IBM z14 Announced, Support Added To LLVM Clang

[kebabbert]

you got some benchmarks for this?

It's one of those things that once you have mainframe software its probably too expenseive to reengineer it to run on commodity hardware, but thought that they were decently powerful for the price you pay for them.

In the link that Dawn cited, there are links. And benchmarks. Read them.

Fact is, Linux on Mainframes and Linux on x86 runs the same source code. On Mainframes, the Linux software is many times slower. Who says that? A guy who ported Linux to Mainframes. He could compare the same software running on Mainframes and on x86. In the links, a consulting firm said that the z9 cpu is slower than a single core 900 MHz Xeon. Every IBM z cpu after that, has been just something like 50% faster than the z9. The z10 cpu is 50%, which corresponds to a single core 1.35 GHz Xeon. And the next cpu, the z196 is again 50% faster which means z196 is as fast as a Xeon single core cpu at 2 GHz. Or a dual core Xeon cpu at 1 GHz. And if you continue increasing performance 50% until you arrive the z14, you will see that z14 is something like 2-3 GHz dual core Xeon. But today Intel have 22 core 2.4GHz Xeons. Guess which is faster?

IBM is famous for posting FUD and false claims. And whenever IBM has a good result, they post it all over the internet. For instance IBM POWER7 was a good cpu and really fast, faster than anything else. IBM posted lot of benchmarks on their website and all over the internet was SPECcpu2006, TPC, SAP, etc benchmarks that really proved that POWER7 had the highest scores. And it did. But for Mainframes? No benchmarks. Nope. Nada. Zilch. No benchmarks nowhere. Why? If Mainframe cpus smokes everything on, say, SPECint2006 scores, IBM should have submitted results to SPEC organization. But there are no results. Nowhere.

I remember IBM claimed that the largest z10 Mainframe with 2TB RAM, could replace 10.000 x86 servers. That is really surprising, if you know that Mainframe cpus are dog slow. So I dug a bit. It turned out that all x86 servers were old and antique (256MB RAM P3) and they all idle at 1-2% cpu load. And the Z10 Mainframe was loaded at 100%. So it is probably true that one Z10 Mainframe can virtualize 10.000 x86 servers that all idle. However, what would happen if 10 of the x86 servers started to do real work? The Mainframe would choke to death trying to keep up.

[kebabbert]

Yeah. I'd like to see at least some comparison...

There are some in the links that Dawn cited. Read them. Fact is IBM Mainframes cpus are dog slow. There are no new customers, everyone are trying to migrate off them. The only one buying Mainframes, are old customers that upgrade. Because of lockin effects. Once you go Mainframes, you are stuck. And IBM sells something like 100-200 Mainframes each year, and these few Mainframes makes something like 10% of all IBM revenue. IBM charges noose bleeding prices for Mainframes. That is sick, when you consider that x86 is much faster.

For instance, there is no stock exhcange running on Mainframes. All are running on x86. If a stock exchange is faster then ever trader go there, because High Frequency Traders earns money on speed. If Mainframes were faster than x86, then every stock exchange in the world would switch to Mainframes, because there are so much money involved in stock exchanges. The price of hardware is irrelevant if a Stock exchange can increase their speed, because then every trader would go there. But no stock exchange runs Mainframes. Why? Is it because Mainframes are much faster than x86? Or is it because Mainframes are much slower?

[Dawn]

There are some in the links that Dawn cited. Read them. Fact is IBM Mainframes cpus are dog slow. There are no new customers, everyone are trying to migrate off them.

There are, actually. A few dozen per generation.

The only one buying Mainframes, are old customers that upgrade. Because of lockin effects. Once you go Mainframes, you are stuck.

For the most part, yep. The mainframe customer base is largely long-term customers.

And IBM sells something like 100-200 Mainframes each year

The number is vastly higher than this. Fujitsu, whose mainframe business is a tiny fraction of the size of IBM's, sells over 200 units per year in Japan alone, not counting their German business. And IBM outsells them in both Japan and Germany - and Fujitsu has basically no mainframe sales elsewhere.

IBM charges noose bleeding prices for Mainframes.

They do. Tens of thousands of dollars list price for activation of one core. The pricing is not favorable, unless a user can secure extremely deep discounts.

That is sick, when you consider that x86 is much faster.

Depends on what you're doing. For single-thread performance, and per-core performance (both of which are real metrics to those who have a job beyond spending years shilling Oracle) z is first-rate.

For instance, there is no stock exhcange running on Mainframes. All are running on x86.

First off, I know of at least one major stock exchange in a Western country running on z (EC12, last I checked.) Second, what does that say about your beloved SPARC if they're all x86 because x86 is the fastest?

[gilboa]

There are some in the links that Dawn cited. Read them. Fact is IBM Mainframes cpus are dog slow. There are no new customers, everyone are trying to migrate off them. The only one buying Mainframes, are old customers that upgrade. Because of lockin effects. Once you go Mainframes, you are stuck. And IBM sells something like 100-200 Mainframes each year, and these few Mainframes makes something like 10% of all IBM revenue. IBM charges noose bleeding prices for Mainframes. That is sick, when you consider that x86 is much faster.

