A glimpse at the difficulties of multi threading

The most interesting thing is some games are multithreaded but do not take full advantage of what they have available.

A good example is VALVe Source engine. It has a consol comman that you can put in (MQM) that allows it wo take advantage of more than just one core or possible advantage of the dual/quads. It boosts the FPS in the games a lot but is not 100% stable.

VALVe has stated they are working on it for their next game Left 4 Dead but thats to be seen.

Well, I hope this will stem some of the 'unicorn wishes' that every new app or game coming out will be quad optimized.

Honestly, you would think the fact that dual core has been around for years yet most programs are still single threaded would get that point accross, but it appears the 'quantity is better than quality' mentality can overwhelm logic.

You think thats un-nerving I read an interview with the president of Intel a year ago or so. He said that there next step after the multiple cores is, they can apply a magnetic field to the computer. This forces all the elctrons to the outer edges of the mbo and the other chips (gcard etc.). By pushing them to the edges you reduce all the paths of the electrons making them straight. He said that this could produce a desktop size PC that ran at 10 GHZ or more! This is no-bull I read it in Discover. You guys discussion reminded me of it, I havent tried researching it in a while, but come on this was out of the mouth of the guy who started Intel. It also showed a graph of the increase in processor speed then and what was coming up (Dual Core) I wish I could find that issue to see if that graphs predictions of the futer came true. Anyway thought you peeps might find that interesting Peace.

Carmack said, give me 1 cpu that delivers 10GHz over a quad that does 4Ghz, or something similar. Its true, the need for speed cant be optimized very easily in an OoO world. Thats why graphics need to take a spot in this solution, as gfx cards run in parallel. Theres apps out there, and plenty, where these cards will help. The cpu and the gpu need to find a cozy coexistence until like the above post comes true. Making the electrons straight would increase speed, reduce friction, and allow for better thruput. Making chips for this would be ideal. HKMG might even be overkill if such a thing was practical today

Hmmm..

Question.

Then why does my system (XP home & Vista 64 & Linux) support up to 32 cores on a single socket?

I mean.. so I have a quad system that will never shine as bright as my old P4 northy and my dad's XPAthlon 2100+?

Ok so if multi threading is so far away for the average programmer then why arent the development tools - ie Havoc working with multi threaded capabilities thrown in..


For example 1 core working on physics, another on colision, another on user interaction and the other on game variables.

Surely this would make the game run faster as processors work on the task together - or am I missing the point of multi threaded applications..

In short, thats not nessasarily the case, not becuase of the core count, but because of the processor generation. While you might pump the P4 to a higher freq than the quad (depending which P4 you have) the odds that you can get it fast enough to 'do more work' than a single core on the quad are low.

The long answer:


Interesting territory that delves into, of all things, licencing and packaging as well as logical processors. Remember the old debates about the costs of XP and Vista in relation to multicore because of the socket issue? And the one about XP home supporting only a single core?

First, How does MS define the number of CPUs? By core count, socket or logical count?

Gotta look at the M$ legal stuff for that:

http://www.microsoft.com/licensing [...] icore.mspx

Now that we know how M$ 'counts' CPUs (by socket) how do they support multicore?

Here is a good chart, taken from Paul Thurrotts "Supersite for Windows"

http://www.winsupersite.com/showca [...] _final.asp





Now, according to this, there is no limit on the number of cores Vista will support per socket. Unfortunately, it doesnt address logical CPUs. The number of physical CPUs has nothing to do (not really true) with the number of logical devices'cpus'. If I recall correctly, Vista is limited to 32 logical devices. I dont have a link to prove that, and I very well may be wrong.

The big thing is, just because the OS supports multicore, doesnt mean it makes use of all those cores. Its like a gas tank. You may have a 30 gallon tank in your car, but that doesnt mean it has 30 gallons of gas in it at the moment, only that it can hold up to 30.

as a programer, i can tell you multi threading is hard, it just not worth it. dont forget that two thread can run on one core or two, it is just 2 line of codes, which can never meet.

the prolem is memory addressing a debugging, one thread cant access another memory. having said for games you only reall need to multithread the engine once, and just reuse it.

on top of that multi core/multithread system are still great as you can run many diffrent apps at the same time.

