Aeshen

CPU's still the key?

Multi-core CPU's are an important development, but they may not be quite as relevant in a few years.  That's not to say that they'll go away, but they won't be as exciting.  I think four cores is a good number for home use, and once we have affordable quad-core processors in most machines, increases won't be as important.  When you think about the average use for a home computer, you think about word processing, web browsing, photo management, maybe home video production for higher-end users.  All of these tasks can certainly take advantage of multiple cores, but it's not very critical (well, more so in video work probably).  The other big home computer use is for gaming.  This is a different world entirely.

Games are much akin to the simulation and analytical processing done for research, only demanded in real-time.  We acknowledge that research requires powerful workstations, but of course the average gamer can't afford a real workstation (though some of them are close!).  In order to bridge the gap between a workstation class machine and a home machine, gamers buy add-on cards.  Modern sound and video cards are actually parallel processors working to offload tasks from the processor.  Most modern video cards contain GPU's (graphical processing units) with more parallelism than many CPU's.

There's still more to go though.  Ageia makes a physics processing unit (PPU), PhysX, on a card that simulates fluid dynamics, cloth, particle effects, and other real-world object interactions.  Not just another core, this is a whole separate processor just for physics calculations.

An even newer product, the Intia processor, by AIseek, promises to bring still more realism to games by offloading artificial intelligence simulation from the CPU.  The Intia is intended to improve terrain analysis, path finding, and line-of-sight sensory simulation.  This is important stuff for dynamically changing scenes when precomputing paths, hiding places, etc. won't help.  The Intia doesn't appear to be available yet, but will really change the playing field in games with lots of NPC's when released.

My point with all of this, is that multiple cores on the main general CPU is definitely still important, but I think that many of the tasks in games and simulation benefit from more specialized processors.  AMD is talking about marrying an ATI GPU to a CPU.  This will reduce overhead considerably.  Perhaps we'll see CPU "arrays" that couple the general CPU, GPU, APU (audio), PPU (physics), and AIPU (AI?) onto a common processor bus, independent of PCI/PCIe/AGP.  This allows the CPU to handle just the general orchestration of the software, dispatching work to be done to the worker processors.  The CPU can perform all of these functions already, but it isn't optimized for many of the tasks.  Heck, some smart folks have used GPU's for additional math computation -- without respect to any graphics.  As with so many advances in computer science, maybe it will just come down to more extensions to make the CPU more adaptable.  As usual, only time will tell!

Arian Kulp
akulp@3leaf.com

PC's can have multiple cores too!

This morning I was looking up information on the three next-gen consoles (Sony, MS, and Nintendo) and I saw something that struck me as funny. On Sony’s site there is a PS3 commercial touting the benefits of the hardware – to the point of relegating the games to the end of the ad. What was amusing to me was the bragging about the Cell processor. Their big point was it works in parallel and that it can do so much more. This is true of course – parallel processing is the way to go for complex tasks like path-finding, physics, AI, etc that are part of modern gaming. The thing is, they bragged about how the PS3 uses eight processors running in parallel while the PC only uses one and can only do one thing at once. With the new Intel Core line with two and four processors per chip, this just isn’t really true anymore. Of course, most PC’s don’t use multi-core or multi-processors yet, and 8-way processors aren’t on the desktop quite yet, but going beyond one core has been an option since before the PS3!

This brings me to my next point though. There are really two aspects of parallelism. I agree with Michael that people just aren’t as interested in upgrading their software (witness Microsoft trying to sell new copies of Office), but they aren’t really upgrading their hardware for multi-core either. Hard-core gamers are always pushing the curve out further with the best hardware, but that doesn’t really trickle down to the masses very quickly. Without parallel hardware on the desktop, most application developers won’t feel the push (from those multi-core consumers) to parallelize, which further reduces the perceived need to buy the hardware. It’s a cycle! It’s also disappointing. Well-done parallelized software is more complex than traditional software, but not by an order of magnitude such that it should be considered a barrier. In many cases, it’s just taking advantage of well-written code. Synchronization can be tricky, but even single-core hardware can perform better when appropriate work is divided between multiple threads.

Not to pick on anyone specifically, but I’m going to mention Google’s acquired Picasa product for photo organization/management. Overall I love it, and there are some clearly threaded portions of the application, but it has some way to go. It imports, uploads, and sends emails on a background thread, but it performs all other image file I/O synchronously. If I select a batch of photos and update keywords on them, it writes each file before returning control to me. Perhaps most people don’t experience much lag, but I store my photos on a network drive – writing even five 2 meg image files can take 10-15 seconds on my 54Mbps wireless connection. I’m sure that for local files it’s nearly instant, but it underscores a bad assumption. Just because testing was done on local files, or perhaps networked files on a fast wired network, doesn’t mean that writing to files will always be a quick operation. I could choose to update 100 local files and it would probably also take some time to complete. Had they parallelized it properly from the start, it wouldn’t matter either way.

