NT vs. Unix

SpecFP/$

I originally wrote this while working as an engineer in the Ford Business Group at Arvin Exhaust. I had taken up being a finite element analyst, but I wasn’t allowed to pick what platform upon which I would do my analysis. A coworker was given that task. I would have chosen some Unix-based workstation setup. He selected Windows NT. He assembled the first real “workstation” in our group, a dual Pentium 100. It was to be his computer, but he had pity on me, since I was wandering aimlessly on the job at the time, and let me have it. It was on this computer I started trying out various bits of software.

I tried COSMOS, Pro/E, and NASTRAN for Windows. I wanted to buy the real NASTRAN, and use Pro/E as a front end, but Ford Motor Company was working a huge multi-million dollar deal with SDRC at the time, and our boss’s boss decided that we should follow suit, so IDEAS was selected. They had a passable linear solver, but, then, everyone else did too. On the other hand, they did indeed have a great modeler. Because Windows was such a wannabe up-and-comer at the time, I  benchmarked my rig against some Unix workstations in our Advanced Engineering department, where they also had a seat of IDEAS. I even went so far as to get a demo Indigo II from SGI on site with which to experiment.

The dual Pentium wasn’t even close.

So we upgraded, this time to a dual Pentium Pro 150 with 144 MB RAM. Now we were talking! With the much better SPECfp scores of the Pro series chips, I thought I could close the performance gap. However, my buddy in Advanced got a new UltraSPARC along the way too, so the comparison needed rerun.

Better, but still a long way off.

From: “David Krider” <dunkirk@slip.net>
To: “ICCON Misc Group” <iccon-misc@kosh.sdrc.com>
Date: Thu, 26 Sep 1996 15:13:16 -0000

The results are finally in. I have looked for a long time, and this is as close to an apples to apples test as I have yet seen. I used Master Series 3.0 running on Windows NT 4.0 and SunOS 2.5. I simply used the p-element analysis example problem in the Exploring I-DEAS Simulation book, Volume 1. The solution times were 308 seconds on the PC versus 195 seconds on the Sun, a 60% faster score. The price of the Sun, however, is 400% percent more. (If you read nothing else, check my final statement about what this is letter is all about.)

The PC is an HP brand dual 150 MHz Pentium Pro with 144 MB RAM, 4 GB disk space (on a SCSI bus), 4 MB Matrox Millenium video card, 4X CDROM and a 17″ monitor. This goes for about $5,000. Add another $2500 for one of the new AccelPro 3D accelerators, and the total comes to $7,500.

(I have an AG500, which — according to Accel — runs a little faster and has a couple more features than the AccelPro, but it cost me $3400, and is no longer being made. The AG500 is net yet running under NT 4.0; AccelGraphics says end of October for the new drivers.)

The Sun is an UltraSPARC 170E with 196 MB RAM, 4 GB disk space, and Creator 3D graphics. Adding a floppy drive and not taking into account maintenance, this one runs $30,000 (with discount for previous sales.)

Please note in this review that I have rounded off the prices. I don’t claim accuracy to 3 decimal places. Also note that I am not discussing graphics. While I have no complaints about the 3D performance of my machine, I realize that the Sun blows mine away. I know UNIX hardware will perform much better in this arena. But hey, you’re paying for it. I have found that after 100,000 DOF’s in a simulation model, you don’t want to run it on either platform tested here. You tend to look for ways to simplify the model. And my computer can handle the graphics of such a model nicely. So I don’t see this as a big deal in regards to FEA. Building assemblies and drafting them is another issue.

When we first went with dual PC’s, the guy who spec’d them was convinced staying on an Intel platform running Windows NT was the way to go. His reasoning was that the performance benefit of running on a UNIX box was negated by the difficulty in running the operating system and the lack of “everyday” applications for it (read: Office 95.) Now someone will gripe that UNIX isn’t that bad and that you can run a Windows emulator on it. Well, tell that to two of the three MSME’s here doing FEA that don’t know squat about UNIX (and don’t _want_ to know squat about it) and all of whom have a PC sitting on their desks. Mind you this is being written by someone who wanted a UNIX box to start out with. I do know a bit about UNIX, but I have grown to think that NT is way ahead of it in many respects. (Flames on this point please direct to everyone@has.their.own.com.)

