Re: zEC12, and previous generations, why? type question - GPU computing.
Mainframes? The 8100 was a series of small machines that grew out of the 3790. They were no more mainframes than their competitor, the S/1. Perhaps you are thinking of DPPX/370, which ran on the 9370. It is debatable (although maybe we shouldn't here!) - 8100 was a DPD product, not GSD like the midranges. -- Will -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
- Original Message From: zMan zedgarhoo...@gmail.com To: IBM-MAIN@LISTSERV.UA.EDU Sent: Wed, September 5, 2012 5:17:37 PM Subject: Re: zEC12, and previous generations, why? type question - GPU computing. On Wed, Sep 5, 2012 at 2:28 PM, Shmuel Metz (Seymour J.) shmuel+...@patriot.net wrote: In a6b9336cdb62bb46b9f8708e686a7ea0115baa1...@nrhmms8p02.uicnrh.dom, on 09/05/2012 at 11:45 AM, McKown, John john.mck...@healthmarkets.com said: If it is because the z architecture is not good at numeric computation, The z architecture is fine for numeric computations. The problem is that the implementation is competing with processors manufactured in bulk. If IBM could sell millions of z boxen then they'd be able to cut the price dramatically. I've always wondered what would have happened had IBM used a 370 instruction set on the PC instead of Intel. 16MB ought to be enough for anybody? :-) Since IBM wasn't manufacturing the chips, of course that wasn't even on the table, but it's still a VERY interesting Gedankenexperiment... -- zMan -- I've got a mainframe and I'm not afraid to use it They could have gone with the Motorola 6800x chips instead. However, I am not sure that Motorola would have committed to producing as many chips as IBM thought they needed. The 6800x is what was used for the 370 part of the PC-370 systems. Lloyd -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
AT/370. On Thu, Sep 6, 2012 at 2:07 AM, Leopold Strauss leopold.stra...@isis-papyrus.com wrote: Yes. It was a microprogrammed motorola-68000-chip, which was used. Name was similar to PC/370, but I am not sure about that. Many years ago the company, where I was employeed at that tim, had one for short for testing-purposes. Ibelieve to remember, it was the time, where 3033-systems came up ( before 3081/3083). On 06.09.2012 07:58, George Henke wrote: I believe IBM produced a pc with a 370 to run VM on a PC. Merrill Lynch had one. Somewhere in the late 80's I believe. On Thu, Sep 6, 2012 at 1:52 AM, Timothy Sipples1 sipp...@sg.ibm.com wrote: Yes, there are organizations that use zEnterprise servers for heavy numeric computation. Like decimal floating point. Cryptography is another excellent example. And you can buy optional CryptoExpress adapters if you want to augment the excellent capabilities found in every machine. You can also buy the optional zBladeCenter Extension (zBX) if you want to add DataPower accelerators, Power blades, and/or X86 blades. You can also add an optional IBM DB2 Analytics Accelerator, to boost many types of DB2 queries. So we're way ahead of you, John. ;-) I think the simple answer is that it depends what you optimize for in designing a server processor (or complex). But IBM has broken a lot of rules already about which server should do what, and I predict more rules will be broken. With respect to the 370-on-a-chip, IBM sort of did that with the 1975 introduction of the IBM 5100 Portable Computer starting at $8,975 (1975 dollars), although it was for a relatively narrow initial purpose (to get APL running). The 5100 sold reasonably well from what I've read, but I think there were three basic problems which prevented it from becoming a blockbuster: 1. The price was not low enough for mass market appeal. (Apple had a similar problem with the Lisa in the early 1980s.) 2. The software selection didn't exactly hit the mark, although it was a good try for the time. (IBM learned the value of software somewhat later in its evolution but not in time for the 1981 IBM PC.) 3. It probably didn't have the right third party marketing and distribution channels. With some very notable exceptions, like typewriters, at that time IBM would have had some challenges with this type of product. Keep in mind that for 1975 this was absolutely amazing technology, but amazing technology required some expense. Being early is pricey. If the 5100 debuted in, say, 1977 or 1978, it would have still been well timed but could have dramatically reduced the chip and board