Thoughts on building a high-performance NAS Setup for data crunching?

Yeah, z196 is very very fast for certain workloads. See this: http://www.realworldtech.com/z196-mainframe/

Granted, they probably would be incredibly ill-suited for what we do, but they definitely do have a niche. I really love reading about cpu/gpu architectural design, and z196 is an interesting chip. Massive floating point performance is not what they're targeted at - there's a reason they measure them in MIPS. Think of the difference between a Semi-truck (mainframe architectures), a Prius (ARM, etc) and a Porsche (x86/GPGPU/etc) - they are all very good for what they are designed for, just most people enjoy the Porsche more :)
 
Last edited:
TheyDroppedMe said:
Yeah, z196 is very very fast for certain workloads. See this: http://www.realworldtech.com/z196-mainframe/

Granted, they probably would be incredibly ill-suited for what we do, but they definitely do have a niche. I really love reading about cpu/gpu architectural design, and z196 is an interesting chip. Massive floating point performance is not what they're targeted at - there's a reason they measure them in MIPS. Think of the difference between a Semi-truck (mainframe architectures), a Prius (ARM, etc) and a Porsche (x86/GPGPU/etc) - they are all very good for what they are designed for, just most people enjoy the Porsche more :)
Click to expand...

madrebel said:
how exactly did you arrive at that performance guestimate? those mainframes are anything but slow. stupid expensive, absolutely. however for what they do they do an extremely good job. not that you can't build something superior for less, you can, however those bigiron boxes do some really impressive shit.
Click to expand...
Are we talking about the same IBM Mainframe chip, z196, "the worlds fastest chip"?
http://www.engadget.com/2010/09/06/ibm-claims-worlds-fastest-processor-with-5-2ghz-z196/
It has 5.26GHz and in total ~250MB cpu cache. Not 24MB cpu cache, but it as half a GB of cpu cache! Impressive specs.

Well, the z196 is a slow cpu. Sure, an entire Mainframe is fast if we talk I/O, but cpu wise they are slower than high end x86 cpus. Again, there is a reason IBM never publishes common benchmarks comparing dhrystone, SPECint, etc - which can be used to assess performance compared to x86. Have you ever seen Mainframe cpu benchmarks, using common cpu benchmarks? No. Why? Because they are slow and IBM tries to hide that fact. You can emulate a Mainframe on your laptop, running TurboHercules. Sure, it will not do I/O wise, but performance wise it is ok.

Using software emulation you can use a 8-socket x86 server to get a midsize IBM Mainframe of 3.200 MIPS. And remember that software emulation is 5-10x times slower. If it would run native code, it would be ten times faster, and give something like 32.000MIPS, almost matching the biggest z196 Mainframe which has 50.000MIPS.
http://en.wikipedia.org/wiki/Hercules_emulator#Performance

Here is a Mainframe expert that happened to compare x86 and Mainframe cpus. He ported Linux OS to Mainframe and could compare same workload on Mainframes and on x86. He concluded that 1 Mainframe MIPS, equals 4MHz x86. The biggest z196 Mainframe, has something like 50.000MIPS. That equals 200GHz. But remember a 10-core Xeon at 2GHz = 20GHz. So you need ten Xeon cpus to match the biggest z196 Mainframe with max nr of cpus. If we talk cpu wise. The Xeon server will not be able to match the Mainframe if we talk I/O.
http://www.mail-archive.com/linux-390@vm.marist.edu/msg18587.html

Again we see that ten x86 highend cpus give 32.000-50.000 MIPS. It is the same ballpark. There are two independent sources saying that 8-socket x86 servers give the same number: 40.000ish MIPS. Almost matching a fully equipped z196 Mainframe. Instead of using 8 Xeon cpus, you could buy two 8-socket servers and surpass the z196 Mainframe. If we talk about raw number crunching ability. There is a reason supercomputers are built up of x86 clusters, and not clusters of Mainframes.

Hence, the "performance" of IBM Mainframes are a well kept secret. And IBM intends to keep that way by refusing benchmarks. Just think a bit and dont trust IBM marketing.