For instance, there is no stock exhcange running on Mainframes. All are running on x86. If a stock exchange is faster then ever trader go there, because High Frequency Traders earns money on speed. If Mainframes were faster than x86, then every stock exchange in the world would switch to Mainframes, because there are so much money involved in stock exchanges. The price of hardware is irrelevant if a Stock exchange can increase their speed, because then every trader would go there. But no stock exchange runs Mainframes. Why? Is it because Mainframes are much faster than x86? Or is it because Mainframes are much slower?

I believe your sales figures are completely false.
AFAIR IBM's revenues from the Z13 greatly suppressed previous generations (some sources claim 130% increase in 2015/2016).

As for the discussion about CPU performance, IMHO is completely irrelevant. Yes, mainframe CPUs are slower, and no, nobody really cares.
Mainframes are not used to drive CPU bound application, thy are used to drive I/O bound applications that require uptimes that measured in years (5+ nines and above).

BTW, most of the companies I work with that use Z, usually use z/OS and not Linux.

[boxie]

In the link that Dawn cited, there are links. And benchmarks. Read them.

Fact is, Linux on Mainframes and Linux on x86 runs the same source code. On Mainframes, the Linux software is many times slower. Who says that? A guy who ported Linux to Mainframes. He could compare the same software running on Mainframes and on x86. In the links, a consulting firm said that the z9 cpu is slower than a single core 900 MHz Xeon. Every IBM z cpu after that, has been just something like 50% faster than the z9. The z10 cpu is 50%, which corresponds to a single core 1.35 GHz Xeon. And the next cpu, the z196 is again 50% faster which means z196 is as fast as a Xeon single core cpu at 2 GHz. Or a dual core Xeon cpu at 1 GHz. And if you continue increasing performance 50% until you arrive the z14, you will see that z14 is something like 2-3 GHz dual core Xeon. But today Intel have 22 core 2.4GHz Xeons. Guess which is faster?

IBM is famous for posting FUD and false claims. And whenever IBM has a good result, they post it all over the internet. For instance IBM POWER7 was a good cpu and really fast, faster than anything else. IBM posted lot of benchmarks on their website and all over the internet was SPECcpu2006, TPC, SAP, etc benchmarks that really proved that POWER7 had the highest scores. And it did. But for Mainframes? No benchmarks. Nope. Nada. Zilch. No benchmarks nowhere. Why? If Mainframe cpus smokes everything on, say, SPECint2006 scores, IBM should have submitted results to SPEC organization. But there are no results. Nowhere.

I remember IBM claimed that the largest z10 Mainframe with 2TB RAM, could replace 10.000 x86 servers. That is really surprising, if you know that Mainframe cpus are dog slow. So I dug a bit. It turned out that all x86 servers were old and antique (256MB RAM P3) and they all idle at 1-2% cpu load. And the Z10 Mainframe was loaded at 100%. So it is probably true that one Z10 Mainframe can virtualize 10.000 x86 servers that all idle. However, what would happen if 10 of the x86 servers started to do real work? The Mainframe would choke to death trying to keep up.

It sounds like you have an axe to grind and most of your evidence appears to be anecdotal at best.

The reality is, if there are people who are paying for these things, then IBM will continue to spend lots of engineering effort on making them. If people are buying them then they are not as bad as you make it out to be.

[bws111]

The problem with those 'benchmarks' is that they provide no useful information on exactly what they benchmarked. All they say is 'IBM z9'. Because mainframe software is priced according to the performance of the machine it is running on, there is a WIDE range of performance options available. The fastest z9 (2094-754) was about 750x faster than the slowest z9 (2096-A01), with hundreds of models to chose from. Even just comparing single-core performance, the fastest single core z9 (2094-701) is 25x faster than the slowest single core z9 (2096-A01), and there are dozens of performance options between those extremes.

The next question is: what enviroment were they in when they did their testing? Did they actually have full, exclusive use of a full-speed model 701? Since, once you add storage, you are talking close to $2M for such a machine that would be highly unlikely and a great waste of resources. Much more probable is that they had a development environment, which generally consists of a virtual machine in z/VM running in an LPAR on the mainframe. And the LPARs can be capped as to how much of the CPU they can use, so as to not interfere with production workload on the same box.

You would have to be nuts to think that a processor that is running at 5.2GHz, in which most of the instructions are implemented in hardware (not ucode), with a tremendous amount of cache, out-of-order execution, and superscalar design is going to be 'dog slow'. It isn't. It is faster than anything out there.

[kebabbert]

1) There are, actually. A few dozen per generation.

2) The number is vastly higher than this. Fujitsu, whose mainframe business is a tiny fraction of the size of IBM's, sells over 200 units per year in Japan alone, not counting their German business. And IBM outsells them in both Japan and Germany - and Fujitsu has basically no mainframe sales elsewhere.

3) Depends on what you're doing. For single-thread performance, and per-core performance (both of which are real metrics to those who have a job beyond spending years shilling Oracle) z is first-rate.

