Is the mainframe just a high performance PC ?

From: "William Zhang" 
Newsgroups: bit.listserv.ibm-main
Sent: Monday, October 30, 2000 1:33 PM
Subject: Mainframe vs. high performance PC

I am a newbie using IBM Mainframe S/390.

I have used PC for many years, and I am familiar with Windows, UNIX. From magazines and newspaper, I know that high-end PCs, workstations are becoming powerful and powerful. Even a single PC or workstation's power is not so great, we can use many of them to build SMP, MPP, or Clusters to greatly increase the power. This let to think, can these machines compete with traditional mainframes?

In my opinion, mainframe's CPUs, DISKs, Type Drivers, I/O channels are like those counterparts used in PCs, SMPs, etc. Because the same techniques can be used. Maybe the major diference is the OS they use, UNIX is time-sharing, and OS/390 can do both batch-oriented and time- sharing.

In fact, there are so many aspects to be considered in comparing the two different kind of machines mentioned above. Can anyone tell me what's the actual, major difference between mainframes and those more-popular machines? And the most difficult question for me is, are mainframes more POWERFUL than the counterparts really (without thinking of the price)?

From: "Doug Nadel" 
Newsgroups: bit.listserv.ibm-main
Sent: Monday, October 30, 2000 5:49 PM
Subject: Re: Mainframe vs. high performance PC

You might want to search the archives since this has come up a couple of times, but it comes down to this. In terms of raw CPU speed, there is probably not a difference anymore. Both CPU technologies are essentially huge computers on a chip or a small set of chips. The differences lie mainly in reliability and I/O capabilities. Mainframes have a lot of redundency and servicability features built in. The way they handle I/O tends to be much more efficient than the 'lower end' boxes since they are able to handle many (hundreds? I don't know) of simultaneous I/Os. Also the S/390's means of doing paging I/O is very, very fast. And it depends on who you talk to, but in recent years, mainframes have been cheaper to operate than large farms of networked PCs when you factor in support costs (personel, infrastructure, backups, etc). The goal of decentralization has turned into a nightmare of cost and managment issues in some cases.

I'd add that mainframe software (at least system software) tends to be written with the old ideas of efficiency in mind. Yes, it has gotten more bloated than it was, but it is nothing compared to the bloatware on PCs. You know, you install a little clock program on a PC these days and you see things like 8MB downloads. If someone in a mainframe environment shipped a dumb utility like that and it took up 8MB, even including doc, runtimes, source, and their mother's favorite fruitcake recipe, they'd be cast out into the land of the burger flippers. But that is a software issue, only related to your question in a more cultural sense.

A more interesting question is what is a mainframe. S/390 and its ilk, probably Alpha and similar high-speed multi-terabyte capable data warehouses, but it basically comes down to I/O: how much and how fast.

Doug Nadel

From: "John S. Giltner, Jr." 
Newsgroups: bit.listserv.ibm-main
Sent: Tuesday, October 31, 2000 4:51 AM
Subject: Re: Mainframe vs. high performance PC

Performance is relative and depends on what you want to do. The UNIX and PC worlds are now looking at implementing the same type of IO structure (channels) as mainframes have used for years. This is because of the IO capacity the mainframe has. The just announce zSystem has increased this capacity even more. The new zSystem can support 24 Gigabit Ethernet connections or 36 100 MB FICON connections at wire speed. There are very few UNIX or Intel boxes that can even get 24 Gigabit Ethernet connections and those that can normally only have 4-8 independent IO busses and therefor every 3 - 6 cards are sharing the same IO bus and the buses are only running at just over 1 Gigabit. If you have 3 Gigabit cards sharing the same gigabit bus, you are not going to be able to drive all 3 cards to a gigabit. I have not seen any test recently, but about a year ago the fastest Intel box could only get about 800 mbps.

There was a company in northern Virginia that took a G5 box, created a LPAR with two processors and 256 MB memory. They ran VM in that LPAR and then ran Linux under VM. They ran Apache Web Server on the Linux systems. They kept starting more and more Linux system until they could not do anything else. They got 40,000 (yes 40,000) Linux systems up and running on a G5 2 way with 256 MB memory. Now is it practical to run 40,000 images, well no because this is where the system stopped. However IBM does state that you can run thousands of production Linux systems. In fact in one of their ads for the new zSystem they quote a price of $500 per Linux image, this is based on 2,500 images. This cost only includes the zSystem (CPU, Memory and IO cards), does not include DASD or any other external equipment.

