Building the ultimate 2008 Radiosoft workstation
None of the forum categories seemed appropriate but hopefully this will serve.
I've been tasked with designing the best possible workstation for running the Radiosoft apps base on what will be available around June of this year. The most obvious route is a dual quad-core system using the Intel Seaburg chipset but there are many other details to consider.
How much does memory size affect ComStudy performance? Is it enough to make going to 64-bit and greater than 4GB RAM recommendable? Will there be native 64-bit versions of the Radiosoft products now or anytime soon?
How well does the threading scale? Will building the 8-core beast be rewarded or would I see major diminishing returns after the first few cores are engaged?
Is there or will there be any support for processor farms, as in enlisting idle machines on the network as supplemental processing? (Similar to Folding@Home on a local scale.) This would be especially useful when running a really big job while everybody else has left for the day.
Does video subsystem performance matter much to the workload or does it only matter when displaying the results? The plan is to have two monitors at 22" or greater. A single card can handle this today but an SLI setup would be considered if there is any real gain to be had.
Answer to these questions and any other performance issues are eagerly awaited.
- Login to post comments
What a great topic, thanks!
The biggest problem you raise is a marketing one here. It seems our products (ComStudy®, RadioForms® and RadioCompass®, among others) are often thought of as "RadioSoft®" (sorry about all the "r's", legal says we have to do that).
Development on ComStudy's 10-year-old 32-bit single processor platform is over. So I'll address your questions as if they were asked about RadioCompass, which is
1) Multi-threaded, so it will use as many processors within the local machine or externally (!) as it is allowed to use. Processor farm support is designed into its engine, though not yet implemented, since Vista has made it much more interesting to do;
2) 64 bit instruction-capable, so lots of RAM will be used where available. To answer your "how much benefit" questions, that of course depends on what you typically ask RadioCompass to do! You see:
3) The RadioCompass engine is capable of processing studies on a scale never before attempted. Consider its minimum cell size of 10 cm. when compared to ComStudy's 80 meters, an area difference of 800 x 800 or 640,000 times the point density, and THAT'S ONLY in 2D! When 3D building data, Satellite photos or LIDAR data, lateral diffraction and Fresnel and Fourier derived multipath calculations are added, you'll need every flop you can find.
Some other daunting processor-intensive additions to RadioCompass will be:
1) Users can download GIS data of their own. That's nice but suddenly all the optimizations we've been so good at for years with terrain and vector data are gone, and using every bit of what used to take 1 MB of drive space now spreads over 200 MB, slowing things quite a bit;
2) Method-of-moments calculations, involving potentially many thousands of "wires" contributing to a side-mounted antenna or intermodulation study;
3) Third party plug-ins, over whose coding efficiency we will have no control at all; and
4) Our final goal of making radio fields visible in real-time 3D projection... these are all enormously demanding, well beyond the capability of a typical Pentium in anything like a reasonable time.
Your final question, about video performance, is too early in the RC development to answer. AFAIK there has been no large project made on the dot-NET platform, so no help there, and I cannot say how Vista or its replacement next year will handle video demands over multiple processors. Seems to me one processor will become mostly useless when running real-time 3D projections, maybe more than one? Don't let that stop you from planning two large displays though--who knows where 3D will be in five years?
Peter
#2 above is ambiguous. Is there 64-bit compilations or just 64-bit extension in the 32-bit version?
While a 32-bit x86 app on Windows can support >4GB RAM usage through the combination of PAE in the hardware and AWE in the OS, I cannot say I've ever heard anyone speak fondly of the experience. It was a workaround to build large systems in the days before true 64-bit functionality on the Intel platform and associated operating systems. I don't recall ever hearing of it used at the workstation level.
I ask this because running 64-bit Windows, although much improved under Vista, still adds some hassle factor to building and supporting a system. I recently built a quad-core, 8GB box to run Vista 64 and just having all of that memory available is wonderful. A huge improvement over XP-64. The problem is that the person I'm building this workstation for is likely to be highly Vista resistant. Without native 64-bit apps at issue the machine will likely just get 32-bit XP.