4) First off, I know of at least one major stock exchange in a Western country running on z (EC12, last I checked.) Second, what does that say about your beloved SPARC if they're all x86 because x86 is the fastest?

1) I doubt that. All articles I have read say everybody flees off from IBM Mainframes. Show us links there are new customers.

2) I read that IBM sells 100-200 Mainframes each year, that was an article in www.theregister.co.uk who covers Mainframes sometimes. I stored that number on my memory, thats it. I will google for the article again and see if I find it.

3) That is a lie. We all know that performance of Mainframe cpus are abysmal. If the Mainframe cpus were super fast - why hide it? Why not post benchmarks? "IBM does not post benchmarks that prove Mainframe cpus are superfast, because IBM does not want to hurt the feelings of competitors?". Or what? What do you think exactly when you keep defending Mainframe cpu performance? What are IBMs motives to not post benchmarks where Mainframes is 100x faster than Intel Xeon? If they were that fast, everyone would rush to Mainframe cpus. Does not IBM want new customers?

4) You are toast now. I have worked in one of the largest exchanges in the world, for many years. We know everything about the competitors. And there is NO exchange running on Mainframes. Large exchanges earns all money on trade volume. And you get volume if you are fast. Mainframes are dog slow, so no exchange would get traffic using Mainframes. And that exchange would go bankrupt. You are toast - I know the largest Exchanges very well. I want you to post links to an exchange that runs on Mainframes. You can not. This proves you are deliberately lying. So, "Dawn" is a liar and FUDer and can not be trusted. He is a paid IBM liar.

[kebabbert]

I believe your sales figures are completely false.
AFAIR IBM's revenues from the Z13 greatly suppressed previous generations (some sources claim 130% increase in 2015/2016).

I read these numbers from theregister.co.uk. I will try to find that article again.

As for the discussion about CPU performance, IMHO is completely irrelevant. Yes, mainframe CPUs are slower, and no, nobody really cares.

People here care and they call me a liar for insisting dispelling the FUD that Mainframe cpus are fast.

Mainframes are not used to drive CPU bound application, thy are used to drive I/O bound applications that require uptimes that measured in years (5+ nines and above).

This is true. This is exactly what I have been saying all the time. Mainframe cpus are dog slow - but they have good I/O. I am glad one person here agrees.

[kebabbert]

It sounds like you have an axe to grind and most of your evidence appears to be anecdotal at best.

Are you kidding? I have posted links to Mainframe experts. One ported Linux to Mainframes. Another wrote a Mainframe emulator. Do you call them "anecdotal"?

What do you call IBM claiming that Mainframe cpus are superfast - with no proof whatsoever to back it up with? No benchmarks, no links, no nothing. What do you call that? Anecdotal?

The reality is, if there are people who are paying for these things, then IBM will continue to spend lots of engineering effort on making them. If people are buying them then they are not as bad as you make it out to be.

I told you, people are bying them because they have no choice. Once you go Mainframe, you are stuck. You can not migrate off them. You need to rewrite all software to Java or something modern programming language that can be run on x86 - and that takes years. It is cheaper to stay on Mainframes and pay multi expensive fees and get slow cpu performance.

However, note that Mainframe I/O is good - I have never claimed the opposite. Truth is truth: Mainframe cpus are dog slow - no benchmarks vs x86 nowhere. Mainframe I/O is easily superior - they have many I/O coprocessors. Truth is truth. Dont lie.

[kebabbert]

You would have to be nuts to think that a processor that is running at 5.2GHz, in which most of the instructions are implemented in hardware (not ucode), with a tremendous amount of cache, out-of-order execution, and superscalar design is going to be 'dog slow'. It isn't. It is faster than anything out there.

If it is faster than anything out there, prove it. If you dont have any proofs, no benchmarks, no nothing - but still believe IBM FUD, then you must be nuts. It is like religious people, they believe in stuff without proof and gladly kill other people for their belief. That is called BELIEF. You show belief now. I am scientific, and I demand proof. If there are no proof, I doubt the claim. And you call me nuts? Science requiring proofs vs blind faith - who is nuts here?

Regarding the 5.2GHz, large cache, etc - well IBM POWER6 had also good specs, running at 5GHz, etc - and one 1.6 GHz SPARC T2+ was 10x faster on official SIEBEL v8 benchmarks. Also Intel Pentium 4 had good specs but was dog slow despite it's high clocks. So good specs does not prove anything - you need to look at the benchmarks to be sure. Are you blindly trusting AMD Vega specs, or do you want to see benchmarks before buying? Science with proofs, or blind faith?

And you call me nuts for not believing in IBM marketing?

https://en.wikipedia.org/wiki/Fear,_uncertainty_and_doubt#Definition

"FUD was first used with its common current technology-related meaning by Gene Amdahl in 1975, after he left IBM to found his own company, Amdahl Corp.: "FUD is the fear, uncertainty, and doubt that IBM sales people instill in the minds of potential customers who might be considering Amdahl products."