count. I also think the small built-in monitor could have been sacrified (at least as an option) in favor of a display port of some kind -- ideally RF for TV hookup. And IBM might have gone with a diskette drive for storage -- the 5100 was too early for the 5.25 inch drive, which debuted in 1976. Finally, if IBM had provided a little more guidance on the 370 subset instruction set they implemented, software developers could have taken over from there. So I think the 5100 could have been a nice 5110 by tweaking the recipe a bit. But history didn't happen that way. IBM had some success with the System/4 Pi avionics processors which are descended from System/360. --**--** --**-- Timothy Sipples Consulting Enterprise IT Architect (Based in Singapore) E-Mail: sipp...@sg.ibm.com --**--** -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN -- Mit freundlichen Grüßen / Kind Regards, Leopold Strauss Research and Development ISIS Papyrus Europe AG Alter Wienerweg 12, A-2344 Maria Enzersdorf, Austria T: +43 - 2236 – 27551, F: +43 - 2236 - 21081 @ leopold.strauss@isis-papyrus.**com leopold.stra...@isis-papyrus.com Visit our brand new extended Website at www.isis-papyrus.com This e-mail is only intended for the recipient and not legally binding. Unauthorised use, publication, reproduction or disclosure of the content of this e-mail is not permitted. This e-mail has been checked for known viruses, but ISIS accepts no responsibility for malicious or inappropriate content. --**--**-- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN -- zMan -- I've got a mainframe and I'm not afraid to use it -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
There were two: the PC/370 and the AT/370. I am not sure that many PC/370s got distributed as they were real SLOW. My old company had both for awhile. Lloyd - Original Message From: zMan zedgarhoo...@gmail.com To: IBM-MAIN@LISTSERV.UA.EDU Sent: Thu, September 6, 2012 8:54:25 AM Subject: Re: zEC12, and previous generations, why? type question - GPU computing. AT/370. On Thu, Sep 6, 2012 at 2:07 AM, Leopold Strauss leopold.stra...@isis-papyrus.com wrote: Yes. It was a microprogrammed motorola-68000-chip, which was used. Name was similar to PC/370, but I am not sure about that. Many years ago the company, where I was employeed at that tim, had one for short for testing-purposes. Ibelieve to remember, it was the time, where 3033-systems came up ( before 3081/3083). On 06.09.2012 07:58, George Henke wrote: I believe IBM produced a pc with a 370 to run VM on a PC. Merrill Lynch had one. Somewhere in the late 80's I believe. On Thu, Sep 6, 2012 at 1:52 AM, Timothy Sipples1 sipp...@sg.ibm.com wrote: Yes, there are organizations that use zEnterprise servers for heavy numeric computation. Like decimal floating point. Cryptography is another excellent example. And you can buy optional CryptoExpress adapters if you want to augment the excellent capabilities found in every machine. You can also buy the optional zBladeCenter Extension (zBX) if you want to add DataPower accelerators, Power blades, and/or X86 blades. You can also add an optional IBM DB2 Analytics Accelerator, to boost many types of DB2 queries. So we're way ahead of you, John. ;-) I think the simple answer is that it depends what you optimize for in designing a server processor (or complex). But IBM has broken a lot of rules already about which server should do what, and I predict more rules will be broken. With respect to the 370-on-a-chip, IBM sort of did that with the 1975 introduction of the IBM 5100 Portable Computer starting at $8,975 (1975 dollars), although it was for a relatively narrow initial purpose (to get APL running). The 5100 sold reasonably well from what I've read, but I think there were three basic problems which prevented it from becoming a blockbuster: 1. The price was not low enough for mass market appeal. (Apple had a similar problem with the Lisa in the early 1980s.) 2. The software selection didn't exactly hit the mark, although it was a good try for the time. (IBM learned the value of software somewhat later in its evolution but not in time for the 1981 IBM PC.) 3. It probably didn't have the right third party marketing and distribution channels. With some very notable exceptions, like typewriters, at that time IBM would have had some challenges with this type of product. Keep in mind that for 1975 this was absolutely amazing technology, but amazing