Funny you should ask this. We are just today removing the last of the third-party software which limited us to compiling 32 bit only. So yes, we can (as of next week, anyhow) release both 32-bit and 64-bit native RadioCompass applications.
Warning! We have found a performance *decrease* when using 64-bit. I am not yet sure why this is so, but betting that a 64 bit native app will always represent a faster app may be intuitive but is probably incorrect.
More fun in the playground.
Peter
Yep, a lot of developers have run in to that, from what I've observed in online discussions. 64-bit is not a magic wand for performance. It raises the issue of how much the access to great heaping gobs of RAM helps an app vs. the processing load increase of 64-bit instructions. If your jobs don't cause heavy disk access on a 3GB machine, then you could be better off staying at 32-bit. OTOH, if memory saturation makes for constant drive activity becoming a limiting factor to throughput, then 64-bit could be the solution.
I've seen articles about how to assess data loads, typically for DB servers, to see if there is a win for going to 64-bit but this hasn't come up much at the level I normally operate. As the hardware gets so incredibly cheap I expect the market to make the transition with most consumers being largely unaware of it happening.
So long as there isn't a big problem putting Vista on the monster box, I can always start off 32-bit and easily reinstall as 64-bit (along with a bunch more RAM) later if there appears to be an advantage.
Well, it's been several months but we've yet to build that workstation. Now, the new architecture from Intel, core i7, has arrived and appears far more cost effective than than the premium one pays for a dual socket workstation using Core 2 processors. With HyperThreading (HT) the processor appears as eight logical cores within the OS, each supporting a separate thread. The benefits, of course, depend a great deal on the apps used and how they're coded.
Seeing as how a high-end Core i7 system could be an incredible beast for far less cost than the equivalent number of threads in a Core 2 solution, I'm eager to know if Radiosoft has had a chance to evaluate its products on one of these machines or gotten feedback from a customer who has.
RadioCompass will take advantage of every core, HT or not, available to it - the same goes for RAM. The i7 or C2Q will be a huge advantage when calculating numerous sites.
Real world testing has only been done on Core 2 Quads - calculating 30 FM sites in GA using 3 second data took under 30 minutes - every core was maxxed out through the study. I've been told that after the optimization phase of the development is completed, this could be down to as low as a couple of minutes - can't wait for that!
We have been running RadioCompass(r) on a Dual-Quadcore system with 9Gig of memory running on Win2003x64 Server for several months.
While we have located several "bugs" (that were readily addressed) we have noticed a tremendous improvement when compared to our Dual-Core system with 4Gig of memory running Win2003x32 Server.
While crunching 13+ sites TX/RX at 0dBu can still take a couple of hours it beats taking a 1/2 a day or more with no noticeable hardware issues.
Where are you at on this? An i7/920 quad core/8 thread brand-name computer with 6 gigs of triple-channel DDR3 RAM and dual-monitor 1 GB 1920/1200 video can be had for under $1,000...in fact one from Dell is sitting right in front of me. It would be nice to be able to use this monster to its full advantage.
I've yet to build the RadioSoft dream workstation. The project kept getting delayed while the economy went into a tailspin, making it a persistently low priority.
Now, Intel is shipping the Core i7 980 EE, which has six HT cores for a dozen threads per socket. The Core i7/i5/i3 line is now mainstream. I suspect we'll end up buying a relatively cheap machine that will still be an immense upgrade over the Boss' current box. A decently equipped 860 with 8 GB of RAM should be down around $700 by the end of the year.
I still think it would be a cool idea to have a compute farm approach, where you'd buy licenses to enlist workstations on your LAN purely for non-interactive processing loads during idle periods, such as after business hours. (Think Folding@Home distributed processing for RadioSoft apps.) The investment in creating the capability would be a consideration, of course, but it could be quite a revenue enhancer in the long term. It also means the customer's investment for improved throughput is spent with RadioSoft rather than a make rof high-end PC workstations.
A budget-minded consumer can find i7 builds for around $600-700 with varying amounts of RAM. These systems are refurbs, but just as good as new. High-end C2Q refurbs can usually be found for $400-$800.
As for the 980x - instead of paying the very high premium for this processor, I would suggest buying one of the more mainstream i7 processors, then taking the leftover money and putting it towards more RAM.