Mac Pro vs. SGI Prism
Hey guys,
Although this question may be too early to ask, I figured I'd inquire what your thoughts are... We currently have an 8-processor SGI Prism with 16GB RAM, two 160GB SATA HDs, Quad monitors, etc. I can't remember the speeds on the processors (maybe 800mhz), but they are Itaniums. Does anyone have a comparison between the Woodcrests and those Itaniums?
Doing a Mac Pro setup with Tiger Server would be pretty cool to augment our current setup, and we can get a Mac Pro 3ghz, 16GB RAM, 2TB, Quadro FX, two 30" Cinema displays, etc for around $15k which is a great price. But I'm not sure how the two dual-core 3ghz Xeon Woodcrests stack up to the 8-processor 800mhz Itaniums. Perhaps waiting for the new Xserves in October may be better for our use? Any ideas?
Although this question may be too early to ask, I figured I'd inquire what your thoughts are... We currently have an 8-processor SGI Prism with 16GB RAM, two 160GB SATA HDs, Quad monitors, etc. I can't remember the speeds on the processors (maybe 800mhz), but they are Itaniums. Does anyone have a comparison between the Woodcrests and those Itaniums?
Doing a Mac Pro setup with Tiger Server would be pretty cool to augment our current setup, and we can get a Mac Pro 3ghz, 16GB RAM, 2TB, Quadro FX, two 30" Cinema displays, etc for around $15k which is a great price. But I'm not sure how the two dual-core 3ghz Xeon Woodcrests stack up to the 8-processor 800mhz Itaniums. Perhaps waiting for the new Xserves in October may be better for our use? Any ideas?
Comments
If x86 then you are not even using the Itaniums really - the hardware emulation on those things is worse than using a software emulator. You would see orders of magnitude increase on the Intel Mac.
Assuming, of course, that this is something you either have the source code for or that exists in a Mac universal binary version.
I was thinking the same thing about the Itanium's hardware-emulation... Although my MacBook Pro is only a dual core, some of the programs I build using our research application seems to run almost as fast (if not as fast) as it does on the SGI (with 8-processors!).
One other concern that I have is the use of MATLAB. I know that Mathworks doesn't have an Intel-compatible version for the Macs yet. Hopefully they will by spring of next year. Currently I've installed it on a Windows and Fedora virtual disk in Parallels on my MBP, but of course its code execution times doesn't compete with native Intel builds. I'm wondering how much faster Parallel's MATLAB may be on the new Mac Pro?
Do you by chance have any code that maybe I could run on the Prism and perhaps a Mac Pro (that someone else can run for me) to compare some speeds? Basically some simple code that can run on all the processors and for a certain number of iterations (on both computers)?
BUT - you need to set gcc to output the VLIW (architecture IA-64) binary. gcc is supposed to support that, but I have not tried it.
The Redhat you are running should be VLIW native, as it also supposedly comes in a VLIW flavor.
If you intend to test it under Redhat, you would have to find out what the Linux APIs are for the number of processors and the threading in Linux and port those calls over to whatever Linux expects.
Running x86 binaries on the Itanium is eight times slower than running them on a regular x86 chip.
What is it your trying to run, a better questioned would have been "I have this program, what do you think the best monster is to run this f-cker".
What's with the exotics? Do you guys just like pretty boxes, you must if your using X86 code with an Itanium machine (especially a Prism). Not harping on you man but even an Apple sounds more like your trying to decorate your office then use the machine for what it's intended for. A simple but powerful Sun-Fire X4200 from Sun would be MORE then suffice for what your doing and at a fraction of the cost. Unless it's graphics work, but you mentioned interest in the Xserver so....
What is it your trying to run, a better questioned would have been "I have this program, what do you think the best monster is to run this f-cker".
Hehe, good question. There are two main programs that we would like to run on it:
1. MATLAB (http://www.mathworks.com)
2. SCIRun (http://software.sci.utah.edu/scirun.html)
MATLAB unfortunately will not be supported on the Prism, however, there is a piece of software called Star-P (http://www.interactivesupercomputing.com) that theoretically allows us to use the parallel processors and RAM capabilities on the Prism to run MATLAB remotely on it. Anyways, MATLAB will supposedly be fully supported on the Intel Macs/Xserves by spring of next year (and a beta version sometime this fall). To answer your question, this software requires a lot of computation, and mostly 2D graphics, some 3D.
SCIRun can be built from source as 32-bit or 64-bit on Intel-based processors. This software is for bioelectric field modeling for my research. It requires a LOT of computational power, which can be spread across multiple processors. An example is that it takes about 10 days on our current Prism setup to run through solutions for a single patient. This uses up all 8 processors 24/7 and about 32GB of RAM. In fact, we eventually hit swap every 4th day and have to restart the computer and resume our calculations from where we paused. We use this software heavily for 3D graphics as well. We currently have a 4-monitor setup on the Prism and use it often. There is a lot of 3D rendering we do on it, as well as a lot of 3D simulations in real-time. Also, I know for a fact this software runs fine on the Intel Macs already.
So given those requirements, what do you guys suggest? Upgrading the current Prism machine by adding more bricks and RAM, or getting an XServe array cluster of some sort? Or perhaps even a Sun-Fire X4200 as Relic suggested?
Thanks in advance for your help!
If you really want to speed things up, just run your simulations in C. At one point during college, (computers were a little slower then) I had a Matlab simulation running for about a week. I realized that it might not be done by the time I needed the results, and wrote a C app in three days to do the same thing. Two days later, the C program was done with the simulation, and Matlab was still chugging with no end in sight.