technology required some expense. Being early is pricey. If the 5100 debuted in, say, 1977 or 1978, it would have still been well timed but could have dramatically reduced the chip and board count. I also think the small built-in monitor could have been sacrified (at least as an option) in favor of a display port of some kind -- ideally RF for TV hookup. And IBM might have gone with a diskette drive for storage -- the 5100 was too early for the 5.25 inch drive, which debuted in 1976. Finally, if IBM had provided a little more guidance on the 370 subset instruction set they implemented, software developers could have taken over from there. So I think the 5100 could have been a nice 5110 by tweaking the recipe a bit. But history didn't happen that way. IBM had some success with the System/4 Pi avionics processors which are descended from System/360. --**--** --**-- Timothy Sipples Consulting Enterprise IT Architect (Based in Singapore) E-Mail: sipp...@sg.ibm.com --**--** -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN -- Mit freundlichen Grüßen / Kind Regards, Leopold Strauss Research and Development ISIS Papyrus Europe AG Alter Wienerweg 12, A-2344 Maria Enzersdorf, Austria T: +43 - 2236 – 27551, F: +43 - 2236 - 21081 @ leopold.strauss@isis-papyrus.**com leopold.stra...@isis-papyrus.com Visit our brand new extended Website at www.isis-papyrus.com This e-mail is only intended for the recipient and not legally binding. Unauthorised use, publication, reproduction or disclosure of the content of this e-mail is not permitted. This e-mail has been checked for known viruses, but ISIS accepts no responsibility for malicious or inappropriate content. --**--**-- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists
Re: zEC12, and previous generations, why? type question - GPU computing.
gahe...@gmail.com (George Henke) writes: I believe IBM produced a pc with a 370 to run VM on a PC. Merrill Lynch had one. Somewhere in the late 80's I believe. re: http://www.garlic.com/~lynn/2012l.html#72 zEC12, and previous generations, why? type question - GPU computing 1984, xt/370 ... later same board was made available on at as at/370. basically a couple M68k executing subset of vm370 ... code-named washington. it didn't support i/o ... so vm370 was modified to communicate with a monitor running under dos on the 8088 for all i/o functions. it provided approx. 100kip 370 with 384kbytes of memory ... little bit faster than 370/115. however, since all disk i/o (paging, cms file, etc) was being done on 100ms (per block) dos hard disk. By that time, vm370 and cms had gotten quite a bit bloated ... much larger than cp67/cms that would run on 256kbyte 360/67. Also any kind of disk i/o (paging, file activity) could become extremely painful ... compared to what one was use to with real mainframe disks. I got con'ed into doing some work on it ... first thing simple paging tests showed almost any cms application would page thrash in the pageable pages available left over after vm370 kernel fixed storage size (from 384kbytes) ... exhaserbated by the paging on dos xt disk. I got blamed for several month schedule ship in the product while they upgraded the memory from 384kbytes to 512kbytes ... to cut down on severe paging problems. However, cms applications that tended to be much more file intensive than (and fared poorly in comparison with) equivalent applications developed for the DOS/XT resource limited environment. I had tried to start a project to implement a super lean and fast vm370 replacement kernel in pascal. As a demo I had re-implemented the vm370 kernel spooling function in pascal running in virtual address space. My objective was to enormously increase the throughput and performance compared to the kernel assembler implemented equivalent. I had another agenda ... I was also doing high-speed data transport project ... and for vm370 vnet ... which was dependent on vm370 spool ... I needed multi-megabyte sustained thruput to drive the links I had. misc. past posts mentioning hsdt http://www.garlic.com/~lynn/subnetwork.html#hsdt I indirectly referenced it in previous post regarding work with NSF on what was to become NSFNET backbone ... also original mainframe tcp/ip product was done for vm370 in pascal ... and I did the rfc1044 enhancements that got sustained channel thruput (between 4341 and cray machine using only modest amount of 4341 processor, about 500 times improvement in bytes moved per instruction executed). misc. past posts mentioning NSFNET backbone http://www.garlic.com/~lynn/subnetwork.html#nsfnet -- virtualization experience starting Jan1968, online at home since Mar1970 -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
