next up previous contents
Next: In Practice Up: Data-Oriented Design Previous: Optimisations   Contents

New version now available here at

This page is from the beta release of the Data-Oriented Design book. There are errors, spelling and factual, and this page is only kept for purposes of maintaining old links.


Any book on games development practices for contemporary and future hardware must cover the issues of concurrency. There will come a time when we have more cores in our computers than we have pixels on screen, and when that happens, it would be best if we were already coding for it, coding in a style that allows for maximal throughput with the smallest latency. Thinking about how to solve problems for five, ten or even one hundred cores isn't going to keep you safe. You must think about how your algorithms would work when you have an infinite number of cores. Can you make your algorithms work for N cores?

Writing concurrent software has been seen as a hard task in the past because most people think they understand threading and can't get their heads around all the different corner cases that are introduced when you share the same memory as another thread. Fixing these with mutexs and critical sections can become a minefield of badly written code that works only 99% of the time. For any real concurrent development we have to start thinking about our code transforming data. Every time you get a deadlock or a race condition in threaded code, it's because there's some ownership issue. If you code from a data transform point of view, then there are some simple ground rules that provide very stable tools for developing truly concurrent software.

What it means to be thread-safe

Academics consistently focus on what is possible and correct, rather than what is practical and usable, which is why we've been inundated with multi-threaded techniques that work, but cause a lot of unnecessary pain when used in a high performance system such as a game. The idea that something is thread-safe implies more than just that it is safe to use in a multi-threaded environment. There are lots of thread safe functions that aren't mentioned becuase they seem trivial, but it's a useful distinction to make when you are tracking down what could be causing a strange thread issue. There are functions without side-effects, such as the intrinsics for sin, sqrt, that return a value given a value. There is no way they can cause any other code to change behaviour, and no other code can change its behaviour either11.1. In addition to these very simple functions, there are the simple functions that change things in an idempotent fashion, such as memset

Thread safe implies that it doesn't just access its own data, but accesses some shared data without causing the system to enter into an inconsistent state. Inconsistent state is a natural side effect of multiple processes accessing and writing to shared memory. It is these side effects that are the cause of many bugs in multi-threaded code, which is why the develpers of the Erlang language chose to limit the programmer to code that doesn't have side-effects. Any code that relies on reading from a shared memory before writing back an adjusted value can cause inconsistent state as there is no way to guarantee that the writing will take place before anyone else reads it before they modify it.

int shared = 0;
void foo() {
	int a = shared;
	a += RunSomeCalculation();
	shared = a;

Making this work in practice is hard and expensive. The standard technique used to ensure the state is consistent is to make the value update atomic. How this is achieved depends on the hardware and the compiler. Most hardware has an atomic instruction that can be used to create thread-safety through mutual exclusions. On most hardware the atomic instruction is a compare and swap, or CAS. Building larger tools for thread-safety from this has been the mainstay of multi-threaded programmers and operating system developers for decades, but with the advent of multi-core consoles, programmers not well versed in the potential pitfalls of multi-threaded development are suffering because of the learning cliff involved in making all their code work perfectly over six or more hardware threads.

Using mutual exclusions, it's possible to rewrite the previous example:

int shared = 0;
Mutex sharedMutex;
void foo() {
	int a = shared;
	a += RunSomeCalculation();
	shared = a;

And now it works. No matter what, this function will now always finish its task without some other process damaging its data. Every time one of the hardware threads encounters this code, it stops all processing until the mutex is acquired. Once it's acquired, no other hardware thread can enter into these instructions until the current thread releases the mutex at the far end.

Every time a thread-safe function uses a mutex section, the whole machine stops to do just one thing. Every time you do stuff inside a mutex, you make the code thread-safe by making it serial. Every time you use a mutex, you make your code run bad on infinite core machines.

