Summary

Weaver is a mod that roughly triples the performance of DSP. The mod does not alter gameplay, does not disable achievements and can be enabled/disabled anytime. In my tests of late game saves, Weaver improves the game's performance by roughly 3.0-3.5x. You'll experience the best performance while flying in outer space. You can read why in the How it works section.

Mod compatibility

Known compatible mods:

• SphereOpt
• DSPOptimizations
• GalacticScale

Known incompatible mods:

• SampleAndHoldSim
• GenesisBook
• Multfuntion mod
• More Mega Structures

Known issues

None. You can report bugs here.

How it works

The general idea is to avoid doing work that isn't necessary. The game spends a lot of time updating, for example, animation data on planets the player isn't on. The mod solves this issue by "optimizing" planets the player isn't currently on. Whenever a player flies to a planet, the planet's logic is restored to the game's default logic to ensure all animation and UI logic functions as expected. Anything being built, upgraded or destroyed on a planet will also, briefly, cause the planet's logic to revert to the game's slower logic. Whenever a research completes, the mod will reoptimize all planets one at a time to ensure the optimized planet includes any stat changes from the research.

Optimizations

• Assemblers - Assemblers, Smelters, Chemical plants, Oil Refineries, Particle colliders,
  • Optimized data format
    • Recipe and assembler tier data is relegated to a separate struct that is shared by all assemblers using the same recipe and tier.
  • Can go inactive when they are unable to function due to missing items to craft with or due to the output being full.
  • Parallelized logic on a sub-factory basis.
• Sorters
  • Two optimized data formats
    • Bidirectional sorters
    • Non-bidirectional sorters
    • Each format relegates tier information to a separate struct that is shared by all sorters in that tier.
  • Optimized data access when interacting with entities
    • Optimized access to belts, assemblers and labs
  • Sorters can go inactive when the entities they interact with are unable to provide/accept the items transferred.
  • Parallelized logic on a sub-factory basis.
• Labs
  • Two optimized data formats.
    • Labs producing science.
    • Labs consuming science.
    • Can go inactive when they are unable to function due to missing items to craft with or due to the output being full.
  • Parallelized logic on a sub-factory basis.
• Belts
  • Parallelized logic on a sub-factory basis.
    • All locks in belt code are gone.
  • Items stored on belts take up less memory.
• Spray Coaters
  • Parallelized logic on a sub-factory basis.
  • Optimized data format.
    • Optimized access to sprayer belt and sprayed belt.
  • Optimized access to belts.
• Monitors
  • Optimized access to belts.
  • Parallelized logic on a sub-factory basis.
• Fractionators
  • Optimized data format.
  • Optimized access to belts.
  • Parallelized logic on a sub-factory basis.
• Power system
  • Optimized data format.
  • Optimized power networks
    • Optimized Power Exchangers.
      • Optimized data format.
      • Optimized access to belts.
    • Optimized Ray Receivers.
      • Optimized data format.
      • Optimized access to belts.
    • Optimized all Power generators.
      • Optimized data format.
      • Wind power generation is now simply windGeneratorCount * planetWindStrength.
  • Calculation of power consumption is now done in an entitys regular update loops.
  • Ray receivers view of a dyson sphere has been parallelized.
• Splitters, Pilers and Monitors
  • Optimized data formats.
  • Parallelized logic on a sub-factory basis.
  • Optimized access to belts.
• Miners
  • Optimized data formats for all miners.
    • Oil Extractor.
    • Water Pump.
    • Mining Machine.
      • Optimized access to belts.
    • Advanced Mining Machine.
      • Optimized access to belts.
  • Removed all locks on veins and mining flags.
  • Parallelized logic on a sub-factory basis.
• Depots
  • Parallelized logic on a sub-factory basis.
• Tanks
  • Optimized data format.
  • Parallelized logic on a sub-factory basis.
• Digital system
  • Parallelized logic on a sub-factory basis.
• ILS and PLS
  • Parallelized belt interaction logic on a sub-factory basis.
  • Optimized access to belts.
• Ejectors and Silos
  • Parallelized logic on a sub-factory basis.
• Turrets
  • Optimized access to belts.
• Multithreading
  • Step by step logic has been replaced with a work stealing worker system.
  • Independent factories on the same planet are considered sub factories. Sub factories are threaded independently from each other even though they are on the same planet.
• Statistics
  • Production statistics
    • Replace locks around updates to production statistics with per-sub-factory statistics that are aggregated afterwards.
    • Track which items a sub factory create and only calculate statistics for those.
  • Traffic statistics
    • Parallelized statistics calculations.
    • Track which planets/solar systems has statistics updates and only do expensive updates on those. The rest do a fast statistics update.
  • Combat statistics
    • Parallelized statistics calculations.
    • Track which planets/solar systems has statistics updates and only do expensive updates on those. The rest do a fast statistics update.