What I meant by shine as bright... usually the single core CPU usage at 100% is easier to obtain, not that it can perform better.

Also, I've mentioned it before, my Vista seems to treat the multi-cores differently, well at least the quad. Kinda like there's a different threshold on the usage. Even Linux has the acts the same to where 1 core has to be fully loaded before the other core takes on another task. Never did take the time to figure out if there's a way to assign a core to a specific task on linux.

Like for example, Super_PI on XP when ran, will only load one core to 100%. On my Vista acts differently, the usages varies from 0-3. The only way to make it act like XP is to set Super_PI affinity to 1 core.

But it's a shame if they can't figure out a way to make mulit-core CPU's better, game wise. Although if you have a 10Ghz CPU, can't imagine how much copper you need to keep it cool.

Well, if super pie runs similarly to Prime, you have to run 1 instance for every core. The embeded 'tasking' apps in TAT are the same. I played with that stuff about 3 years ago on an E6600 (yes, E, not Q). They wouldnt share the load of the instance, but they would spread the indvidual instances themselves accross the cores.

I've been writing multi-threaded applications since the late 80s; while it can be complex, there are various methods of dramatically reducing that complexity.

The simplest is to use message queues between threads, with one thread producing messages while the other thread processes them (e.g. on one side might be your physics thread(s) determining where objects are, the other side might be the render thread(s) actually drawing those objects on the screen). That eliminates the synchronisation problems since the only place where you have to synchronise is in the queue code; the downside is that if the system is poorly designed the overhead of the message passing can be greater than the benefit of the extra thread.

So yes, if you keep trying to traditional write single-threaded code and stick multiple threads into it, that's a huge pain in the ass; but that's like complaining that coding in assembler is a huge pain in the ass compared to C++... you're using the wrong tools for the job.

On Medal of Honour - Airbourne Assault it installs a Ageia Physics card driver wether or not and Ageia card is fitted...

Its a fun game btw..

Now I recommended this game to a mate of mine who is into wargames and such and had a P4 3.2 Ghz and a 7800gs AGP card and it ran like an absolute dog... Wasnt even playable, unlike bf2 and bf:2142 which runs like a dream on his system...

I put this down to the fact that physics will not run on a single core system... The 7800GS all though slow compared to some still cuts some mustard...

Anyway hes got him self a PS3 and the game for this, but the games shabby compared to my pc version - I guess down to the physics "emulation"as I am sure its not programmed in to the PS3 version..

On my C2D 6700 with a 8800GTX it runs unbeliveably smooth with no slow down, so to me it would seem that physics is either running on the second core or with the guts of the 8800GTX..

So this goes further to prove that Ageia cards are an unecessary for a dedicated physic card if the system is hi spec enough.. I mean how many THG frequenters boast about having a Ageia card...

Its about as much use as a 200 dollar network card..

There was talk of an Indiana Jones games with physics built in to the games engine - but in the light of Lucasarts going the way of many of the others, the titles will be passed on to either EA or Activision ( I cant remember who ) it might not even see the light of day..

What would do well for Ageia is to set the standard and license the code in a game creation tool for multi core processors.. Ageia Physics Compliant for example like Nvidia did with TWIMTBP. This maybe a re-invention of the wheel we are looking for whilst the rest of the gaming publishing communities are working out how to get the most of a Q6600 let alone a Skull Trail ( who in their right mind would buy on of these when Nahelem ( oops i7) will out run these) system to give us an enhanced experience over the consoles which are dominating our release schedules..

Unfortunatly, I hate to say it but there is not that much appart from Consoles to PC nowadays - yeah the graphics are better on pc but the game almost plays the same.....Some games like Car racing games, soccer and other sports games are better than pc due to the control pad and steering wheels which never caught on to the pc market as it should of done..

On a second note, I do honestly believe that THG work very hard at giving us benchmarks that are honest, non biased ( some who I would like to mention im sure would disagree ) , true and believable compared to some other un reliable sources which I am sure we all could mention...

All though I'm not a chip, or even a electronic engineer - I have been in this game for nearly 24 years now as a profession diagnosing faults etc and building pcs and have seen things come and go, but the pc is at a bit of a stigmata right now which I would never thought I would see, and needs some clever tools and software to move to the next level to take advantage of what we got, other wise its at a stale mate...