Thread-safe, therefore, is another way of saying: not concurrent, but won't break anything. Concurrency is when multiple threads are doing their thing without any mutex calls, semaphores, or other form of serialisation of task. Concurrent means at the same time. A lot of the problems that are solved by academics using thread-safety to develop their multi-threaded applications needn't be mutex bound. There are many ways to skin a cat, and many ways to avoid a mutex. Mutex are only necessary when more than one thread shares write privileges on a piece of memory. If you can redesign your algorithms so they only ever require one thread to be given write privilege, then you can work towards a fully concurrent system.

Ownership is key to developing most concurrent algorithms. Concurrency only happens when the code cannot be in a bad state, not because it checks before doing work, but because the design is such that no process can interfere with another in any way.

Inherently concurrent operations

When working with tables of data, many operations are inherently concurrent. Simple transforms that take one table and generate the next step, such as those of physics systems or AI state / finite state machines, are inherently concurrent. You could provide a core per row / element and there would be no issues at all. Setting up the local bone transforms from a skeletal animation data stream, ticking timers, producing condition values for later use in condition tables. All these are completely concurrent tasks. Anything that could be implemented as a pixel or vertex shader is inherently concurrent, which is why parallel processing languages such as shader models, do not cheaply allow for random write to memory, and don't allow accumulators across elements.

Seeing that these operations are inherently concurrent, we can start to see that it's possible to restructure our game from an end result perspective. We can use the idea of a structured query to help us find our critical path back to the game state. Many table transforms can be split up into much smaller pieces, possibly thinking along the lines of map reduce, bringing some previously serial operations into the concurrent solution set.

Any N to N transform is perfectly concurrent. Any N to <=N is perfectly concurrent, but depending on how you handle output NULLs, you could end up wasting memory. A reduce stage is necessary, but that could be managed by a gathering task after the main task has finished, or at least, after the first results have started coming in.

Any uncoupled transforms can be run concurrently. Ticking all the finite state machines can happen at the same time as updating the phsyics model, can happen at the same time as the graphics culling system building the next frame's render list. As long as all your different game state transforms are independent from each other's current output, they can be dependent on each other's original state and still maintain concurrency.

For example, the physics system can update while the renderer and the AI rely on the positions and velocities of the current frame. The AI can update while the animation system can rely on the previous set of states.

Multi stage transforms, such as physics engine broadphase, detection, reaction and resolution, will traditionally be run in series while the transforms inside each stage run concurrently. Standard solutions to these stages require that all data be finished processing from each previous stage, but if you can find some splitting planes for the elements during the first stage, you can then remove temporal cohesion from the processing because you can know what is necessary for the next step and hand out jobs from finished subsets of each stage's results.

Concurrent operation assumes that each core operating on the data is free to access that data without interfering, but there is a way that seemingly unconnected processes can end up getting in each other's way. Most systems have multiple layers of cache, and this is where the accident can happen. False sharing is when data, though actually unrelated, is connected by the physical layout of the hardware. For example, the cachelines of a CPU might be 16 to 128 bytes long. If two different CPUs try to write to neighbouring bytes, or words, at best the cores will lock up as the memory is shunted around trying to keep memory consistent, but worse, could end up losing data if the cache is not coherent11.2. To reduce the chance of this happening, processing of small element tables should consider this and split any cooperation into cacheline size jobs.

Queues as gateways, and "Now"

When you don't know how many items you are going to get out of a transform, such as when you filter a table to find only the X that are Y, you need run a reduce on the output to make the table non-sparse. Doing this can be log(N) latent, that is pairing up rows takes serial time based on log(N). But, if you are generating more data than your input data, then you need to handle it very differently. Mapping a table out onto a larger set can be managed by generating into gateways, which once the whole map operation is over, can provide input to the reduce stage. The gateways are queues that are only ever written to by the Mapping functions, and only ever read by the gathering tasks such as Reduce. there is always one gateway per Map runtime. That way, there can be no sharing across threads. A Map can write that it has put up more data, and can set the content of that data, but it cannot delete it or mark any written data as having been read. The gateway manages a read head, that can be compared with the write head to find out if there is any waiting elements. Given this, a gathering gateway or reduce gateway can be made by cycling through all known gateways and popping any data from the gateway read heads. This is a fully concurrent technique and would be just as at home in variable CPU timing solutions as it is in standard programming practices as it implies a consistent state through ownership of their respective parts. A write will stall until the read has allowed space in the queue. A read will either return "no data" or stall until the write head shows that there is something more to read.

When it comes to combining all the data into a final output, sometimes it's not worth recombining it into a different shape, in which case we can use conc-trees, a technique that allows for cache-friendly transforming while maintainingg most of the efficiency of contiguous arrays. Conc-trees are a tree representation of data that almost singlehandedly allows for cache oblivious efficient storage of arbitrary lists of data. Cache oblivious algorithms don't need to know the size of the cache in order to fully utilise them and normally consist of some kind of divide and conquer core algorithm, and conc-trees do this by either representing the null set, a concatenation, or some data. If we use conc trees with output from transforms, we can optimise the data that is being concatenated so it fits well in cache lines, and we can also do the same with conc tree concatenate nodes, making them concatenate more than just two nodes, thus saving space and memory access. We can tune these structures per platform or by data type.

Given that we have a friendly structure for storing data, and that these can be built as concatenations rather than some data soup, we have basis for a nicely concurrent yet associative transform system. If we don't need to keep the associative nature of the transform, then we can optimise further, but as it comes at little to no cost, and most reduce operations rely on associativity, then it's good that we have that option.

Moving away from transforms, there is the issue of now that crops up whenever we talk about concurrent hardware. Sometimes, when you are continually updating data in real time, not per frame, but actual real time, there is no safe time to get the data, or process it. Truly concurrent data analysis has to handle reading data that has literally only just arrived, or is in the process of being retired. data-oriented development helps us find a solution to this by not having a concept of now but a concept only of the data. If you are writing a network game, and you have a lot of messages coming in about a player, and their killers and victims, then to get accurate information about their state, you have to wait until they are already dead. With a data-oriented approach, you only try to generate the information that is needed, when it's needed. This saves trying to analyse packets to build some predicted state when the player is definitely not interesting, and gives more up to date information as it doesn't have to wait until the object representing the data has been updated before any of the most recent data can be seen or acted on.

The idea of now is present only in systems that are serial. There are multiple program counters when you have multiple cores, and when you have thousands of cores, there are thousands of different nows. Humans think of things happening at a certain time, but sometimes you can take so long doing something that more data has arrived by the time you're finished that you should probably go back and try again.

Take for example the idea of a game that tries to get the lowest possible latency between player control pad and avatar reaction. In a lot of games, you have to put up with the code reading the pad state, the pad state being used to adjust animations, the animations adjusting the renderables, the rendering system rastering and the raster system swapping buffers. In most games it takes at least three frames

read the player pad.
respond in logic to new pad state.
animate the character.
queueForRender the new bone positions.
FRAME happens, moves the render requests into the procesing list while calling:
renderToBackBuffer on the old lists.
FRAME happens again which starts to render our reaction to the pad state.
swapBuffers then shows our change as

and many games have triple buffering and animation systems that don't update instantly.

In this case, the player could potentially be left until the last minute before checking pad status, and apply an emergency patchup to the in flight renderQueue. If you allowed a pad read to adjust the animation system's history rather than it's current state, you could request that it update it's historical commits to the render system, and potentially affect the next frame rather than the frame three buffer swaps from now. Alternatively, have some of the game run all potential player initiated events in parallel, then choose from the outcomes based on what really happened, then you can have the same response time but with less patching of data.

next up previous contents
Next: In Practice Up: Data-Oriented Design Previous: Optimisations   Contents Beta release of Data-Oriented Design :
Expect errors, spelling and factual. Expect out of date data, or missing stuff. Expect to be bored stiff in some sections, and rushed in others, but most of all, please send any feedback on any of these and any other things that you spot, to

Richard Fabian 2013-06-25