Comment by abeyer | Hacker News Reader

Comment by abeyer | Hacker News Reader

Parent Full thread

> Games aren't optimized well beyond 2 cores, with games that are optimized for more than 4 cores being particularly unheard off.

How much of that is a chicken-and-egg problem? Gamers buying hardware look for clock speed over core count because that's what today's games benefit from, therefore developers don't optimize for higher core counts because customers don't have them now.

> Games aren't an application that supports parallelism that well since your sound, physics, AI and graphics threads all have to be synced within a single frame otherwise everything falls apart.

I've often heard that repeated, but I'm not sure I totally understand why. I admittedly haven't looked at anything resembling a modern game codebase -- my last time in that space was eons ago in programmer years -- but my instincts would say that doesn't necessarily have to be true, or at least not for that reason.

Graphics and sound kinda make sense, in that you're trying to push out pixels/samples at a fast & regular rate and they need to be perfectly aligned. Though does anything not primarily rely on GPU (which is itself massively parallel) for most graphics work nowadays? Given that, I'd assume the CPU's part of graphics is mostly a support role moving stuff on and off the GPU.

For aspects like AI and physics, I'm not convinced they do have to be locked to frame rate like you say. Why can't they run independently and continuously and let the game take advantage of the "most recent" state of the world as often as needed to match the frame rate? There may be design/architectural reasons this isn't done, but I don't see any fundamental reasons that it couldn't.

Maybe there are factors I'm just missing or not aware of, I'd love to see a good solid analysis of why this is true, if it is.

Dual Core CPUs are still 50% or so of the current gaming PC audience based on steam hardware surveys, chicken and egg or not but until 4+ core CPUs make the bulk of the market game developers will optimize their games for 2-4 cores.

>For aspects like AI and physics, I'm not convinced they do have to be locked to frame rate like you say. Why can't they run independently and continuously and let the game take advantage of the "most recent" state of the world as often as needed to match the frame rate? There may be design/architectural reasons this isn't done, but I don't see any fundamental reasons that it couldn't.

AI and Physics define what is going to be displayed on a screen if say you doing cloth physics this will change the animations of a flag weaving in the wind, if there are thread locks the animation will be breaking up and be choppy. It's even worse if the physics have actual game implication does a barrel hit a player or not if a frame is skipped? And then we get into the realm of multiplayer where you don't just need to sync things within a single computer but between multiple computers all over the world.

Everything has to be synced to make the game work, can it be split across 10 cores probably, but since most gamers don't have 10 cores you better work with 2-4 cores and make sure it works well.

If you've already optimized your game for 2-4 cores, is it that much harder to make it run well on more than 4 cores?

Yes.

People down vote this, so I will provide a bit more explanation: Parallelism is 'easy' as long as you have natural separated problems. This is the reason embarrassingly problems are so great for many-core systems. Every single sub-problem can be done without having to synchronize with other parts of the system. Unfortunately, most problems aren't that way, but instead are interlocked on multiple axes:

- Minimal granularity: You can never run more cores than you have problems at the same time, but you also cannot just split problems as long as you want or the overhead of splitting will kill all performance gains you'd had (x+y in an extra thread/process is technically possible, in reality pretty stupid)

- Synchronization points: Most problems are not completely separate from each other, so you can do part of the problem in parallel but at some point you have to converge and do something with the combined result.

- Comparable task size: In an ideal world all of your tasks would take exactly as long as the other, because if one task takes far longer than the others you have to wait for it. So, the minimal run time of your parallel problem is bound by the time of your longest sub-problem. If one of the problems takes far longer than the others (and they depend on each other) you've lost.

The last one is more of a "meta" point: Complexity doesn't scale linear for dependent problems. That's another reason embarrassingly parallel problems are so nice. You cannot only run them all in parallel, you can think about each one as a completely separate problem. The moment you have dependencies you have to think about how to bring them all together and that gets hard very fast if you have many moving parts.

So, to sum it up: Many small things conspire against just scaling something which works great for two or four cores up to eight, 16, 32 or whatever. If you happen to have an embarrassingly parallel problem you're golden, but unfortunately only a small subset of interesting problems are that way and for other problems scaling is hard.

"In an ideal world all of your tasks would take exactly as long as the other, because if one task takes far longer than the others you have to wait for it. So, the minimal run time of your parallel problem is bound by the time of your longest sub-problem. If one of the problems takes far longer than the others (and they depend on each other) you've lost."

Could the cores that would otherwise be waiting for the longest sub-problem to complete start working on some new sub-problems to make the following set of sub-problems finish sooner?

If you have a set left that doesn't depend on the result of the problem still in computation you can do this, yes. Though in that case you don't actually have two sets, you have one set which is bigger than your core count. I simplified a bit before, but if you have more problems than cores your longest sub-problem shouldn't run longer than the time it takes you to compute all other sub-problems (i.e. that includes computing five sub-problems on one core while another core computes one longer problem) or you'll have to wait.

Unfortunately, I had this happen to a program of mine last year. All other tasks were long done but one kept running and running. The run time of the program escalated to hours with the first few minutes being marked by high usage of the system and then only one core in use, while everything else had to wait. It sucked.

Thanks for expanding on this. My answer of 'Yes.' deserved the downvotes.

Reply on news.ycombinator.com