EDIT: I/O wise the Mainframes are fast. They can have somthing like 296.000 I/O channels. They have lots of I/O co processors. But, what happens if we add all those I/O help processors on a highend x86 server? It would be faster, both cpu and I/O.

And, IBM claims a big Mainframe can virtualize 1.500 x86 servers. How is that possible, knowing how slow the Mainframe cpus are? If few x86 servers start to do some heavy cpu work, then the Mainframe will crumble. Ergo; dont trust IBM marketing. A slower Mainframe can not replace 1.500 faster x86 servers.
 
Last edited:
you cant compare apples to apples. there is no comparison. however if you think IBM Zseries are slow you're fucking high.

that said i think mainframes are stupid but they arent slow by any stretch.
 
madrebel said:
you cant compare apples to apples. there is no comparison. however if you think IBM Zseries are slow you're fucking high.
Click to expand...
I dont think they are slow. They ARE slow. Cpu wise. But I/O wise they are fast. What is it you dont understand? Have you read my links? Here is another link:

http://www.microsoft.com/presspass/features/2003/sep03/09-15LinuxStudies.mspx?
"we found that each [z9] mainframe CPU performed 14 percent less work than one [single core] 900 MHz Intel Xeon processor running Windows Server 2003."

The z10 is 50% faster than z9, and the z196 is 50% faster than z10, which means a z196 is 1.5 x 1.5 = 2.25 times faster than a z9. This means a z196 corresponds to 2.25 x 900MHz = 2 GHz Intel single core Xeon.

Thus, we have three independent sources saying the same thing: the Mainframes are not suited for number crunching.
-One source is an Mainframe expert, because he built a Mainframe emulator.

-Another source ported Linux to Mainframe, so he could compare Linux on x86. He would not be able to port Linux, unless he had deep understanding about Mainframes. So he is also an expert.

-The third source saying the same thing, is a consulting company sponsored by Microsoft. A consulting company does not say much, but maybe they are technically knowledgeable. All of them says the same thing, and give the same ball park estimates of Mainframe cpus.

So do you reject those Mainframe expert's claims? What claims are errorneous? Can you point them out, and I will mail those experts telling them they are wrong.
 
Since I was the one who first mentioned the Z-Series boxes in this thread........
A common theme you will find in most of my comments is to buy what makes sense for any particular application (application meaning entire workflow, not a particular program). I mentioned that for that particular application, the Z-Series infrastructure would have cost about 20x what we built the client/server infrastructure for. That said, I have a friend who works high up in IT for a large bank. They have a number of current mainframes (Z-Series included) and for their particular applications (including lots of legacy code (including COBOL) which would cost tens/hundreds of millions to replace). Much of this code runs MUCH faster on these boxen than the equivalent "high end" clusters (They have done trials with just about everything out there (I helped in a few), trust me he would love to dump this legacy code if he could).
As to brutalizer, you point out 3 links that supposedly make ANYTHING better than a mainframe. One is a post from 2003 (almost a decade ago). One is a Microsoft-financed PR piece that shows (big surprise) that a WINTEL solution is better than a mainframe (and btw they tested mostly consolidation of servers for web serving and file serving, not exactly where mainframes make their mark) and the third is a puff piece from Engadget with very little specifics but a nice pic of the CPU module. Before you hold out those three as "experts" on the Client/Server vs Mainframe battle you should have more experience and background with both.
 
Last edited:
mwroobel said:
A common theme you will find in most of my comments is to buy what makes sense for any particular application (application meaning entire workflow, not a particular program). I mentioned that for that particular application, the Z-Series infrastructure would have cost about 20x what we built the client/server infrastructure for. That said, I have a friend who works high up in IT for a large bank. They have a number of current mainframes (Z-Series included) and for their particular applications (including lots of legacy code (including COBOL) which would cost tens/hundreds of millions to replace). Much of this code runs MUCH faster on these boxen than the equivalent "high end" clusters (They have done trials with just about everything out there (I helped in a few), trust me he would love to dump this legacy code if he could).
Click to expand...
Again, I am talking about CPU performance. And CPU wise, the Mainframes suck big time. There is a reason IBM never releases cpu benchmarks. When IBM has something good, like the POWER7, they release benchmarks all over the place and boast. Regarding Mainframes, IBM has always been silent. Why?

