Print 112 comment(s) - last by MrPoletski.. on Jan 27 at 11:45 AM

Sandia simulations reveal memory is the bottleneck for some multi-core processors

Years ago, the hallmark of processor performance was clock speed. As chipmakers hit the wall on how far they could push clock speeds processor designs started to go to multiple cores to increase performance. However, as many users can tell you performance doesn't always increase the more cores you add to a system.

Benchmarkers know that a quad core processor often offers less performance than a similarly clocked dual-core processor for some uses. The reason for this phenomenon according to Sandia is one of memory availability. Supercomputers have tried to increase performance by moving to multiple core processors, just as the world of consumer processors has done.

The Sandia team has found that simply increasing the number of cores in a processor doesn't always improve performance, and at a point the performance actually decreases. Sandia simulations have shown that moving from dual core to four core processors offers a significant increase in performance. However, the team has found that moving from four cores to eight cores offers an insignificant performance gain. When you move from eight cores to 16 cores, the performance actually drops.

Sandia team members used simulations with algorithms for deriving knowledge form large data sets for their tests. The team found that when you moved to 16 cores the performance of the system was barely as good as the performance seen with dual-cores.

The problem according to the team is the lack of memory bandwidth along with fighting between the cores over the available memory bus of each processor. The team uses a supermarket analogy to better explain the problem. If two clerks check out your purchases, the process goes faster, add four clerks and things are even quicker.

However, if you add eight clerks or 16 clerks it becomes a problem to not only get your items to each clerk, but the clerks can get in each other's way leading to slower performance than using less clerks provides. Team member Arun Rodrigues said in a statement, "To some extent, it is pointing out the obvious — many of our applications have been memory-bandwidth-limited even on a single core. However, it is not an issue to which industry has a known solution, and the problem is often ignored."

James Peery, director of Sandia's Computations, Computers, Information, and Mathematics Center said, "The difficulty is contention among modules. The cores are all asking for memory through the same pipe. It's like having one, two, four, or eight people all talking to you at the same time, saying, 'I want this information.' Then they have to wait until the answer to their request comes back. This causes delays."

The researchers say that today there are memory systems available that offer dramatically improved memory performance over what was available a year ago, but the underlying fundamental memory problem remains.

Sandia and the ORNL are working together on a project that is intended to pave the way for exaflop supercomputing. The ORNL currently has the fastest supercomputer in the world, called the Jaguar, which was the first supercomputer to break the sustained petaflop barrier.

Comments     Threshold

This article is over a month old, voting and posting comments is disabled

RE: This is not all that surprising...
By Motoman on 1/17/2009 8:01:46 PM , Rating: 0
OK, maybe my example wasn't the greatest. But the point stands...vast amounts of what any program needs to do in iherently serial in nature, and can't be spread across multiple cores.

RE: This is not all that surprising...
By kkwst2 on 1/17/2009 10:27:26 PM , Rating: 3
As someone who does a lot of computer modeling, I'm certainly biased, but I'd disagree with your point. A large portion of users who really need high performance computers benefit greatly from multiple cores. This includes images processing, video processing, physics modeling, 3D rendering, biological modeling, stochastic modeling, etc.

In my case (fluid modeling) the programs I use scale very nicely well above 50 cores using clustering. In a single node, two 4-core Xeons are nearly twice as fast as a single one, so scaling is quite good all the way up to 8 cores. The number of cores in each node depends on the architecture used and the efficiency of more cores per node certainly is quite dependent on the memory architecture. So, the article has a point but is probably an oversimplification. It seems to assume that memory architectures are not going to advance and scale with increasing cores, which I'm not sure is true.

By masher2 (blog) on 1/18/2009 12:12:18 AM , Rating: 2
Back when I did MHD modeling, the massively parallel supercomputer we used supposedly had several times as much silicon devoted to node-to-node communication as it did to actual computing on each node itself. I can see Intel having to make serious architectural changes to get decent performance from a 16+ core cpu.

In the case of your two 4-core Xeons, though, you have to remember that this is slightly different than one 8-core CPU. The 2x4 option gives you twice as much cache bandwidth, which if you code fits in cache, is going to negate pretty much all the bandwidth crunch from scaling beyond 4 cores.