Could DSP developers implement these optimizations?

Spray coaters and statistics

Parallelizing spray coaters is relatively straightforward and provides a significant performance improvement in most saves. Same goes for the statistics optimizations this mod does.

Data layout

The general implementation of DSP is quite well thought out. The use of structs of entities makes the code already quite optimized. The biggest performance problem is that the structs are too large. Even if a sorter is on the other side of the star cluster, the sorter's animation-relevant data still has to be loaded in when the game updates the inserter. A lot of the optimizations in this mod could be implemented by the DSP developers by making even more use of struct-of-arrays design. That's what this mod uses to reduce memory bandwidth requirements and improve cache hit rates. Simply adding the attribute [StructLayout(LayoutKind.Auto)] to each of the game's large structs should improve performance and reduce memory usage.

Deduplicating data

Additionally, a lot of the information in each struct is duplicate constant values stored in each struct instance. For example, this includes recipe requirements and entity tier data. This can easily be moved to a separate array of structs that can be referenced, just like how this mod does it for assembler recipe information. This type of optimization can (so far from what I've seen) be applied to the following components: assemblers, inserters, labs, consumers, AnimData, PowerConsumerComponent, PowerGeneratorComponent and probably a lot more. This would also reduce the game's memory usage.

Data access

Other optimizations are a bit more difficult. These would be the data access optimizations. These optimizations reduce the number of indirections necessary to get to the data an entity actually cares about. It is quite easy to add an additional array that stores the network ID for each inserter so the network ID doesn't have to be fetched from EntityData with a random access memory request. But doing so means duplicating information in multiple places. Updating this kind of information would require updating it in multiple places. It can quite quickly result in very complicated update logic. This mod deals with that by re-optimizing a whole planet at a time as described in the How it works section.

Multithreading

Before explaining how Weaver improves DSP multithreading it is important to describe how the game currently utilizes multithreading. This will make it easier to explain why the multithreading logic was changed.

Existing multithreading

The game multithreads its logic by splitting the simulation logic up into discrete steps. A step could, for example, be to update all sorters in the game. The game manages multiple threads by creating a worker for each thread. Each worker is delegated a static amount of work up front and all workers has to complete their work before all workers can start on the next step in the simulation. This means all workers are always waiting for the slowest worker before they can start the next step. Not all CPU cores are equally fast and it is not possible to perfectly distribute the work beforehand. The end result is a mutithreading system that is not capable of 100% CPU utilization. Certain steps are also not multithreaded, like spray coaters, as previously explained.

Weaver multithreading

Weaver analyzes how entities on each planet are connected and splits up independent factories into sub-factories. A sub factory consists of all entities that are connected. Commonly in larger games, a sub factory consists of an ILS/PLS, the belts connected to it, sorter connected to the belts and assemblers connected to the sorters. Simulation of a sub-factory is done on a single thread. As only one thread updates a sub-factorys entities, the entities no longer need to be thread safe. Their locks and atomic operations has therefore been removed. Splitting simulation up into sub-factories allows the game to more effectively utilize multiple cores. Cache hit rate improves as well as all data related to a sub-factory usually fits inside a cores cache.

To split the simulation up into sub-factories, the power system has to not connect entities together. Weaver handles this by updating a planets power system before updating the planets sub factories. A few other things can also not be simulated on the sub-factory level. A notable example is logistics distributors which must be updated in a certain order on each planet. To simulate a planet, weaver splits it up into multiple steps. Firs the power system is updated. Then all sub-factories are updated. Lastly logistic distributors are updated. To efficiently execute these steps for all planets, Weaver implements a work stealing work pool where workers compete against each other to complete steps as fast as possible. There is no synchronization of step execution between planets like there is with the games multithreading logic.

