your use case, what devices, and also what are the specs of the server running this?
I'm curious, at the moment our topology is around 60 devices with a memory usage ~100GB, which means we've got another 100GB for more devices.

[Muddyblack]

@sdargoeuves I just use them to test real devices behavior as well ospf behavior on small changes.
So with 3k frr devices the RAM usage is at ~55GB constantly and I have 200 more free to use. With ~70 cores (I don't even know what cpu is used sry can't tell you that 😅) once neighborhood is established its about 10% CPU usage on the host machine. But the startup neighborhood establishing uses all and needs exponential more the more devices you add ( at least that is my impression). Probably all fighting to get some cpu time.

[ipspace]

FWIW, here's the timing data for a really simple lab topology with 300 nodes from a server with AMD Ryzen 9 8945HS:

$ time netlab inspect 'len(nodes)'
308

real	0m5.540s
user	0m5.449s
sys	0m0.091s

The snapshot YAML file has 1.8MB

[Muddyblack]

With 1800 devices the snapshot has 10.32MB in size.
The netlab create command with my not yet finished PR took 13 Minutes.

[ipspace]

And here's the timing data when using the pickle file:

$ time netlab inspect 'len(nodes)'
308

real	0m0.415s
user	0m0.378s
sys	0m0.037s