Unfortunatly Vista which promised everything has failed miserably in doing so in becoming the next big thing - more like a little jump and still doing the same thing but looking prettier but with a few complications thrown in ( UAC anyone )

All because your new girlfriend is better looking, doesnt mean her cookings any better !

Well, I dunno what to tell you. Vista may handle things differently, though I doubt it. I suspect your doing something errr....off key, or youre looking at the CPU usage which is the summation of both cores and not the individual loading. You have to look at the CPU usage history to get individual core loading....at least in XP

Im not home, but I just ran prime on my LT, and without further chatter, here it is:



Note ther was no noatble load sharing. Core 1 was loaded, core 2 was runnin g everything else, until I stared running the second instance

Well, on the XP run, only core one is loaded, in the Vista run it does look like its distributing the load amongst the cores---but it shouldnt. Try running 3 and 4 instances in vista and see what it does to the loading. Make sure youre running the FPU intensive in and not loading the memory

Oh come on, you know I rarely take anyones word for anything unless they have the proof to back it up

But your results are indicative that vista does distribute the load....which I find difficult to believe because it is Vista and it is M$. But the thing is, and why I singled out FPU, the memory games Vista plays......I dunno...I just cant see Vista having some form of inherent fine Multithreading....its M$

LOL. in order to make Vista run like XP you'll have to add enough bloatware to consume 5% more processor resources, 25% more memory resources, incur a 5~10% perfromance hit and take up 10 more gig of HDD space

no

For some applications multi threading is nearly impossible, and some just eat it up

Its always going to be that way.

Multi threading in software seems to me that it would require at least 3 threads 2 threads doing work and 1 thread doing traffic control or being memory management thread.

Although I still often wonder why cores cannot be organized more like a bucket brigade for software each core carries the bucket a little ways. I can see this as being a great where the source and destination are only in one place but when you move the target/source around one bucket would work best....

np.. based on what you showed, im going to have to dig around for a liitle somethinng..we'll see what I find. See its still doesnt doesnt add up. Prime should load all of of all your cores to 100%. That its only loading to to 1/4 total usage still indicates that its still behaving normally, but the question is...how would it "know" what the capctiy is...its confusing.

EDIT--------

BTW, I dont think your trying to pull anyones leg...

Yeah I think trans coding video is about the only ap I can think of that can make use of almost unlimited cores (you can break a movie up into frames and assign them to a core for processing with out corrupting what another core is working on)

Edit: Hmm maybe mirroring the data in ram so each CPU has its own complete data set for some programs ?

This is an interesting topic to say the least

Edit 2: I think different techniques would be needed for different problems (reinventing the wheel here I am sure lol but its a good mental exercise and you never know when you may stumble onto something new)

Branch prediction across CPU's is a cool idea to tell the truth it might have problems with some applications but others would work really well. (easily predicted things like video)

np.. based on what you showed, im going to have to dig around for a liitle somethinng..we'll see what I find. See its still doesnt doesnt add up. Prime should load all of of all your cores to 100%. That its only loading to to 1/4 total usage still indicates that its still behaving normally, but the question is...how would it "know" what the capctiy is...its confusing.

EDIT--------

BTW, I dont think your trying to pull anyones leg...

----NEW ENTRY-----

Grimmy,

Ive been doing some digging, but havent found anything yet.

Heres the thing, and its difficult to explain textually....

If Vista is doing what you beleive its doing, if you run 1 instance of prime95, single threaded, you should be seeing 100% load on all four cores.

If Vista is load distributing the actual single thread routine, then immediatly after you start prime, as the load on the primary core builds, some of it will be shunted to the other cores, leaving some of the prim ary cores resourses free. But because prime will consume all available resourses, those resourses 'freed' in the primary core by distributing the load to the other cores should still be consumed, and as the load builds it should continue to be distributed, until the resources of all 4 cores are completely consumed.

In short, given the way you think Vista is handling apps, what you should see is a near instantanious load cascade with all four cores sequentially building to full capacity. UNLESS, there is a limit to the resources prime95 will consume. Now, I have never heard of a limit to the FPU calcs prime will run, but I have heard of memory limitations, which is why I noted to run FPU only. So, if there is an FPU limit, then what you have shown should prove what you beleive to be true. But then, if there was a FPU limit, we should have known about it before now. That 'limit', would have shown for example, as a 50% distribution for an E6600 running prime in Vista. It should also show as something less than ~25% for say, a Q9550...maybe 17`19% accross all for cores (a loose guesstimate).