What has really surprised me is that SDRC really is showing an attitude of platform independance. They really do — as a company — think of Wintel as a viable solution. I confess that I expected a different story when I went to the HQ for training. While I am unimpressed with their overall knowledge of the OS, I am quite impressed with their commitment to making a good solution to run on it. (I have on many occasions solved questions regarding I-DEAS interaction with the OS myself when support didn’t have an answer. And don’t even get me started about the new 3.0 install program on NT…)

I wrote this in order to show the UNIX zealots of the world and the prospective corporate specifier that SDRC on NT/PPro is a good solution for running SDRC. There is nothing to be intimidated about when running linear static and dynamic structural analyses on Wintel. I’m _not_ saying that it is the empirical best solution. I _am_ saying that it is better solution bang for buck.

Regards,
dk

David Krider
Product Engineer, Arvin North American Automotive
dunkirk@slip.net or 102227.2300@compuserve.com
For in Him we live, and move, and have our being. (Acts 17:28)

So that was my summary of the situation. As an epilog, I didn’t get much response, but then I had taken so much time preparing and writing it that it didn’t leave much room for argument.

Not long after this, my coworker talked the big boss into a fantasy that all engineers in our group could do FEA at their desks, if they only had the computer with which to do it. You know, because it was so easy; just point and click! So we bought dual Pentium Pro’s for everyone, and I used some parts from the next shipment to get to a dual Pro 200 with 384 MB RAM.

Again, better, but no match in terms of raw performance. However, the price per performance ratio was still heavily in my favor.

Unix Rules

Parenthetically, another guy was hired in the Chrysler Business Group to do FEA during this time. He bought a high-end SGI workstation (a short-lived R8000-based model with stratospheric SPECfp scores) upon which he ran the non-linear FEA package MARC. (I still strongly disagreed with that direction for Arvin’s needs, but that whole business is long gone now. After he left, his computer got transferred to Advanced where I traded it in on a new HP workstation.) Anyway, this being a time before our T1 to the internet, he hired a local ISP to come in and wire up his workstation with an always-on modem connection. When the consultant came in, I started a conversation with him about the whole NT versus Unix thing. I was “into” NT, to be sure. It was the “new hotness” at the time. He was into Unix, and felt the pressure from Microsoft. We had a lively debate about the merits of each.

He told me how NT was so slow, and so lame, and so insecure. He said he was so good with computers that he could hack our network just by using some listening devices from outside the building and recording every keystroke we made. Yeah, whatever. After about a half hour of fiddling, trying to get this R8000 connected over the modem, I asked him what the hold up was. I told him that it took me all of five minutes to configure my NT box to dial up Sprynet. He told me that there were over 200 options he could set in the UNIX dialer program, and that he was just setting everything for optimum performance. And after this, he said, “Unix rules.”

Well, an hour later, he left to go back to his office and configure some things on his side. A week later, and he finally had it working, but every time the modem got connected, it would crash the entire workstation. It was some months later that they finally got a motherboard firmware update that resolved the problem.

Now you might be thinking that that should be enough for me to have a good laugh at this jerk’s expense and move on, but that’s not the end of it. See, a couple months after they got it working, and after we got our permanent connection to the internet (which made this whole exercise moot), we found yet another problem with the way this “expert” had configured that poor machine. It would advertise itself to our internal network as a route to the internet. The real routers would pick up this advertisement and tell the PC’s on the LAN that they could get through the workstation to the internet. Since the workstation was on the local network segment, it would be selected over the real internet connection, which was a couple more hops away. Besides the severe difference in bandwidth between a modem and a T1, the real problem was that when the engineer would shut off his modem, every computer that had the workstation as their default route would fail to get to the internet, leading to some very weird support calls to the help desk.

Choosing to advertise your dial-up connections to the rest of the network was a simple check box in Windows NT 4.0…

As the use of NT for doing this sort of thing was still basically new, there was a lot of debate on the IDEAS users’ email list about it. Many people were against it for various reasons, all of which were subjective and relative. So I took the side for NT as an analysis platform, almost to play the devil’s advocate, even though I still would have liked to have been using a real workstation.

The market now

I originally wrote the email back in 1996, and wrote the article not terribly long after that. The entire landscape has changed dramatically since then. Given what I was seeing, I predicted this, but I would have expected it to take much less time. As of 2009, the world of CAD/CAM and FEA is basically all Windows. SDRC is going to go away; CATIA is taking over the world, and their platform of choice is Windows. I think it’s safe to extrapolate that if this juggernaut has finally thrown in the towel on Unix, everyone else — if they haven’t already — will soon follow suit. The solver side is another matter, for a lot of this was going to Linux. Around 2004, I submitted a proposal around for buying a “compute farm” of Xeon boxes running Linux. Almost every one of the solvers we used had a Linux version, though few had front ends. I suspect that situation has changed now too.