Sparse array means two things. One a usage pattern. The current spatial index already uses the sparse array usage pattern. The second is a common implementation detail. The key difference in my change that new Array(2500) creates an array with 2500 slots filled with undefined. My patch skips that pre-allocation. The behavior is unchanged all indices are undefined until assigned to. I have also read some docs that suggest Map might be faster than a sparse array (the optimizer can make more assumptions with a Map and therefore make faster code).

As for lazy init. I was referring to how the runner builds the spatial index at the start of every tick for each room. With zero overhead we could lazily create that index for each kind of LOOK_* the first time it is need on a room. This reduces CPU because we avoid wasting CPU on spatial indices that will never be queried. HOWEVER, the little profiling I've done didn't show that the construction of Game was a huge CPU drain. My procedure was to grab Game.cpu.getUsed() at the top of my loop, access Memory and grab it again. I should repeat that profile, but when I did it with 10 rooms my Memory was <200k and cost twice as much as the Game construction.

My fix doesn't cost CPU. It simply avoids wasting memory that was going to be unused anyway.

Another tradeoff that might help is Lazy calculation of the the FIND and LOOK registers. Construct each Room object with a list of unindexed room objects, then create the indices as needed. That change potentially decreases CPU and RAM. However I think most players run at least a Room.find over every room every tick anyways, so the FIND registers saving are ephemeral. But I think it might be a net win for the LOOK registers.

@wtfrank Oops. Yeah I lost the context somewhere in there. You're correct. I use for..of all the time to loop over the results of Room.find

@o4kapuk Here is a minimal change. https://github.com/screeps/engine/pull/110.

I think style wise using a Map would be better but a sparse array should also avoid the junk memory.

Roughly I expect this to save about 100KiB per room per tick. In @Totalschaden's 50 room case that would be 5MiB.

The savings should compound since the the entire Game object is junked every tick, but only garbage collected every few ticks.

@wtfrank Game.constructionSites is not an array. @Orlet for..of does not work with simple objects.

Looping over the collections in the Game object (e.g. Game.constructionSites) either need to use for..in or _.forEach.

I have a patch that could help if @o4kapuk is ready for the PR.

Currently the spatial index for each room is 2500 array. Swapping that index for a Map (or another sparse index) should help a lot since most rooms are far from full.

What's special about the Buffer class? As I understand the API memory still needs to be stored as a string in MongoDB so the conversion has to take place somewhere.

The best approach I've found for stuff like that is to keep a copy in global and drop checkpoints into Memory.

A concrete example. I store CostMatrix in global but serialize it to a sparse array for storage inMemory.

I really REALLY like the idea of bulk reactions! It seems like an interesting trade off game play wise. Smaller reaction amounts with easier set up plus ability to change as needed vs Large reaction that lock a lab/factory into a resource type until the cooldown is over.

This offers a CPU savings for players AND it would lessen the intent burden in the processor. One of few cases where what's good for players is also good for the servers.

I agree with @Xenofix. Screeps is a great way to pick up javascript or typescript as a second programming language. I haven't heard of any success stories where they picked up screeps and programming cold.

I think it's possible to could craft online course around screeps but you definitely would want to run on a constrained server until the students were further along. I think PvP might not be the best idea in CompSci 101.

It's possible to use the call stack to determine which file you're in. https://github.com/deft-code/typescreeps/blob/master/src/debug.ts#L9 This is very similar to the code I've using now. I use this trick to add a file and line number to my console.log messages.

Using this trick you could write an alternative require proof of concept entirely in game code rather than in the engine. You might have use a different separator; I don't know what the api will do with files with / in the name. (the current require implementation ignores leading ./ so might have to work around that also.