Compute budgets, rate limiting, throttling, server operator performance levers

[TeamDman]

token bucket algorithms

[TeamDman]

Discussion from discord

https://discord.com/channels/967118679370264627/967119184209264790/1540141337452617820

---

Teamy — 7:10 PM
hoping the new features will make it easier to build tooling that helps newbies
there's tooling in IDEA and stuff for showing ANTLR grammar parse trees, showing the grammar in-game as a visual way to show what the intellisense is drawing from will be neat
Other — 7:12 PM
Maybe one day you will switch to wasm, and the alternative screens, so one person can make a drag and drop support addon lol
Other — 7:13 PM
Hmm I'm curious how the SFM interpreter works now
I think I'll look into it
Teamy — 7:13 PM
it's pretty neat
ANTLR is super fun for creating custom grammars, lets you parse your stuff into an AST then SFM just executes the AST fairly linearly
user-defined grammars entirely in-game would be fun
having the entire state of a network sent to a client for the client to then issue commands back to the server for how to move items would probably be preferable to having a server try and safely compile user code if we want to support languages like lua or rust without forcing them to go through ANTLR
Other — 7:17 PM
Safely as in cyber security, or memory safety or?
Teamy — 7:17 PM
Greater risk to increase the surface that way versus just having the arbitrary code live on the client side and letting it operate via commands
part of the unreleased is the command palette; basically everything that SFM can do is being mapped to client-side slash commands that will be exposed through a CLI and become usable by agents and programs
Other — 7:18 PM
I don't quite get it, would the player be constantly sending the factory manager commands on what to do?
Wouldn't that spam packages like crazy as the thing worked?
Teamy — 7:19 PM
yeah in the worst case probably
SFM is currently an ANTLR grammar that's trusted on the server side, we'd have clients be allowed to manipulate the contents of manager disks. We could add support for more computationally complex grammars but SFM does the bulk of what you want to be doing anyways
SFM is currently unbounded on how complex the programs can become which means that one user can consume all the resources on a server, as part of the role-based access control I'd be implementing there would also be a way for enforcing a computational budget/rate limit on users from the server operator perspective
if you have a server with 5 players online, each player should be entitled to 1/5th the server's resources, and if those players are on teams then they should be able to pool their budget for their base
could do dynamic rate limits where if only 1 of 5 people on a server is using SFM then they get all the budgets until other players start using theirs lol
Other — 7:23 PM
Sounds very restrictive for the player, can they not build giant systems in this case?
Teamy — 7:23 PM
If the system is lagging the server, then one would want the mod to gracefully throttle the system instead of letting the server suffer
there are public servers that disable SFM in modpacks because there's not enough safeguards for mitigating sometimes-accidental laggy sfm constructions
everything will naturally be configurable so that singleplayer or more private environments will be more forgiving with the limits, I'd want the defaults to be very generous
such schemes would make it reasonable to remove the minimum 20 ticks interval as a default for example; if you are wasteful and putting all your programs on 1 tick intervals when 200 would do then you will find yourself out of budget
adding monitoring and advanced introspection will help SFM indicate to the player that "hey, this program could have its interval increased to save budget"
I'm going to conway's law SFM into being as capable as the cloud computing platforms I use in my day job lol
Other — 7:29 PM
Hmmmm
Maybe there would be a compute gauge in the manager screen haha
Teamy — 7:29 PM
yeah the current screen has its charm but I'd want to support a more sophisticated one
keeping both would be part of the incentive to make the screens in the sfm mod interchangeable
Titop54 [CATS], — 7:30 PM
Or parallel managers :hehhh:
Other — 7:30 PM
What about burstable managers haha
AWS style where if you idle it you gain credits, then you can burst when there is high load to consume the credits you gained
Teamy — 7:31 PM
yeah
though it's preferrable to spread the load across multiple ticks so you don't feel a big 2 second stutter every few minutes lol
Other — 7:39 PM
Crazy idea, what if SFM ran on a separate server-side cpu core?
Teamy — 7:40 PM
would need to synchronize to the server ticks but should be possible to have programs that aren't using the same inventories tick at the same time
however, you would immediately run into concurrent modification exceptions if you have two managers using different sides of a double chest perhaps
Other — 7:41 PM
Hmm dunno if its worth it then
Didnt think of that
Teamy — 7:42 PM
it would work, might get tricky if you consider entangloporters are candidates for the same problem of two blocks being able to dereference to the same inventory
but that doesn't mean we can't try
in physics engines, they parallelize by identifying islands of computation that are independent and then it chugs on them all at the same time, no reason SFM couldn't do the same
just gets harder because you may have to curate a ruleset of things that could consider two graphs to become connected even at a distance
having 2 managers on the same network interacting with exclusive sets of inventories for example should be parallelized rather than blindly making single islands out of single networks
Other — 7:44 PM
Sounds like a ton of complexity
Teamy — 7:44 PM
not an insurmountable amount
Other — 7:44 PM
How much could it help with the performance?
Teamy — 7:46 PM
there is a PR from Qther that has some nice performance wins as well #576
the easy wins for performance are done, such creativities may be necessary to let things grow while keeping performance
GitHub
perf: filter caching by Vonr · Pull Request #576 · TeamDman/Super...
as discussed in #571 and various discord DMs and other channels
as discussed in #571 and various discord DMs and other channels
Teamy — 7:47 PM
the wins could be very big if things gracefully work multithreaded while we ensure managers don't step on each others toes
2x+