Now, assuming there is no limit to the FPU calcs on prime, then what youve shown indicates to me that something fishy is going on somewhere, either with prime or with Vista and this is the confusing part, becuase if vista is load sharing as you beleive, it shouldnt be limiting prime to 25%/core.

Guess where I think the fish stink is coming from. And no, its not you, you know better than that. I dont think youre pulling anyones leg, but I think M$ may be pulling all our legs. It wouldnt be the first time and it certainly wont be the last.

Grimmy,

I understand how prime works. I used the same version to run the demos on my laptop the other day.

What Im trying to explain is that the program for the prime calculation is linear. You can run multiple threads, but those are only individual instances of the same calculation.

How you seem to be seeing this is that the prime FPU cacluation is finite, that is that there is a limit to the number of loops it will generate, like a game. Again, if this was so, we would have seen that limit a long time ago, on any core CPU, single, dual or quad. But prime consumes all the resourses it can if you set it to.

Vista cant 'load share' a program of its own accord...it can only do what the program tells it to do. If the program is written in such a manner that it will allow the OS to distribute the load (multithreaded) AND the OS is capable of such tasking, then that can be accomplished. Regardless, Vista itself cant 'break' apart a program to distribute operations among cores, it can only distribute what the program 'tells' it can be distributed. In the case of prime, were Vista to run like you think it is, shunting calculation loops of its own accord, then all 4 cores should load to 100% on a single thread. We both know this is not happening....thats not neither the question nor whats confusing.

Whats confusing is how Vista is presenting the information on the loading...essentially, it looks like Vista is lying, which frankly would not be suprising.


Here the math behind the prime calcs, from the mersenne site itself.

http://www.mersenne.org/math.htm

From the Great Internet Mersenne Prime Search

If Vista were to load share of its own accord, 1 threads worth of the above should load every core to max.

Heh.. I'm okay at math, but thats quite abit for me to take in. So... the only thing I understand, is that you think Vista lying. I suppose that could be true, but it seems when I watch the temps, the corresponding load does cause the same cores temp to go up. And using the affinity can put the entire load on what ever core is chosen.

Just wish I had my quad system up and running. I'll be getting my MB replacement this week, and it's my vacation. Perhaps I could do some more tests once I get it back up and running.

If you can think of any other test to try, I'll try to do it with a screen cap.

I'm surprised I'm the only one seeing this. Can anyone else with a quad on Vista contribute some info or your experience?

Grimmy,

If the primary prime (no pun intended) calc is fine threaded, it would allow Vista to shunt those loops, which would explain distribution, but Vista should not be able to 'cap' the number to 1/4 usage per core.....soemthin wierd is going on....I think I have to aks someone to run some tests on this, though I doubt they will.

That wasnt my intent, and I appologize if it came off it that way.

My meaning was this: Every day, here in the forum, anytime asks the question "quad or dual core?", invariably there are several people who will answer "get the quad, all new games are being to be written for quads"

This is simply untrue, and will always be untrue. It is the result of the "more is better" mentaility...making statements without knowing the facts. Eventually many games will be optimized for multicore., but not all. The simple little games will never need multicore. The more in depth games will see benefits from multicore. But as so many others stated in the thread, writing mutlithreaded is more difficult, time consuming and costly. If, for a given game, a single thread will provide the same results as a multithread, there is not point to expending the extra resources.

What this means for the time being is that rather than just running out and buying as many cores as you can, you should look at the apps/games you want to see what they will benefit the most from. If someone wants to play FSX, or render heavy 3D, or transcode, a slower quad will benefit (depending how slow) them more than a fast dual. Conversley, if they want to play Doom 3, RTCW, soloaire, surf the web or run a wordprocessor, they will get more beneift from the faster dual. Contrary to all the "everything is going to be quad" kiddies 'unicorn wishes', by the time there is sufficient tri or quad threaded software to undeniably offset the advantages of clock speed, every CPU now on the market will be obsolete....thus negating the 'future proofing' aurgument.