In my opinion, developers need to plan better. It was a paradigm shift years ago to imagine your application functionality in terms of objects. It’s time now for another shift where developers think in terms of processing pipelines – background threads ready to perform operations in discrete steps. If game consoles are there, why isn’t the desktop?

Arian Kulp
akulp@3leaf.com

Pipeline Threading with Intel’s Threading Building Blocks Library

Recently, I was tasked with writing a multi-threaded C++ sample application as a way of highlighting Intel’s multi-core processors. I decided to write a file compression program that compresses the contents of a specified directory branch. Once the branch to compress is specified, the processing can be broken down into 3 clearly defined sub tasks: file scrub and input, block compression, and output to the compressed file. I envisioned each of these tasks running in its own thread, with thread-safe queues to pass intermediate work between them. Of the code I had to write, the trickiest, I thought, would be the 2 queues.

Before I finished coding the application, a new requirement was added to the project. Intel has released a new C++ runtime library called “Threading Building Blocks.” I was asked to use this library in the sample application. So I downloaded an evaluation version of TBB from here, and headed back to the design board.

TBB comes with great documentation, including a “Getting Started” guide, a “Tutorial” that goes into more depth, and an exhaustive reference guide. I read the tutorial first, which gave me all I needed for my purposes. TBB offers support for several threading paradigms, including loop parallelization, task pipelining, thread-safe containers, and much more. The TBB pipeline was the best choice for what I was doing. I put each task in its own C++ class, registered the classes with the pipeline, and let it run. The best part was that I didn’t need to write the queues; TBB provided for passing the intermediate data between tasks.

I must say that dealing with the Threading Building Blocks was an unexpected pleasure. My experience using software tools for the first time has been a mixed bag at best; it’s often extremely frustrating to understand and utilize these tools quickly. For instance, the public-domain compression engine I originally intended to use cost me hours of configuration and building headaches, and I never did get it to build in a way that I could use. I ended up tossing it out and using a different compression engine altogether. TBB on the other hand was well documented, easy to understand and implement, and worked right out of the box. I must say that the next time I need to write a C++ application using multiple threads, I’ll likely reach for TBB.

The sample application, along with a readme file that does into the TBB code in some depth, can be found here.

Richard Bowler
rbowler@3leaf.com

Entertainment & MultiThreading

Does the phrase “what is this multi-core stuff all about” sound familiar?  “Why do I care”, say you? 

Personally, I think one of the biggest areas for multi-core hardware and multi-threaded software design today is in digital media applications; specifically entertainment applications.  Do you know what Microsoft^TM Windows XP^TM Media Center edition, Intel’s Viiv^TM platform, the XBOX 360^TM and several other entertainment platforms all have in common?  They all take advantage of multi-threaded software in a fairly big way.  And why not, right?  I mean surely few of us can think of a more taxing process for a computer than running video and playing games right?  I know somewhere out there a comp-sci post doctoral is saying “uh yeah, protein folding, weather pattern analysis and global search engine software”.  To which I say, yes, but few consumers are going to assemble a Google knock-off in their living room (although I admit that would be fun).  The lion’s share of computers sold today for corporate or home use are going to see a lot of office application, web browsing and maybe some occasional digital media (pictures, video, etc.) use.  But when someone invests into one of the aforementioned platforms they’re getting a bona fide digital media workhorse.  This means these machines are intended for some heavy duty processing.  The typical tasks that are expected of them include displaying pictures, updating media information from the internet, managing scheduled tasks, decoding live video and sound, and playing games – sometimes, all at the same time.  And this isn’t even accounting for the usual autonomous tasks a computer is doing as part of just being turned on.

While Windows XP MCE is also quite capable in a single core environment, it really sings in a multi-core setup.  Windows XP itself comes out of the box multi processor capable.  This has been true since the Windows NT days.  But layer on a digital media platform such as Viiv or MythTV and you have even more ways to tax as many processors as you can squeeze into your chassis.  Intel has positioned its Viiv platform to be the choice du jour of PC media centers.  With highly tuned hardware and software meant specifically for digital media applications it’s certainly a compelling platform.  Imagine recording the latest showing of 24 while the rest of your family is streaming music to other computers.  Or maybe you’re recording two shows while watching a DVD and also downloading a movie from NetFlix.  These scenarios beg for multi-core setups.

Nowadays with High Definition becoming more prevalent a system’s cpu(s) are going to be leaned on even harder.  According to the Xbox platform team nothing maxes out the three cores in the Xbox like the new HD DVD player.  This entirely software based solution replaces the DSP chips in other consumer players and allows MS to offer HD DVD to the 360 owner for $199 MSRP as compared to the $500 - $800 for the standalone players.  Like playing Xbox 1 titles on your 360?  Thank the three cores present in the 360.  One of them dedicates itself to emulating the original Xbox operating environment while another runs your game.  Heck, there’s even a core to spare to update your Xbox Live statistics as you set the single game passing yards record in Madden NFL.

So now you have a few reasons to care about all this multi-core stuff, or at the least a few more selling points to convince your someone to get on board the purchase for Christmas.

Jason Shigley
jshigley@3Leaf.com