In caccgc5dh-2kebjfpnmt0apwdcjjaondevelz6wnufcla7og...@mail.gmail.com, on 09/06/2012 at 01:58 AM, George Henke gahe...@gmail.com said: I believe IBM produced a pc with a 370 to run VM on a PC. XT/370 and AT/370 used a 68000 with custom microcode and a second 68000 with standard microcode. The software for it was VM/PC. Note that the later P/370 and R/370 cards implemented the full architecture and ran stock operating systems. -- Shmuel (Seymour J.) Metz, SysProg and JOAT Atid/2http://patriot.net/~shmuel We don't care. We don't have to care, we're Congress. (S877: The Shut up and Eat Your spam act of 2003) -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
XT/370 and AT/370 used a 68000 with custom microcode and a second 68000 with standard microcode. The software for it was VM/PC. Note that the later P/370 and R/370 cards implemented the full architecture and ran stock operating systems. re: http://www.garlic.com/~lynn/2012l.html#72 zEC12, and previous generations, why? type question - GPU computing http://www.garlic.com/~lynn/2012l.html#74 zEC12, and previous generations, why? type question - GPU computing between xt/at/370 and p370/p390 was a74 (7437) done in POK by the same group that had done the 3277GA (i.e. large tektronics graphics tube that plugged into side of 3277 terminal) ... a74 was their POK dept. I got con'ed into making the vm370 modifications for them ... including a74 only supported 4k storage key. this is long-winded post ... with bunch of old a74 press at the bottom http://www.garlic.com/~lynn/2002d.html#4 IBM Mainframe at home post with some of the discussion of the vm370 changes I did for a74 http://www.garlic.com/~lynn/2003f.html#56 ECPS:VM DISPx instructions old internal A74 email on the announcement of a74 http://www.garlic.com/~lynn/2000e.html#email880622 in this post http://www.garlic.com/~lynn/2000e.html#56 Why not an IBM zSeries workstation? a74 was 350kips (370, compared to 100kips for xt/at/370) and the ROMAN chip set was 168/3mips ... mentioned in previous post. -- virtualization experience starting Jan1968, online at home since Mar1970 -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
sipp...@sg.ibm.com (Timothy Sipples1) writes: Keep in mind that for 1975 this was absolutely amazing technology, but amazing technology required some expense. Being early is pricey. If the 5100 debuted in, say, 1977 or 1978, it would have still been well timed but could have dramatically reduced the chip and board count. I also think the small built-in monitor could have been sacrified (at least as an option) in favor of a display port of some kind -- ideally RF for TV hookup. And IBM might have gone with a diskette drive for storage -- the 5100 was too early for the 5.25 inch drive, which debuted in 1976. Finally, if IBM had provided a little more guidance on the 370 subset instruction set they implemented, software developers could have taken over from there. So I think the 5100 could have been a nice 5110 by tweaking the recipe a bit. But history didn't happen that way. re: http://www.garlic.com/~lynn/2012l.html#72 zEC12, and previous generations, why? type question - GPU computing http://www.garlic.com/~lynn/2012l.html#74 zEC12, and previous generations, why? type question - GPU computing http://www.garlic.com/~lynn/2012l.html#77 zEC12, and previous generations, why? type question - GPU put all logic in microcode http://en.wikipedia.org/wiki/IBM_PALM_processor 5100 had enuf 360 microcode emulation to run apl/360 http://en.wikipedia.org/wiki/IBM_5100 from above: The 5100 was based on IBM's innovative concept that, using an emulator written in microcode, a small and relatively cheap computer could run programs already written for much larger, and much more expensive, existing computers, without the time and expense of writing and debugging new programs. Two such programs were included: a slightly modified version of APL.SV, IBM's APL interpreter for its System/370 mainframes, and the BASIC interpreter used on IBM's System/3 minicomputer. Consequently, the 5100's microcode was written to emulate most of the functionality of both a System/370 and a System/3. IBM later