Workers will prioritize executing work for a specific planet, unless the planet has no work immediately available. In that case the worker will attempt to find available work on other planets. Only if it can't find any immediately available work in any planet, will the worker have to wait for other workers to complete their work. A worker will attempt to reserve work in the next step before it waits. If it can't reserve work then it will find another planet to try and reserve work in. This ensures 10 threads aren't waiting in a planet where the next step only consists of 2 work chunks.

Power system

Weaver calculates the energy consumption of an entity right after updating the entity. Although the same work is being done, the work is cheaper due to the entities power data already being in cache. This effectively eliminates the performance impact the power system previously had in the game. Additionally, wind power generators are now calculated in constant time. This is possible because all wind generators generate the same amount of power. Counting the number of wind generators and multiplying that by a planets wind strength is all that is needed to calculate their power generation.

Potential additional power system optimizations

This same methodology could be used for geothermal power generators once they've all warmed. The performance impact of ray receivers and calculating their view of dyson spheres could equally be optimized by excluding ray receivers from this call if they've been determined to always view the dyson sphere. This check could simply be done by determining if the entire planet is inside a dyson sphere, by determining if the ray receiver was created close to the planets poles or by checking if they are placed on a tidally locked planet. Solar power could make use of the same optimization.

Statistics

There is three primary types of statistics that weaver optimizes. First one is production statistics. This type of statistics keep track of how much each planet has produced/consumed each game update. The two other types of statistics are combat statistics. They track how many enemies have been destroyed and traffic statistics which tracks how many items are imported/exported from each planet. True for all types of statistics is that their calculations, for each planet/solar system, are fully independent of each other. Weaver has multithreaded the statistics calculations for all the statistic types.

Production statistics

Internally the game stores production and consumption statistics in two arrays of size 12.000. An items item id is the index into the arrays that the item uses to store how much was produced and consumed. Each tick the game goes over these two arrays of size 12.000 and updates some statistics history data about how much was produced on that game update. Then the game proceeds to clear all data in the two arrays for each planet. There is roughly ~200 items in the current game. Checking for all the remaining ~11.800 non existing items and then clearing them for every single planet is unnecessary and expensive. Weaver optimized production statistics by letting each sub factory figure out which item ids it produces and consumes. Each sub factory then creates two smaller arrays, one for production and one for consumption, where each array only keep tracks of the items the sub factory produces/consumes. Once all sub factories have been updated, a follow up aggregation step runs that sums the production and consumption data for all sub factories on the planet and adds it to the historical statistics. The game was modifying the 12.000 item large arrays in parallel which required locks around updates to the arrays. Since updating a sub factory is handled by a single thread, these locks no longer exist. The result is an implementation that is more cache efficient and which allows multiple threads to update production statistics on a single planet in parallel without any need for synchronization primitives.

Traffic and combat statistics

Traffic and combat statistics are very similar to production statistics but with an important difference that will be described shortly. The arrays for traffic statistics are also 12.000 large while combat statistics use arrays of size 2.048. There exist as many items for traffic statistics to keep track of as there exist items for production statistics to keep track of. I would guestimate there exist less than 50 types of enemies.

Weaver splits updating statistics for these two types into an expensive and cheap update. The expensive update is where both arrays are fully iterated. Secondly is the cheap update where the historical statistics are simply updated with no changes to the import/export and destroyed historical data. The cheap update can only be used when there is no changes to the statistics of a planet/solar system. The important difference between production statistics and traffic/combat statistics is that traffic and combat statistic arrays are accessed through methods. This allows weaver to observe which planets/solar systems these methods are called for and then run the cheap update for all planets/solar systems that were not updated.

The same method used for production statistics would be more efficient than what is currently used but it's also a lot more complex to implement. The optimizations applied for statistics are already making the statistics performance impact irrelevant so there was no need to go overboard.

Base game

I hope most of these optimizations are eventually merged into the base game. Then I don't have to maintain them and more players would be able to enjoy them.

Future plans

There are still quite a few optimizations I've yet to implement. I believe it is possible for the mod to improve the game's performance by ~4x.