In a CFD conference this year, a young professor from the University Heideberg/Germany hold a presentation about the benchmarking of different hardware platforms for CFD simulation. His conclusion is that if a PC has larger cache it will work faster than a workstation and hence one should not forget to use PC for doing the CFD simulation. How are your opinoins and experience about this topic?

X. Ye

What pc and what workstations were compared ?

Not only speed is an important issue in cfd but also stability. Thats why many people try to avoid using NT for this kind of work.

It is also a problem to put more than 2 or 4 processors into a pc or to achive a good speedup because all processors share the same bus.

So for low end work a pc might be ok but for larger problems (several million cells or moving mesh) also a multiprocessor workstation might be to slow for having real fun. Then one has to use vector supercomputers.

He compared a wide range of PC's and workstaions. We should constrain that only single processor will be used and the RAM of PC is sufficient large to save the large amount of cells. The tact frequence of a PC processor is nowadays very high. The message of this young professor is:

1. Not forget to use PC for CFD simulation (PC is cheap); 2. One should combine software and hardware well

X. Ye

This just means that when I put a large engine into a VW-Golf it might be as fast as a Mercedes.

So if you don't have enough money you can by a VW but it is not that comfortable, safe ...

Now it is up to you to decide.

I also doubt that a high-end pc will beat a new workstation in single processor performance. There will be a factor of 2 to 4

I am currently doing a little tour of some of the parallel computing sites on the net. For some time now I have toyed with the idea of running a parallel version of say my own CFD code (if I had one) on several PCs operated under a free operating system such as Linux. Surely this is not an original idea and if not what am I missing? I do not see anybody else doing something of this nature.

I am currently doing research on cooling towers using a PC-based code on grid sizes of up to 250 000 elements. As a result I have time to think of "cheaper" ways of going faster!!

If you have to take the normal costs of an engineer into account (approx. $1000 per day) then you can decide if it cheaper to buy a ready to use solution or to build something by your own.

On this way a pc might become very expensive at the end.

This is very crucial conclusion. How can I get the paper of this study? I use powerful PC (2 Pentium II CPU, 1 GB RAM,512 Kb Cache, 450 Mhz ) installed with NT. As far as I know cache memories of PC's can be extended to 2GB. At the near future PC with 64 bit adress bus (i hope so) will work faster than today's 32 bit one.

I have no problem with Windows NT.

Nuray

the PC is more cost effectively for number crunching than most workstations. for a given peak speed disk space and RAM the PC is often more than five times cheaper than a workstation. the big drawback on PC is the windows OS. but you can use Linux or PC versions of SunOS or IRIX (i think) which are good if you have precompiled software. there are other operating systems for the PC most of which are better suited to high performance computing than windows. also as Chris said you can link PC's to form powerful parallel computers running Linux at low cost. at penn state we have one here in the aerospace dept (see cocoa.ihpca.psu.edu) which is a 50 (25 2 processor pcs) processor machine which cost about $100000 to put together and ther performance is on the same order of magnitude as a 32 processor IBM SP2 (see performance and description at http://bart.ihpca.psu.edu/thesis/). i read a page where a group of aero engineers at sandia (i think) put together a 100+ processor cluster over a weekend. these types of computer solutions are representative of what can be done if we escape the workstation/production supercomputer paradigm. just because you pay more doesn't mean it works better. the volkswagen- mercedes comparison Joern made is not applicable because a cheap Flop is as good as an expensive flop

It is possible to make a supercomputer by connecting many PC's. Connecting of say eight PC's (distributed memory)or one PC with eight processor(shared memory), which one is attractive? Parallel processsing or multithreading ? These are related to hardware part of parallel computing. It looks very promising. But let's consider also modelling side of parallel computing. There are crucial problems for example, load imbalance over the PC's. Actually this is another discussion point.

nuray

it is not possible to say which is better it depends on your application and/or algorithm whether shared memmory or distributed memory is better. however it is easier i think to make distributed memory computers and cheaper too. to make any parallel computer perform well the communications between the memory and the processors has to be fast. in shared memory computers the need is even more critical since the communications between the cpus and the memory is larger. that is why you don't see too many homemade shared memory parallel computers because proprietary equipment (like what you see on SGI Origins) is needed to speed up the communications. on disributed memory machines, like the beowulf clusters i referred to in my last post to this thread, the inter-processor communications is less (usually)so communications speeds are less critical so they make do with fast ethernets. for CFD where you have flexibility to choose algorithms most people use explicit codes on these clusters and take advantage of the higher bang for the buck (see the thesis i referred to in my last message) since explicit codes work better on these machines. in other applications (like a friend of mine who does molecular dynamics, which is not parallelisable) shared memory or even vector machines become more efficient

Hi,

PC's are really best value for money. If you compare the difference in performance between an UNIX workstation and a PC, this becomes quite evident. For instance:

The latest Alpha workstation (Compac XP1000, which is the fastest simple processor machine available) produces an impressive SPECfp95 score of 52.4. Far ahead of the floating point scores of HP's 367MHz PA-8500 (28.1 SPECfp95), SGI's 250MHz R10000 (26.6 SPECfp95), and Sun's 360MHz UltraSPARC II (23.5 SPECfp95). While a 450MHz Pentium II Xeon CPU, scored a mere 15.2 on SPECfp95 (remember that Intel Xeon are already at 600 MHz, and the Xeon scales vary well, due to its large L2 cache). It is important to remember that these are the high specs systems of each manufacturer. If we look to the price, the Compac XP1000 cost 'only' US$ 13,000.00, while you could by a high spec Xeon PC for 3,000 or less. Therefore, I think we can say: 1 - If you are not going for the very high and is trying to go for a entry level machine, you should stay with a PC. Since the price for floating point operation is highly favorable. 2 - If you want really high performance and have cash to spare (which is not always the possible), you should go for a Alpha or SGI system.

So depending on the amount of money you have, there is litle to think about. Isn't it ?

You can check the performance of the entry level UNIX systems, they are all rated at about 10 SPECfp95 or less (except Alpha).

Cheers

Neyval

(1). I think, the use of compter depends on several factors. (2). If you are a consultant ( or a person interseted in CFD), you are likely to use a PC. This is especially true for IBM compatible, because of the cheaper cost of the hardware and software. (3). If you are a small business, the PC approach is also very attractive, because you can link other business software together easily on PC systems. So, PC is a cost effective solution with everything-in-one. (4). For larger organizations, especially those have been around for a while, the change over to PC system is going to take a while. This is because their software used to be on the main frame computers or super-computers. It is not easy to port the software to a new platform. (5). In 90's, most have ported the software to workstations in the design and analysis groups. But on the business side, not everything is on workstation. So, for large companies, they have to consider more factors. (6). I have seen companies with MAC, PC, workstation on engineer's desk at the same time. They also have to access the business software through workstation on the minicomputers by way of network. (7). In terms of performance, we can say that the workstation is a factor of two faster than the PC. The difference is not large. I normally upgrade my PC when there is a factor of 4 to 5 faster on the new system. This is also true at the workstation level. It is fairly common to upgrade the company's workstation after 3 years of usage. Many workstations of 4 years old are still in use today ( which are definitely much slower than the new PC). (8). The general guideline is to write the software as portable as possible. In this way, one can easily move the software to the new computer systems. (If you are using a commercial CFD software, then you are likely to buy the particular brand of computer for the particular software. This is also a common practice.)

The decision as to what machine to use really depends on the code's performance on a machine, and, obviously, budget. The only way to decide is to benchmark.

Results of recent benchmarks we performed tended to say that STAR-CD more or less scales with SPECfp and that Fluent and CFX don't (see http://www.fluent.com/software/fl5bench/ for fluent benchmark data). This difference is probably due to the language used, numerical methods and data structures used, attention paid to optimisation, etc. It does not say that one is necessarily better than another merely that they make different demands of a machine/compiler system.

It does however say that knowing the performance of one code on a machine is not necessarily a good indicator of how another, apparently technologically similar, code will perform.

As ever it points to the need to test your specific applications on a particular machine before buying.

Hi, That a very good observation. CFD is not just equations, numerical method and solver. But nowadays more data structures, implementations and what compiler.

I have compared CFD v. 4.2 with a LES impemenation and my own code. the difference is very large. One the same computer (PII Xeon 512Mb) I solve 10^6 points with LES (with dynamic subgrid model) in less than 60 sec. per timestep.(Geometric Multigrid)

If one compared the difference sparse data format in for example the SPARSKIT by Prof. Y. Saad one will get numbers from 5 Mflops to 29 Mflops (still PII Xeon 450 Mhz). (Linpack is above 100 Mflops)

When using structured grids in some way, the matrices are well order and banded, exploiting this will get you above 55 Mflops.

Once should look at U Rüde's work is this case !

Regards

jens