Slashdot Mirror
IBM Announces First Linux-only Mainframes
A reader writes "The new Z-series mainframe for Linux, which costs $400,000 and is aimed at processing transactions at large businesses, is IBM's first mainframe computer sold without IBM's traditional z/OS mainframe operating system. More info at the IBM zSeries page" This is something that IBM and others of Big Iron vendors of *NIX have said - as Linux grows in maturity, they want to replace their *NIX with Linux. However, there's still work to be done in that area.
No. z/VM is the 'meta-OS'. It's pretty much analagous to VMware in what it can do, in terms of hosting other OSs underneath it.
z/OS is geared at high volume transaction, database, batch processing. it runs either z/VM or more typically natively or in an LPAR.
An LPAR is a 'logical partition', a way of dividing a m/f up into several virtual machines.
for now, these are static and implemented when a partition is 'booted' - IPL'd (initial program load) in m/f terms.
VM on the other hand supports hundreds, even thousands of dynamically generated virutal machines. You can run VM inside an LPAR providing two levels of partitioning. I expect VM and LPAR technologies will converge at some future point.
meanwhile everyhting can talk to each other over 'hipersockets' - memory to memory pipes that looks like a tcp/ip network to your software - blindingly fast
Architecture is the key. What's the difference between a 120 MIPS mainframe and 3000 MIPS desktop, and why is the 120 MIPS mainframe faster in mainframe type applications?
Architecture. Specifically, things like I/O, process handling, etc.
Don't get me wrong, I'm a strong believer that "desktop" type hardware can compete with the big boys, especially considering the cost diferences and the extra speed, boxes, redundancy, etc that you can buy with all that cash you save. But... there are times when the big mainframe architectures really do have a reason for being.
Just my $.05 (inflation, you know).
-- "Other than that, how was the play Mrs. Lincoln?"
MIPS = Meaningless Indicator of Processor Speed
..."
The mainframe is MIPS per CPU, so the 16-way box is 16*120. Also, 120 MIPS is slow these days for a mainframe.
Write a simple memory intensive program and try it on a mainframe and try it on a PC. I guarantee that you won't get 3000 MIPS out of a desktop, even if the data fits in cache. Many reasons for this....
The s390 ISA is definitely CISC, you can copy a whole string with MVCL, that count's as one instruction. Do this on RISC machines and it might take a loop and execute dozens of instructions. Hence "Meaningless
About 2 years ago I wrote some C code to recursively quicksort 20M random integers and tried it on a bunch of platforms. A mainframe that was about 1 cycle behind fastest available gave me about the same single processor performance as a 1GHz PC, both a little slower than Alpha.
The big differentiator is memory architecture. How much time do you lose for a cache miss? Most processors only operate at 20-30% of theoretical maximum speed on big problems.
Memory speed has not kept up, that 2GHz box you dream about is not twice as fast as a 1GHz box, particularly if you're crunching a lot of data.
(Shot of IBM's new server standing alone in a server room)
ANNOUNCER: "If you think we're overcompensating for something with our really, really big mainframe running linux..."(Cut to shot of a dozen small servers being carted off) "...You're absolutely right."
It's not the maintenance that is the problem, things like configuration management and data integrity are more important. If you have a hundred servers, then you have a hundred places to check that everything is in sync. If you are running a small shop with a dozen or so machines and one administrator then they can keep all the state in their heads. When you get up to hundreds then the state is larger than one person can easily cope with and you start having to communicate state to others. With hundreds of boxes, it is easy to overlook things, with fewer boxes, the communication is easier, and cheaper.
The other thing is CPU residency. Lots of small boxes wastes CPU power because they tend to be devoted to one task and are only capable of that task. The problem is, they are so small that you can't add other tasks to them so you need a new box... Generally, CPU residency on small boxes runs about 10%, with mainframes, this can rise to 90%. Take two tasks - one runs during the day, one runs during the night. Conventional wisdom would allocate two small boxes, one per task wasting them for most or their life. Mainframe usage would run them both on the mainframe - this gives each process more power when they run and doesn't waste the box when they don't. Most traffic tends to be peaky but only for a short period of time so if the box is large enough to hold them both, you get a saving whilst still making all the tasks faster.
Small boxes are good when you need maximum cycles per buck and the task is easily partitionable with minimal interprocess communication and the tasks are continuous. When the tasks are not easily partitionable, need lots of IPC or are peaky then larger boxes make sense.
The thing to remember is that where the scale is large, you need to make use of that scale to get maximum performance. You don't see chemical plants using hundreds of small vats, they use a few really big ones. With these systems they are used at a scale where communications and simply keeping track of what is going on is a major exercise and hence a major expense.
My Experience? Well - put it this way, the SunFire 6800 turned up a few weeks ago, the 4800 turns up on wednesday as part of a plan to replace a Tandem mainframe and they will be sitting next to quite a few racks holding Sun E3500s, E450s, E250s, t1s, HP netservers, IBM RS6000s and SGI Origin 2000s and indeed a MacOS server or twenty. A lot of our comms talk to Stratus mainframes and the machine room cooling plants are a more pressing problem than CPU speed.