Probably your friend has some batch I/O workloads, for which the Mainframes are much faster, yes.

There is a reason no stock exchange runs on Mainframes. The latency is not good at all. In the Stock exchange business, money is of no concern, and if Mainframes were faster than stock x86 running Linux - all stock exchanges would switch to Mainframes in a blink. The fastest stock exchange, gets the most customers. Latency is crucial to attract customers.

Latency wise and cpu wise, the Mainframes suck big time. Just think a bit.



As to brutalizer, you point out 3 links that supposedly make ANYTHING better than a mainframe. One is a post from 2003 (almost a decade ago). One is a Microsoft-financed PR piece that shows (big surprise) that a WINTEL solution is better than a mainframe (and btw they tested mostly consolidation of servers for web serving and file serving, not exactly where mainframes make their mark) and the third is a puff piece from Engadget with very little specifics but a nice pic of the CPU module. Before you hold out those three as "experts" on the Client/Server vs Mainframe battle you should have more experience and background with both.
Click to expand...
Those links you talk about, are not the relevant ones I pointed out. I have showed more links, and you chose the wrong ones. I am trying to prove that Mainframe cpus are slow, and the Engadget link says they are worlds fastest - which is not really correct. Instead, it is a proof of IBM marketing lies: claim it is the worlds fastest, when it is slow.

Here are the three relevant links that show that Mainframes are slow:
-One with a Mainframe expert who developed an emulator. If you have written an emulator, then you know all the ins and outs of the target platform, right?
http://en.wikipedia.org/wiki/Hercules_emulator#Performance

-One with someone who ported Linux to Mainframes. If you port an OS, then you delve deeply into architectures and assembler. So that guy knows what he is talking about, especially since he could compare same workloads.
http://www.mail-archive.com/linux-390@vm.marist.edu/msg18587.html

-And the last one, the consulting firm sponsored by Microsoft who benchmarked a Xeon vs a Mainframe. Only benchmark I have seen that compares x86 to Mainframes. IBM forbids such benchmarks.
http://www.microsoft.com/en-us/news/features/2003/sep03/09-15LinuxStudies.aspx

If you suggest these Mainframe experts are wrong, please point out the errors they did?


EDIT: If you know of Mainframe benchmarks/studies showing superior cpu performance, I would like to see them - and then I will stop saying that Mainframe cpus are slow. But unless there are some indications showing the opposite to my links, I see no rreason to change my mind? Science should be built on observations and studies, not belief?
 
brutalizer said:
As to brutalizer, you point out 3 links that supposedly make ANYTHING better than a mainframe. One is a post from 2003 (almost a decade ago). One is a Microsoft-financed PR piece that shows (big surprise) that a WINTEL solution is better than a mainframe (and btw they tested mostly consolidation of servers for web serving and file serving, not exactly where mainframes make their mark) and the third is a puff piece from Engadget with very little specifics but a nice pic of the CPU module. Before you hold out those three as "experts" on the Client/Server vs Mainframe battle you should have more experience and background with both.

Those links you talk about, are not the relevant ones I pointed out. I have showed more links, and you chose the wrong ones. I am trying to prove that Mainframe cpus are slow, and the Engadget link says they are worlds fastest - which is not really correct. Instead, it is a proof of IBM marketing lies: claim it is the worlds fastest, when it is slow.

Here are the three relevant links that show that Mainframes are slow:
-One with a Mainframe expert who developed an emulator. If you have written an emulator, then you know all the ins and outs of the target platform, right?
http://en.wikipedia.org/wiki/Hercules_emulator#Performance

-One with someone who ported Linux to Mainframes. If you port an OS, then you delve deeply into architectures and assembler. So that guy knows what he is talking about, especially since he could compare same workloads.
http://www.mail-archive.com/linux-390@vm.marist.edu/msg18587.html

