----- Original Message ----- From: "John S. Giltner, Jr." <[EMAIL PROTECTED]> Newsgroups: bit.listserv.ibm-main To: <[email protected]> Sent: Tuesday, July 17, 2007 7:21 PM Subject: Re: PSI MIPS (was: Links to decent 'why the mainframe thrives' article)
> > Can the mainframe z900/z990/z9 compete head to head with Intel? > IMHO, yes. Well, I think that depends upon the definition of 'head-to-head'. > > Head to head. You are talking about the same number of processors and > same amount of "RAM/Central Storage". Not really. For example, I could compare the slowest Sun SPARC against the fastest Itanium, and while one might call that a 'head-to-head' comparison, it really would be skewed in favor of the Itanium. In addition, if you have two z990s vs. 48 x86 machines, you will necessarily require more memory because you will be running at least 46 more copies of an operating system. So no, I wouldn't call this a head-to-head comparison of processors. > > > Are we? NOPE. In the end to replace 80% of 7 z990 CPU's and 20GB of > RAM the Intel side will have > > Right now the plan is to ONLY have 96 total "cores" (some systems will > have single core processors and some will have dual core) CPU's and 456 > GB of RAM. This is a est. and they beleive that they may need to > increase this by as much as 50%. > > So the "head to head" comparsion is 7 CPU's and 20GB to 96 CPU's and 456 > GB of RAM. Doesn't seem to "head to head" to me. It isn't, but it is really cool information. What it seems to indicate to me is not that the z9 CPU is faster (which is may or may not be), but that the z990 system design has much better throughput. Of course, without any details on the Intel based system, it is hard to determine what the real limiting factor is. For example, whoever made the decision might have gone with the cheapest solution (hardwarewise), rather than a more robust and more expensive solution. Since the application may be very I/O bound (typical business application), the CPU speed may not even be the bottleneck (and probably isn't in this case). This would mean you need more systems to handle the same transaction load, which means more CPUs just because you need at least one for each system, and of course, more memory. I mentioned that Stratus makes very robust, fault-tolerant x86 based systems but they are really expensive. Sun makes systems that are able to handle more users/transactions as well, but still aren't as cheap as commodity server systems. A true head to head comparison would have to be done holding as many factors constant as possible. If one were to determine how many users a single, well-designed x86 box could handle, and measure performance for both it and a z990 with the same number of users/transactions, that would be a better indicator of processor speed... but that is also presuming they are running the same application. For example, if you are moving from z/OS to Windows, you are likely not running the same application code and probably not even from the same vendor. So the OS and application differences can account for quite a bit of performance difference. However, if both are running Linux and the same application - then you have a better head-to-head test. So, I suggest that a real head-to-head CPU performance comparison would be to put the exact same load on a single processor system (for example) and measure the response time, or how fast a problem is solved, etc. That is what CPUs do (calculations). However, where the mainframe shines is its ability to support many users and perform massive amount of I/O without bogging down the system. This is not a processor issue, but a system design issue, as has been mentioned before. > > That is a about a 13:1 ratio on CPU's and 23:1 on "RAM". If Intel was > "faster" then we should be able to do more work on less processors. No. As explained above. It would be true if all other factors were the same, but they are not. > > Please show me a site that has migrated off a modern day mainframe to > Intel using the same number of CPU's and same amount of Central > Storage/RAM as they had on the mainframe. You know head to head. I would like to see just that kind of comparison myself, but we probably won't see it. But, as I said, the information you posted is really interesting. If you can provide more details about the Intel systems, that would be useful (yes, I am curious. No, I am not working for a hardware vendor of any kind). My conclusion regarding this situation is that it seems obvious that the processor speed isn't really the issue. And in that case, whether the z9 is slower than an x86 (or RISC) processor is not even important. This is where appropriate benchmarks are useful (such as the SPECjbb or TPC benchmarks I've mentioned, or any other that can be identified) - to show where the real strengths of a system are. Regards, Dean ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html