When you look at the price you need to look at a lot of thing. Number of users, amount of work done by box, number of support personal per user. A single NT box is cheaper that a single mainframe, however you may need a lot of NT boxes to do the same amount of work as a single mainframe. IBM and Compaq has recent set some records with their Intel servers running Windows NT. I believe the Compaq solution cost about $12 million and the IBM cost $14 million. Not so cheap.

Now is the mainframe right for everybody, no. UNIX and NT have their places. Heck, even Microsoft knows that NT can't do it all. They use Unix for some of their internal systems and for some of their services. Microsoft runs Hot Mail, Hot Mail runs on Unix, not NT.

The other thing that mainframe have had traditionally is hardware and software reliability. As Intel and UNIX boxes have some of the same hardware monitoring functions built into them, they are becoming more and more expensive.

From: "Joseph Katnic" 
Newsgroups: bit.listserv.ibm-main
Sent: Tuesday, October 31, 2000 3:20 PM
Subject: Re: Mainframe vs. high performance PC

Actually, as you may have seen from various posts, the CPU is not the issue in a mainframe it's the IO.

I go one step further. PARALLEL SYSPLEX.

What's that you say? Well it's a way of writing SERIAL code and having it work PARALLEL across multiple mainframes WITHOUT THE PROGRAMMER HAVING TO KNOW ABOUT IT.

It's miles/kilometers ahead of anything in the non-mainframe world. It takes the mainframe to a new level that the others can't match at this time, and probably never will. This is because it requires control of the middleware that is not available on the PC - due to too many different suppliers and no agreement on how to achieve the result.

Joe Katnic

Date:         Sat, 2 Dec 2000 11:11:46 +0800
Reply-To:     Victor Zhang 
From:         Victor Zhang 
Subject:      MIPS versus MHz

Hello mailing list,

Is there any method to compare MIPS used in mainframe world and MHz or GHz in PC or Unix world?

--Best regards, Victor mailto:zsl263@263.net

Date:         Sat, 2 Dec 2000 12:32:49 -0500
Reply-To:     IBM Mainframe Discussion List 
Sender:       IBM Mainframe Discussion List 
From:         Doug Nadel 
Subject:      Re: MIPS versus MHz
 

Not really, primarily because both are useless numbers. It would be like comparing rotten apples with decaying oranges (so to speak).

MIPS measures instructions (opcodes) per second which is a pointless thing to do because what you really want to know is which machine will get the job done the fastest and that can't be determined by MIPS because of the variations in instruction types (amount of work done by each instruction, CISC vs RISC), instruction mix, memory reference patterns, cache behaviour, I/O waits, multiprocessor dispatching overhead, etc, etc. MIPS is a measure that seems easy to grasp but it doesn't really measure anything useful except in the broadest sense (a G5 has a higher MIPS rating than a 370/125 so the G5 must be faster). It does not measure throughput.

MHz measures cycles per second, but in addition to all of the problems with measuring something in MIPS, you also have differences in architecture, pipeline lengths, pipeline invalidations, prefetch differences, branch prediction, microcode levels, internal data movement, external data movement, components on the die (eg internal vs external cache), etc. Again, it is useful in broad terms (a 1.5 GHz Athlon is faster than a 4.77Mhz 8086) but useless in narrow terms. For example, clock tick for clock tick, an Athlon is much faster than a Pentium III - a 700Mhz Athlon is faster than a 750MHz PIII at most tasks). Again, it doesn't measure throughput.

It is tempting though to think about it. I suspect that a lean operating system and lean code on a modern PC could do most of the basic, non I/O intensive work most companies need. But then again, when is the last time you saw lean code (operating system or otherwise) on a PC?

Doug Nadel , ISPF and OS/390 Tools & Toys page: http://somebody.home.mindspring.com/

From: "Bill Ogden" 
Sent: Sunday, December 03, 2000 10:39 PM
Subject: Re: 360/370 instruction cycle time

Instruction cycle times made sense on 1960s and earlier 1970s machines. They really make no sense now because the interactions between cache effects (at multiple levels), parallel/lookahead execution, various addressing modes, TLB overhead, etc, produces a very wide potential variation in the times for a given instruction.  Taking a single instruction, in isolation, and timing its steps through its execution would produce a misleading result.  Even for RISC machines, the instruction cycle times are only distantly related to the power of the machine due to all the parallel work being done.

Many of us somewhat miss the simplicity of simple, predictible instruction timing on the older machines --- but that time is long gone.

Bill Ogden