-And the last one, the consulting firm sponsored by Microsoft who benchmarked a Xeon vs a Mainframe. Only benchmark I have seen that compares x86 to Mainframes. IBM forbids such benchmarks.
http://www.microsoft.com/en-us/news/features/2003/sep03/09-15LinuxStudies.aspx
Click to expand...
Those ARE 2 of the 3 links I was referencing, the ones you re-posted. One is a post from 2003 (almost a decade ago). One is a Microsoft-financed PR piece that shows (big surprise) that a WINTEL solution is better than a mainframe (and btw they tested mostly consolidation of servers for web serving and file serving, not exactly where mainframes make their mark) and the third makes this specific comment "It is difficult to determine exactly how Hercules emulation performance corresponds to real mainframe hardware, but the performance characteristics are understandably quite different. This is partially due to the difficulty of comparing real mainframe hardware to other PCs and servers as well as the lack of concrete, controlled performance comparisons. Performance comparisons are likely legally impossible for licensed IBM operating systems, and those operating systems are quite different than other operating systems, such as Linux."
Again, all I said was buy and use what makes sense. Whether that is from a $$$ standpoint, from an existing infrastructure standpoint or from a pure performance standpoint based upon your current workflow and/or applications. I'll reiterate what I have said to you a number of times. Nothing is perfect for any circumstance. Sometimes the x86 C/S architecture is best, sometimes the Mainframe architecture is the better choice. Buy what make sense for the particular application you need to service.
I don't have any Z-Series mainframes. For our particular application they don't make sense. But for Government, hospitals and many other large businesses with batch based workflows and with very high connection/user counts they can be obscenely fast, especially when running legacy code. Mainframes don't suck any more or less than x86 sucks, it depends on the application. Obvioulsy if x86 C/S were perfect for any application, nobody would buy anything but because the numbers wouldn't work. You need to look less into absolutes (Mainframes are slow, etc) and more into "the right tool for the right job" mode of thinking.
Click to expand...
 
Last edited:
mwroobel said:
Those ARE 2 of the 3 links I was referencing, the ones you re-posted. One is a post from 2003 (almost a decade ago). One is a Microsoft-financed PR piece that shows (big surprise) that a WINTEL solution is better than a mainframe (and btw they tested mostly consolidation of servers for web serving and file serving, not exactly where mainframes make their mark)
Click to expand...
Yes, true. But it is a benchmark, right? And benchmarks can be reproduced, right? And the benchmark says that generation of Mainframe cpu is as fast as a single core Xeon at 900MHz. The thing is, several Mainframe experts give the same ball park numbers of performance. If all these persons made up the numbers, they should differ, not be the same?


and the third makes this specific comment "It is difficult to determine exactly how Hercules emulation performance corresponds to real mainframe hardware, but the performance characteristics are understandably quite different. This is partially due to the difficulty of comparing real mainframe hardware to other PCs and servers as well as the lack of concrete, controlled performance comparisons. Performance comparisons are likely legally impossible for licensed IBM operating systems, and those operating systems are quite different than other operating systems, such as Linux."
Click to expand...
Yes, and the co-founder of the company also says that a 8-socket Nehalem-EX gives 3.200 MIPS under software emulation. That is quite respectable and a decent sized Mainframe. Of course, if you would rewrite the software for x86, it would run 5-10x faster, giving 16.000-32.000 MIPS, which is a high end Mainframe.


Again, all I said was buy and use what makes sense. Whether that is from a $$$ standpoint, from an existing infrastructure standpoint or from a pure performance standpoint based upon your current workflow and/or applications. I'll reiterate what I have said to you a number of times. Nothing is perfect for any circumstance. Sometimes the x86 C/S architecture is best, sometimes the Mainframe architecture is the better choice. Buy what make sense for the particular application you need to service.
I don't have any Z-Series mainframes. For our particular application they don't make sense. But for Government, hospitals and many other large businesses with batch based workflows and with very high connection/user counts they can be obscenely fast, especially when running legacy code. Mainframes don't suck any more or less than x86 sucks, it depends on the application. Obvioulsy if x86 C/S were perfect for any application, nobody would buy anything but because the numbers wouldn't work. You need to look less into absolutes (Mainframes are slow, etc) and more into "the right tool for the right job" mode of thinking.
Click to expand...
I have not said something else. You are saying the same thing as I say, in different words.

I say: cpu wise the Mainframes suck big time. So, you should not run number crunching or something similar cpu heavy on a Mainframe. Use a x86 cluster for that. If you are running I/O things and batch loads, then use a Mainframe because a x86 can not match the I/O of a Mainframe. This is exactly the same thing as I am saying: use the right tools (but Mainframes are to be avoided for cpu heavy things). Good that we have agreed on this.