By Fritzr on 1/18/2009 1:27:03 AM , Rating: 2
When scaling using clusters each node has it's own dedicated memory. The article is talking about multiple cores using a single memory which is what you get with the current multicore processors.

There is one memory connection reached through the memory controller and each core has to share that connection.

Assuming all cores are busy and reading/writing main memory then for a dual core the memory is half speed per core, quad core is quarter speed per core, eignt core is 1/8 speed per core ... as the number of cores goes up, the average available memory bandwidth per core drops.

One work around is larger unshared cache. The bigger the cache dedicated to each core the less that core is likely to need to go to main memory. As new code is written that is optimized in such a manner as to minimize main memory access the performance of multicore will go up.

For now when comparing multicore CPUs you need to look at per core dedicated cache. Larger cache boosts performance of multicore CPUs by reducing memory contention. This was the original solution used for supercomputers...each processor node has a large dedicated memory.

RE: This is not all that surprising...
By Motoman on 1/18/2009 11:06:43 AM , Rating: 1
...your case is wildly different from the normal desktop usage of the average consumer, which is the point I was trying to make. And I was thinking about applications like yours when I mentioned an ERP and an RDBMS. What you're doing is highly specialized, and is well-suited for multi--core/CPU usage.

For the normal, average statement applies. And frankly, so does the guy's post currently at the top of the list, who has now been rated down to zero.

It's like people are already zombified to the idea that more cores is always going to be better.

Pick your favorite game, all you gamers out there. Benchmark it on a single core, dual-core, quad-core, and then 8-core CPU (all of the same family to keep things even, at the same speed, etc.). It's virtually certain that the benchies will fall on their face at the quad-core...and even as newer games come out, there will not be things that can be spread over 8 cores, or 16, or whatever.

...unless, as I've said, some kind of currently unconcievable technomagic can be invented to allow serial processing to occur over multiple (parallel) cores. Which as far as I know, is impossible.

But look at the benchmarks. *Look at them.* And please don't start with the wonky synthetics that have a PC do 40 things at a time - that proves nothing. Average conusmers and gamers don't encode video while ripping MP3 tracks while folding proteins while compiling C# code while typing a letter to Grandma. Bench your favorite individual games and applications.

RE: This is not all that surprising...
By mathew7 on 1/18/2009 3:00:49 PM , Rating: 2
While you are right with the games, there is one point that I have not read until now: the SW will have to adapt.
The games from last 2 years all are adapted to 2-cores (at least those which need CPU performance). I could even say that they ignored quad-cores, because not many gamers had quad-cores (they were developed while quads were very expensive). Switching from 1 to 2 cores was easy for games. But now splicing the workload again will not benefit as much. So doing this on currently released games would have been a waste of time/resources. Probably the games that are half-way in development now can benefit from 4-cores. But that has to be decided from an early stage.
One of the problem is that the current programmers are not used to think with paralel algorithms. Also, paralelism cannot be applied to everything.

Current desktop applications do not require much performance. I mean you could have a big excel file with lots of data, which I'm sure MS had it designed to benefit of as many cores as you have. But the point is that the file should be very big and very complex for you to be affected by current processors (I mean the CPU workload to be timed in minutes, not seconds). At that dimensions, you would be better with a DB application.

By William Gaatjes on 1/19/2009 1:56:22 PM , Rating: 2
True, Since windows NT Version 6 (yes vista) microsoft updated the schedular, interrupt and thread handeling mechanisms to take use of hardware features modern processors have since K7 or the P4 at the least. Windows XP (NT5) uses an anciënt schedular, interrupt and thread handeling mechanisms based on software loops togther with interrupt timers while vista does these things in hardware.

See this link :

The multimedia class service is useless tho in my opinion.
If microsoft would just use a large enough memory buffer for audio data and the audio chip DMA's the data from memory and the cpu get's to update that data before the audiochip runs into the end of the memory region it was assigned to DMA, then you will never notice a glitch.

As is readyboost useless.

Superfetch seems handy but we need more bandwidth from HDD to main memory before superfetch is really interesting.

"We basically took a look at this situation and said, this is bullshit." -- Newegg Chief Legal Officer Lee Cheng's take on patent troll Soverain

Related Articles

Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki