[syhprum (OP)]

When I first started working with the Siemens R30/R10 (1972 design IBM360 clone ?) which had the performance of the mini MAC they called it a mainframe, how are the different classes of computer defined.   

[Dr.IC]

it differs with respect to the number of process it has and the data handling capability too plays major role. Servers have less data handling capability compared to main frames, usually ATM machines are monitored by mainframes. servers are used has dedicated system for networking purpose, but PC is not like that..

[LeeE]

Yeah, it's a difficult difference to define.  In my experience, mainframes have been characterised by their data throughput; they aren't necessarily fast computational systems but they can shift a lot of data around, from a lot of different sources, to a lot of different destinations.  On-line teleprocessing (TP) systems, like an airline booking system, would be an example; bookings linking available seats on available flights can be made from many booking agencies around the country/world simultaneously and with little delay.  A PC workstation, on the other hand, might be capable of running graphics very quickly but would struggle to support several thousand TP clients.

[DoctorBeaver]

I agree. The differences are not clear-cut. But, as has been stated, the amount of different data sources and output devices the machine can handle is a factor.

LeeE mentioned a flight booking system. I actually worked on 1 of those on an IBM 360 mainframe (AAARRRRGH - it was written in COBOL!). That was in my early days as a programmer and PCs or servers hadn't been thought of back then. All that were around were mainframes and midis. I was told then that the major difference was in the number of I/O devices the machine could handle. We had over 100 terminals in about 20 locations around the country all sending data to the mainframe.

I later worked on a banking system that connected to over 500 terminals.

When I worked on a Data General Nova midi, that could only handle about 20 terminals.

I would, though, say that the number of instructions per second that a computer can perform is equally as important. The average desktop PC can perform about 100 million instructions per second. That may sound impressive but the IBM BlueGene/L can perform 360 trillion operations per second. That is officially the world's fastest computer. I think it would be pretty obvious which of those would be classed as a mainframe.

Incidentally, there is a computer, the RIKEN MDGrape-3, that is almost 3 times as fast as the BlueGene/L at 1 quadrillion operations per second. However, that computer cannot run the officical software that is used to rate speed so that is why the BlueGene/L is officially the fastest (I wonder who wrote that test software - IBM. maybe?)

But neither of those can match the human brain. That has been estimated to perform 10 quadrillion instructions per second.

[LeeE]

Hmm... the likes of the IBM BlueGene/L systems really fall under the 'super-computer' banner, rather than being a Mainframe, imho.  Although there are still a few super-computers that use custom hardware, most of them now use massive arrays of commodity x86(64) or IBM Power cpus, and as such are closer to PCs than mainframes, in many cases actually being clusters of PCs (or even PS3s in one system).

Once upon a time though, there was a person named Seymour Cray...

[DoctorBeaver]

Interesting story about Mr Cray. He was talking to Steve Wasniak of Apple fame. Wasniak told him that the new Apple PC had been designed on their Cray supercomputer. Seymour Cray replied that the next Cray supercomputer was being designed on an Apple Macintosh!

I take your point about arrays and that raises another point - do arrays class as a single computer? Think of all the PCs at the LHC handling all that data. There are hundreds of them all linked together and controlled by custom hardware. What, really, is the difference between multiple CPUs and an array of linked PCs with an uber controller? Both are distributed processing but just on a different size scale.

[LeeE]

I think that the difference between a number of interworking systems and an array or cluster is that each of the individual systems in the interworking collection performs a specialised role whereas the individual systems in a cluster all work on a single task.

If you look at something like the LHC, or indeed any large business or organisation where many computers are linked together, you could view the organisation-wide collection of systems as a single distributed system, the single role of which, is to run the organisation.  Within that distributed system though, there are specialised sub-systems devoted to things as diverse as controlling the building's environmental systems, such as heating, hot water and air conditioning to the PCs used in the legal department to layout and print their contracts.

In a cluster though, each system is identical in terms of s/w functionality, if not in terms of processing power, and each individual system is generally doing exactly the same processing as each other, just with different data values but not different data types. 

You can sometimes partition clusters, to split the collection in to sub-collections of systems, which can then run as several individual clusters.

[DoctorBeaver]

LeeE - What you have said is describing the uses to which the different systems are commonly put, not the systems themselves.

My Sony PC has 2 processors (not dual core). Under normal circumstances what each of those does at any given time is controlled by system software and would, commonly, be assigned on a prioritised, timesliced basis. I can, though, assign tasks and priorities to each manually. There is no reasons why I can't have 2 instances of the same program running simultaneously, 1 on each processor (the only problem could be I/O conflicts).

The same could be done on a true distributed system. The uses that each component is put to does not describe the functionality of the system.

[LeeE]

The use to which a system is put is the defining factor, not the hardware (as long as it's s/w compatible)  A collection of computer hardware, correctly connected and configured, does nothing without the software - it's inert.

To an extent, what you say regarding running multiple copies of a program on multiple cpus has been done, for example, with Seti and Folding at Home.  However, although those schemes all used a single set of data, it was parceled out and distributed by their server system.  You, on the other hand, have no way of making your two instances of your program work on a single set of your data unless you go in and split it.

I don't think it's really the same at all.

[DoctorBeaver]

On reflection, I think we're saying the same thing but looking at it from 2 different angles (our own mini M-theory!)

I was referring more to the potential uses to which different hardware configurations can be put and you were talking about the uses themselves. I don't see that it makes a lot of difference whether the CPUs are in the same box, in different boxes at the same location, or scattered across the world as with FAH & SETI. Some hardware configurations do not have the potential to be put to certain uses (or they would be extremely inefficient).

It is the controlling software that would try to make use of the resources available. In the case of my Sony the system software could be written so that 2 instances of the same program could work on the same set of data simultaneously. A single processor system could not do that.

[yor_on]

DB, wouldn't you need to have separate buses all over then, and perhaps two graphics processors etc? To get that true 'parallel' computing? Do parallel computers exist anywhere? Isn't all computing (digital) serial to its nature. I love the idea of beowulf systems, If I ever get that money I would like to try it with some ps3:s just to see it working. Didn't the Russians use this concept on 'ordinary' PC:s for their space program?

[DoctorBeaver]

In a distributed system the data can only be passed to 1 CPU at a time anyway as the controller would be linear.

As for multiple graphics cards, the PC array at the LHC doesn't have any graphical capability at all. It is only when that data is sent to a central hub that it can be looked at. The PCs in the array merely check the data that's coming in and decide whether to send it on or discard it.

Most computing is serial, but not all. There are parallel systems in use (neural networks for instance) but they are only a tiny proportion of the total number.

London Underground uses neural network computers to monitor how many passengers are on a platform and to close the gates if it is getting too crowded. That's the only 1 I know of that is in commercial use rather than at a university or other research establishment.

I know nothing about the computers used for the Russian space program. It wouldn't surprise me if they used abaci.

[DoctorBeaver]

A Beowulf system? Which part would be Thrunting?

[LeeE]

It is the controlling software that would try to make use of the resources available. In the case of my Sony the system software could be written so that 2 instances of the same program could work on the same set of data simultaneously. A single processor system could not do that.

When you say that the system software could be re-written to allow two instances of the same software to work on the same data set, are you talking about the Operating System software or the Application software?  If you're talking about the Application s/w then I agree but if you're talking about the Operating System s/w I disagree.

Each individual cpu core, since the Pentium 4, has been able to do a degree of parallel processing, but once again, the application software generally needs to be written, or at least compiled, to make effective use of it.

[DoctorBeaver]

In many specialised or embedded systems the OS and the application software amount to the same thing.

[LeeE]

Hmm... I'm not sure about "many" specialised or embedded systems, at least not nowadays.  Many simple electronic devices would need hard-coded logic to work e.g. T.V. remote controls, or multi-meters etc. but you couldn't really view their logic as software and they certainly weren't mainframe/server class systems.

The only systems that I can think of where the OS and the application really amouted to the same thing were the PICK and AS400 database OS systems.  These days though, the applications and OS are clearly delimited, permitting the application to be cross-platform (OS).

[DoctorBeaver]

Hmm... I'm not sure about "many" specialised or embedded systems, at least not nowadays.

You've not worked in defense?

I will concede, though, that those systems would probably not be classed as mainframes or servers.

[LeeE]

Good point - I've not worked in defense but you've reminded me of the typical LRUs you find in military stuff.  Afaik, these do operate as you suggest, with the application running directly on the hardware and without an intervening OS layer.

[yor_on]

A Beowulf system? Which part would be Thrunting?

Thrunting?
Don't know what you mean DB?
I was going to try it some years ago, but I had only PC:s of different cpu:s HD etc, so it became a to complicated project to play with.

But there is some that use PS3:s and gets good results. http://www.ps3cluster.umassd.edu/

[DoctorBeaver]

Thrunting was the name of Beowulf's sword. I thought everyone knew that!