Of course, if you ever find benchmarks comparing x86 to Mainframe cpus, I would be very interested. I have never found anyone, because IBM refuses such benchmarking.
 
Originally Posted by brutalizer
Of course, if you would rewrite the software for x86, it would run 5-10x faster, giving 16.000-32.000 MIPS, which is a high end Mainframe.
Click to expand...

madrebel said:
you make this sound so trivial.
Click to expand...

You are right mad. I would also love to find out where he got the 5x - 10x performance increase stats. What he is saying is that if I wrote an app in C and compiled it for (lets say a Z-Series CPU) that if I change the compile flags to build for x86 it would run 10x faster? As someone who has run quite a number of projects that involved architecture changes in both hardware AND software, he has NO clue what he is talking about.
 
Last edited:
As an additional comment to brutalizer, since you seems to value MIPS (which really should stand for Meaningless Indicator of Processor Speed :D ) you should know that any benchmark such as that is completely meaningless when comparing CPU's of different processor families. If you want to read up on benchmarks you can go here.
 
Last edited:
RPG ...

not to mention some of these internal software packages were written originally by folks who are no longer alive.

mainframes don't make sense for modern programming if you're starting from scratch. you would be a fool to choose a mainframe over distributed elastic compute.

however you would be equally foolish to take decades worth of biz processing and rewrite it all so you can get off IBM mainframes.
 
madrebel said:
RPG ...

not to mention some of these internal software packages were written originally by folks who are no longer alive.

mainframes don't make sense for modern programming if you're starting from scratch. you would be a fool to choose a mainframe over distributed elastic compute.

however you would be equally foolish to take decades worth of biz processing and rewrite it all so you can get off IBM mainframes.
Click to expand...

So two data points:
1. I was at a client recently that used a mainframe plus a blackbox application to generate plus four zip codes. Seriously.
2. Typically where I see mainframes being purchased these days is in the cannot fail and cannot be even slightly incorrect space, especially when there is lots of legacy code.
 
madrebel said:
RPG ...

not to mention some of these internal software packages were written originally by folks who are no longer alive.

mainframes don't make sense for modern programming if you're starting from scratch. you would be a fool to choose a mainframe over distributed elastic compute.

however you would be equally foolish to take decades worth of biz processing and rewrite it all so you can get off IBM mainframes.
Click to expand...

Not only that, in many cases the users don't even have source code to the apps, and in some cases don't even know exactly some of the apps work.
 
madrebel said:
you make this sound so trivial.
Click to expand...
It can be difficult to port software to another cpu architecture, but it is often viable. Not trivial.


mwroobel said:
You are right mad. I would also love to find out where he got the 5x - 10x performance increase stats. What he is saying is that if I wrote an app in C and compiled it for (lets say a Z-Series CPU) that if I change the compile flags to build for x86 it would run 10x faster? As someone who has run quite a number of projects that involved architecture changes in both hardware AND software, he has NO clue what he is talking about.
Click to expand...
5-10x performance increase? That is a common number that "everybody" uses.

Look, if you software emulate a different cpu architecture then you take a big hit in performance. Say I emulate an Amiga or a Nintendo on my x86, then I take a big hit. The emulation of a different cpu architecture will cost much. Typically everything will run 5-10x slower. That is consensus.

If I software emulate a Mainframe architecture with all its quirks and weird 31/35bit stuff, then my x86 server will take a big hit. Typically everything will run 5-10x slower because software emulation of a different architecture adds a big penalty. I mean 31 bits. Not 32 bits, but 31. Or is it 35 bits? I dont remember. But there are many pecularities in the z/OS world, stemming back from arcane 1960. Providing full backwards compatibility to 1960.

For instance, VMware Workstation/ESXi is not software emulation. Every asm instruction is more or less just sent unmodified and directly down to the x86 cpu to execute, so you can run nearly full speed. But emulation is another thing, you need to emulate registers, ALU, everything. That takes a lot of horsepower. So, you will not see VMware virtualize for instance SPARC cpus on x86, because that would be _emulation_ and a different thing that hogs performance on its own. Everything would run 5-10x slower. Then it would be faster to run Oracle DB natively on x86 instead of running it on SPARC, emulating it on x86.

Agreed?




EDIT: I read an article of a developer doing Nintendo 8-bit emulators for Windows. He said, to properly emulate the 8-bit (everything, timing between different components, etc) would require huge performance, and is not feasible in the near future. He talked about... 100x slower penalty? Let me see if I can find the article again....
Here it is
http://arstechnica.com/gaming/2011/...-3ghz-quest-to-build-a-perfect-snes-emulator/
"Take the case of DICE, the digital integrated circuit emulator. Here is an emulator that works at the transistor level for absolutely perfect recreation of the very first video games ever created. To run Pong at about 5-10fps, DICE requires a 3GHz processor. Yes, you read that right: no computer processor at this time that can run Pong at the circuit level at full speed. It's not that DICE is a slow program; indeed, it is very well optimized. It's that there is enormous overhead to simulating every last transitor propagation delay.
...
As much as I would like every emulator to support every last transistor propagation delay, in truth this simply isn't possible. It becomes fair to say that we will likely never see hardware capable of emulating an N64 at this level at playable framerates within our lifetimes."
 
Last edited:
madrebel said:
RPG ...
Click to expand...
I also like Role Playing Games! :)


however you would be equally foolish to take decades worth of biz processing and rewrite it all so you can get off IBM mainframes.
Click to expand...
I am not saying that. I am saying: Mainframe cpus are slower than a highend x86 cpu. The Mainframe server has very fast I/O, but the cpu is slow.

If you want to rewrite all legacy Mainframe code to x86, for a huge cpu increase, is up to you. But often the Mainframe workloads are not cpu bounded, but I/O. And if you need I/O, then you should stay with Mainframes. But if you are going to do number crunching, then use a cluster of x86 instead. Ok?

Mainframe cpus = slow.
Mainframe servers = good I/O
x86 cpus = fast
x86 servers = bad I/O
 
Wow, just wow. That's the number that "everybody" uses? You, who demands MULTIPLE sources of linked items before you will believe ANYTHING, uses "everybody" as a reference.
In fact, the new CPU that IBM is using is the fastest CPU in the world. No current x86 is anywhere near as fast. Even with liquid nitrogen, current x86's can clock up that fast for about long enough to grab a screenshot before the nitrogen has boiled away and the cpu clocks back down. The mainframe processors run 24/7/365 at these speeds. Second, when you absolutely, positively need hardware and software that has uptimes measured in years (Banking, Airlines, Hospitals etc) as well as ironclad security then that is where you look. Next, in some cases it isn't as obscenely expensive as you would expect. Go license 64 cores of Oracle DB (for example) and get less overall database performance than one core on a zEC12, as you say. I'll help you out: you'd probably pay about $1.5M in database software licensing plus $300K+ in annual software maintenance for your 64 X86 cores versus $47K and $9.4K on a zEC12 core. And that's one cost factor among many, not the only one. So which server is "cheaper"? If you have a workload that is of a type best serviced by a mainframe, that is what you do. If you don't, you look elsewhere. If you claim that current Mainframe CPU's like in the zEC12 are slow, you are (with all respect) nuts.
 
Last edited:
mwroobel said:
Wow, just wow. [5-10x] That's the number that "everybody" uses? You, who demands MULTIPLE sources of linked items before you will believe ANYTHING, uses "everybody" as a reference.
Click to expand...
Did you miss my link on the Nintendo emulator developer who says it is impossible to completely emulate a Nintendo64 today? Even 3GHz cpus does not suffice to do a correct and full emulation.

I remember when I had an 7MHz Amiga, and ran a C64 emulator - it was dog slow. The games were unplayable. I had a 16bits M68000 running at 7MHz, and still it could not emulate a 8bit 1MHz 6502 cpu. The M68000 was probably more than 14x faster (roughly: double the number of bits gives twice the power, and 7x higher clockspeed). In this case the performance penalty was greater than 14x. If you ever had tried to emulate other computers to play games, you would know that you need a much faster cpu. Typically you need a much faster PC to emulate an old game consol such as Megadrive or Amiga. You could never use a 8MMHz 286 to emulate a 7MHz Amiga, even a 16MHz 386 did not suffice. You need to get to 700MHz or more, i.e. 100x faster.