used the same approach for its 1983 introduction of the XT/370 model of the IBM PC, which was a standard IBM PC XT with the addition of a System/370 emulator card. ... snip ... part of the issue was apl code was fairly dense ... and apl\360 workspaces were typically 16kbytes (some systems offerred 32kbytes). cambridge science center had taken apl\360 ... stripped out all the multitasking and swapping stuff and got it to run under cms workspace as large as virtual memory ... for cp67 cms\apl. some amount of work had to be done on how apl\360 storage since it tended to use all available workspace ... which resulted in page thrashing in virtual memory environment. there was also an cms\apl API to access system services (including file i/o). The combination of large workspace and file i/o allowed doing a lot of real-world applications (that couldn't be done with apl\360). The business planners in Armonk loaded the holiest of holy data (detailed customer profiles) on the cambridge system for business modeling in cms\apl. This also created something of security issue since cambridge also allowed non-employee access from various institutions in the cambridge area (students, staff, faculty). palo alto science center then did the enhancements to make vm370 apl\cms ... they also did the 370/145 apl microcode assist and the 5100. the person that did 370/145 apl microcode assist was also instrumental in many of the fortan hx performance enhancmenets. -- virtualization experience starting Jan1968, online at home since Mar1970 -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
360s, 370s, etc ... have been microcode implemented on variety of other kinds of engines. circa 1980 there was an effort to replace the wide variety of internet microprocessors used for controllers, lowmid range 370s, the planned as/400 replacement for s/38, etc ... all with 801/risc Iliad chips. For various reasons the efforts floundered and they went back to doing custom processor implementations. Was this effort in some way related, or in competition with, the UC series of controllers? Quite a lot of machines used those internally, and they even popped out with the 8100 series (the mainframes that have fallen into the memory hole). -- Will -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
wdonze...@gmail.com (William Donzelli) writes: Was this effort in some way related, or in competition with, the UC series of controllers? Quite a lot of machines used those internally, and they even popped out with the 8100 series (the mainframes that have fallen into the memory hole). re: http://www.garlic.com/~lynn/2012l.html#72 zEC12, and previous generations, why? type question - GPU computing http://www.garlic.com/~lynn/2012l.html#74 zEC12, and previous generations, why? type question - GPU computing http://www.garlic.com/~lynn/2012l.html#77 zEC12, and previous generations, why? type question - GPU computing http://www.garlic.com/~lynn/2012l.html#79 zEC12, and previous generations, why? type question - GPU computing uc controllers were much simpler, earlier (and underpowered) processors, 3705, 8100, service processor for 3081, etc. early on before 3705 was announced there was a strong effort at the science center to get cpd to use peachtree for 3705 (instead of uc) ... peachtree was much more powerful processor and was used in series/1. UCs would have been part of the internal microprocessors replaced by 801, the 801 replacement effort was circa 1980 ... but for various reasons the efforts floundered (the as/400 quickly did a cisc chip to replace the planned 801 ... but in the 90s eventually migrated to 801/risc power/pc). The followon to 4331/4341 (aka 4361/4381) were suppose to be iliad (801/risc) ... but there was a white paper (that I contributed to) that shot that down that effort (even tho I was working on 801/risc for other things). In the wake of the failure of those efforts in the earlier 80s, some number of 801/risc chip engineers left and showup working on risc efforts at other vendors (I've posted various old email from people worried that I might be following in their footsteps). bo evans had asked my wife to audit 8100 and shortly later it was effectively canceled (although continued to linger on for quite some time) ... has some amount about UC also: http://en.wikipedia.org/wiki/IBM_8100 old email referencing mit lisp machine group asking ibm for 801 processor ... and evans offering 