IBM z14 Announced, Support Added To LLVM Clang

[Dawn]

If we ask *really* nicely can you try?

You would have a marketing case, since lots of these things run some form of Linux and I am sure you would have some clients like universities that might use your phoronix-test-suite

so - pretty please? (it would be pretty awesome)

In general, IBM is sensitive about public benchmarking of z. Power is fine, but not z. Customers who are interested can, of course, ask for benchmarks directly from IBM, or run them themselves - but that's different from random websites running benchmarks fully publicly.

Sure, it's paranoid, but it's IBM.

I can say, though, based on my own z experience: It's fast.

[boxie]

In general, IBM is sensitive about public benchmarking of z. Power is fine, but not z. Customers who are interested can, of course, ask for benchmarks directly from IBM, or run them themselves - but that's different from random websites running benchmarks fully publicly.

Sure, it's paranoid, but it's IBM.

I can say, though, based on my own z experience: It's fast.

naw - i wanna see how much it spanks desktop systems by :'(

[kebabbert]

If we ask *really* nicely can you try?

You would have a marketing case, since lots of these things run some form of Linux and I am sure you would have some clients like universities that might use your phoronix-test-suite

so - pretty please? (it would be pretty awesome)

Actually, ibm z14 cpus are much slower than a decent intel xeon. You can emulate a mid sized ibm mainframe on x86 using Turbohercules. A high end 8-socket x86 server would beat the shit out of the largest z14 server. This why ibm never publishes benchmarks of mainframes vs x86. Ibm always publishes benchmarks of Power cpus vs x86 cpus, because Power is not shit. But you will never ever find benchmarks of mainframe vs x86. Why? Because mainframe cpus are dog slow. Mainframe have good I/O, but the cpus are shit slow. Your x86 is much faster than z14.

[boxie]

Actually, ibm z14 cpus are much slower than a decent intel xeon. You can emulate a mid sized ibm mainframe on x86 using Turbohercules. A high end 8-socket x86 server would beat the shit out of the largest z14 server. This why ibm never publishes benchmarks of mainframes vs x86. Ibm always publishes benchmarks of Power cpus vs x86 cpus, because Power is not shit. But you will never ever find benchmarks of mainframe vs x86. Why? Because mainframe cpus are dog slow. Mainframe have good I/O, but the cpus are shit slow. Your x86 is much faster than z14.

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.

[Zucca]

you got some benchmarks for this?

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

[Dawn]

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.

Kebabbert is an infamous pro-Sun troll that's been posting copypasta about how slow z supposedly is for a decade. https://arstechnica.com/civis/viewto...26ec#p24960491 is an example of what he's going to post.

He also happens to be wrong - despite his utterly nonsensical methodology to the contrary, z is *damn* fast, especially on single-thread. It's pointlessly expensive (tens of thousands of dollars list price for a single core activation - which imo makes z/Linux pointless unless you already have a z investment) but I can say from personal experience it's not slow.

[boxie]

Kebabbert is an infamous pro-Sun troll that's been posting copypasta about how slow z supposedly is for a decade. https://arstechnica.com/civis/viewto...26ec#p24960491 is an example of what he's going to post.

He also happens to be wrong - despite his utterly nonsensical methodology to the contrary, z is *damn* fast, especially on single-thread. It's pointlessly expensive (tens of thousands of dollars list price for a single core activation - which imo makes z/Linux pointless unless you already have a z investment) but I can say from personal experience it's not slow.

Thanks for that - any chance you have access to a Z or point us in the direction of some benchmarks? - I am still kinda curious

[Dawn]

Thanks for that - any chance you have access to a Z or point us in the direction of some benchmarks? - I am still kinda curious

I do have access to a z mainframe (z13 - regrettably no z14 yet) but like I said, IBM is sensitive about public numbers and prefers users to benchmark their own workloads. I can say, however, that single-thread performance is excellent (at least on integer - I haven't done much with floating point code.)

[boxie]

I do have access to a z mainframe (z13 - regrettably no z14 yet) but like I said, IBM is sensitive about public numbers and prefers users to benchmark their own workloads. I can say, however, that single-thread performance is excellent (at least on integer - I haven't done much with floating point code.)

fair enough - thanks for taking the time to share you knowledge/experience!

[kebabbert]

Kebabbert is an infamous pro-Sun troll that's been posting copypasta about how slow z supposedly is for a decade. https://arstechnica.com/civis/viewto...26ec#p24960491 is an example of what he's going to post.

He also happens to be wrong - despite his utterly nonsensical methodology to the contrary, z is *damn* fast, especially on single-thread. It's pointlessly expensive (tens of thousands of dollars list price for a single core activation - which imo makes z/Linux pointless unless you already have a z investment) but I can say from personal experience it's not slow.

Well, if you read on this a bit, you will see that lot of people say that Mainframe cpus are dog slow. They compiled the same source code for Linux, onto Mainframes running Linux and compared to Linux running on x86, and Mainframes were several times slower. I have given three sources in my link you cited, do you have any objections to those Mainframe experts? One of them wrote a Mainframe emulator, and he says Mainframes are much slower than x86. Another Mainframe expert ported Linux to Mainframes and also says that Mainframes are much slower. And the third is a consulting firm, also saying the same thing.

Are you saying all of these Mainframe experts are lying? Have you ever written a Mainframe emulator or ported Linux to Mainframes?

I also wrote that Mainframes have good I/O and are damn fast at I/O - but their cpus are dog slow. You must differentiate between Mainframe cpus and Mainframe I/O. One of them are dog slow, the other one is fast. And everyone agrees: the Mainframe cpus are dog slow. Mainframes have lot of I/O help cpus, for instance one Mainframe had 296.000 I/O channels. Of course Mainframes will have superior I/O to a x86 server. But, if you give the same amount of I/O help cpus to a x86 server - the x86 server would have faster I/O as well.

When you go out on the internet and say that Mainframe cpus are fast - then you are just not telling the truth. Both you and I know that. If you were correct, you could show us any evidence; benchmarks, links, whatever. But nope. There is no evidence for your false claims - some would call it FUD. Are you deliberately FUDing? If you do not agree on this; show us some benchmarks proving the opposite. But guess what, you will never find any such links. Why? Because Mainframes are slower than x86! That is why you can not find any benchmarks that shows Mainframes faster than x86 - because that is not true. You can not prove false things. Simple as that.