So the number 5-10x slower is not made up. For people used to emulate other computers, it is well known. Maybe that number is even on the low side. You need more than a 5x faster computer to do a good emulation, surely. 10x faster will not suffice either, you might need 20x faster or so. Depending on how exact the emulation is. Emulation is not free, it costs.




In fact, the new CPU that IBM is using is the fastest CPU in the world. No current x86 is anywhere near as fast.
Click to expand...
Interesting claim by IBM. Have you seen any benchmarks showing the superiority of the EC12 cpu? IBM said the same thing, everything was the same, about the previous generation z196 cpu. And still it was dog slow.

If you can find any proof/evidence/benchmarks/whatever at all, that shows that the slow EC12 cpu is faster than a high end x86 cpu, I will change my mind. That will end the discussion. Just keep it simple: show us some benchmarks and you will win. If you dont find any benchmarks (which will not find) then maybe you can show us some other links? It is ok if those links even are from IBM. Then we can discuss the links, instead of you just repeating IBM marketing material with no proofs, benchmarks or whatsoever.

I didnt expect you, mrwroobel, to blindly trust what a company says. I mean, you ARE clearly very knowledgable, and you know your stuff. But have you ever seen any benchmarks? No? Why dont you think a bit, instead of repeating what IBM tells you? Use the knowledge you possess, and think a bit? "Can those links brutalizer shows, be true?", "why have IBM never released any benchmarks? IBM likes to boast how fastest in the world these cpus are, why are there no corresponding benches?" etc. I am sure you can make up some better questions on you own. Just use your vast knowledge.



If you claim that current Mainframe CPU's like in the zEC12 are slow, you are (with all respect) nuts.
Click to expand...
Of course they are slow. But the Mainframe server has great I/O. But the cpu is slow. Why are there no corresponding benches from IBM? IBM has just released the "worlds fastest cpu" (again) without any benches. When a company releases a new product and claims things, they should back it up with benches or something, right? Otherwise, it is just FUD. "Our product XXX is faster than product ZZZ from the other company - but we will not show you any benches, you will just have to trust us" - this is FUD.

How can people blindly trust others, without even critically think a bit? I always analyze what I hear/read: "can it be true?". No matter who says it. Even Noble prize winners in physics have said really stupid things that are plain wrong. And people dont analyze what they hear? Whats wrong with people? If the USA said things, would you not think "can it be true?"
 
You are welcome to believe what you like. In one thread, you state that modern CPU's are so powerful they can accelerate ANY hardware functionality (you were talking about ZFS vs. RAID at the time) and now you say that modern CPU's aren't even capable of emulating a N64 (Which, it seems they aren't, at least with the current software based emulator). Regardless, believe what you like.... It comes down to buy what works, buy what makes sense... For a particular workload.
 
mwroobel said:
You are welcome to believe what you like. In one thread, you state that modern CPU's are so powerful they can accelerate ANY hardware functionality (you were talking about ZFS vs. RAID at the time) and now you say that modern CPU's aren't even capable of emulating a N64 (Which, it seems they aren't, at least with the current software based emulator). Regardless, believe what you like....
Click to expand...
I dont base my beliefs on blind trust, I want to have some basis for my beliefs. Things like, research papers, benchmarks, white papers, etc. That is the reason I believe ZFS has the best data protection, because of all the research.

To trust IBM marketing without any proofs - is religion. The question is, why dont anybody ask for proofs from IBM? No one does. Really really weird.

I remember when IBM claimed the POWER6 had... 240(?) GB/sec bandwidth. Considering it is a extremely high number compared to SPARC Niagara, I dug a bit. Well, it turns out that IBM added all bandwidth in the POWER6 cpu: L1 cache + L2 cache + .... Well, this is clearly wrong to do. If you have studied higher discrete math, you know that MAXFLOW = MINCUT.
http://en.wikipedia.org/wiki/Max-flow_min-cut_theorem
This theorem basically says that the maximum bandwidth can never be greater than the smallest restricted flow. If you have a restriction of 1GB/sec somewhere, then the restriction is the fastest it can be. The POWER6 can never have more than 1GB/sec bandwidth, if that were the case. It is like, if there is a small pipe in your house which can transfer 1 litre/sec, then you can never get more water than 1 litre/sec. And you can not add the flows in all pipes in your house, like "1l/sec + 100l/sec +50l/sec = 151l/sec, wow, I can sustain 151 litre/sec water!!". Wrong.

So, clearly, IBM marketing is either ignorant of higher math, or are just lying. You pick your choice.

BTW, the word "FUD" was invented because of IBMs foul behavior to its competitors. http://en.wikipedia.org/wiki/Fear,_uncertainty_and_doubt#Definition
"FUD was first defined with its specific current meaning by Gene Amdahl the same year, 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."



It comes down to buy what works, buy what makes sense... For a particular workload.
Click to expand...
Agreed. But dont buy a Mainframe for cpu heavy work. Even wikipedia says so:
http://en.wikipedia.org/wiki/Linux_on_System_z#Appropriate_workloads
"...mainframes in general, and Linux on System z in particular, do not perform well for single task computations. Examples include most scientific simulations, weather forecasting, and molecular modeling. Supercomputers, including Linux-based supercomputers, excel at these workloads."
 
mwroobel said:
You are welcome to believe what you like. In one thread, you state that modern CPU's are so powerful they can accelerate ANY hardware functionality (you were talking about ZFS vs. RAID at the time)
Click to expand...
No, I did not say that. I said something like:

When cpus are fast enough to emulate hardware in pure software, then I dont see the need for buying hardware.

And I said, for example, when cpus are fast enough to do good cryptography in real time, you will not need cryptographic accelerators. And when cpus are fast enough to do good graphics in real time, you dont need to buy a separate gpu. And today, cpus are fast enough to do XOR parity / ZFS checksum calculations so you dont need a dedicated separate hardware raid card.

What can be emulated in software, need not be bought as hardware. I never said everything can be emulated fast enough. But that can be emulated fast enough in software, makes hardware obsolete. So thus, hardware raid cards are obsolete.

This is what I said. We can go back to the thread and I can quote myself if you wish.
 
mwroobel said:
You are welcome to believe what you like. In one thread, you state that modern CPU's are so powerful they can accelerate ANY hardware functionality (you were talking about ZFS vs. RAID at the time) and now you say that modern CPU's aren't even capable of emulating a N64 (Which, it seems they aren't, at least with the current software based emulator). Regardless, believe what you like.... It comes down to buy what works, buy what makes sense... For a particular workload.
Click to expand...

well said :).....
I believe your statement!
 
If I was doing this:

Put together a data storage box (hard drives and RAID hardware) with an external SATA3 bus, push the box from server to server as needed.

Have as many boxes as necessary.

---

I am an old timer and remember when it was faster to move tape reels by airplane from the North Slope to the processing center. Now people might try to use that internet thing.
 
brutalizer said:
So thus, hardware raid cards are obsolete.
Click to expand...

<sarcasm>
I am sure that Areca, Adaptec, LSI, IBM, HP, Dell, Intel and any of a dozen other companies which sell and/or design (or supply the underlying silicon/firmware for) Hardware RAID cards will just turn off the lights and stop designing/selling them since you have deemed them obsolete.
</sarcasm>
 
mwroobel, in his defense, most of the new 'flashy' storage products/vendors (omg so punny) are using software based parity/redundancy. netapp, emc, ibm ... sure they are still moving volume on hardware controllers as is dell and hp in their servers. the storage market is trying to move towards a more dumb hardware with smart software approach.
 
madrebel said:
mwroobel, in his defense, most of the new 'flashy' storage products/vendors (omg so punny) are using software based parity/redundancy. netapp, emc, ibm ... sure they are still moving volume on hardware controllers as is dell and hp in their servers. the storage market is trying to move towards a more dumb hardware with smart software approach.
Click to expand...

I agree (to a point), but that does not make the current crop (or upcoming models) obsolete in any way, shape or form.
 
Like everything else, it all depends on your particular needs and applications.
 
You must log in or register to reply here.
Back
Top