8100 instead: http://www.garlic.com/~lynn/2006o.html#email790711 later one of the baby bells did a NCP VTAM (both) emulation on series/1 ... and outboard of mainframe ... carried sna traffic over real networking infrastructure (mainframe vtams were told all resources were cross-domain ... which was actually simulated outboard in redundant infrastructure). I did a deal with the baby bell to turn it out as an IBM product ... as well as concurrently porting from series/1 to rios (801/risc processor used in rs/6000). Because I knew that communication group would be out for my head ... I cut a deal with another baby bell to underwrite all of my development costs ... with no strings attached (their business case was that they would totally recover all my costs within the first year just replacing 37x5/NCP with new product). The internal politics that then happened could only be described as truth is stranger than fiction. part of presentation that I did at sna architecture review board meeting in raleigh, fall of 1986: http://www.garlic.com/~lynn/99.html#67 System/1 ? part of presentation by baby bell at series/1 common meeting http://www.garlic.com/~lynn/99.html#70 Series/1 as NCP past posts mentioning 801/risc, iliad, romp, rios, fort knox, power, power/pc, etc http://www.garlic.com/~lynn/subtopic.html#801 In the previous reference about using large number of 370 3mip roman (three) chip sets in racks ... the 801 chip was blue iliad ... which was first 32bit 801 chip ... and design for 20mips ... although it was never put into production (and it was a very large hot chip). Biggest design problem bottleneck was increasing problem with getting all the heat out of the rack as ever increasing numbers of chips were packed into the rack. old post http://www.garlic.com/~lynn/2004m.html#17 mainframe and microprocessors mentioning series of documents that I did on the roman/iliad rack cluster design RMN.DD.001, Jan 22, 1985 RMN.DD.002, Mar 5, 1985 RMN.DD.003, Mar 8, 1985 RMN.DD.004, Apr 16, 1985 old email discussing 801, risc, romp, rios http://www.garlic.com/~lynn/lhwemail.html#801 there was huge amount of communication group FUD about my 3725 numbers used in comparison/presentation ... which I pulled directly from HONE 3725 configurator ... HONE configurations (world-wide virtual machine based online salesmarketing) were used by IBM salesmarketing for configuring hardware. In the case of 3725 configurator ... performance modeling had official communication group sanction. misc. past posts mentioning HONE http://www.garlic.com/~lynn/subtopic.html#hone one of my hobbies was enhanced production operating systems for internal datacenters ... HONE was (also) one of my long time customers since cp67/cms days in the early 70s. that hone was actually virtual machine based was obfuscated from most
zEC12, and previous generations, why? type question - GPU computing.
I guess that I should preface this with another question. Does anybody use a z for heavy numeric computation anymore? Or has that all gone to Intel and Power boxes? Why is that? If it is because the z architecture is not good at numeric computation, I have a question. The internals of the z has used the PCIe bus for some time. Wouldn't that imply that it would be at least theoretically possible to plug in any PCIe card? OK, I understand that there would need to be some way to access it. But, in conjunction with the previous question about computational ability, couldn't this be used for GPU computation. It seems to be that the really heavy computation is being moved from the CPU (Intel) onto the GPU (AMD or Nvidia graphics processor). Would it make any monetary sense to enable GPU computation on a z? Long ago, there were vector instructions. Why not some sort of interface instruction(s) which allow loading a GPU processing program to be placed in a GPU on a PCIe card and then have some way to suspend the unit of work until the GPU computation is complete? Likewise have an API to request access to a GPU which would suspend the unit of work until a GPU was available and assign the GPU to the unit of work? Weird thoughts from a weird person. -- John McKown Systems Engineer IV IT Administrative Services Group HealthMarkets(r) 9151 Boulevard 26 * N. Richland Hills * TX 76010 (817) 255-3225 phone * john.mck...@healthmarkets.com * www.HealthMarkets.com Confidentiality Notice: This e-mail message may contain confidential or proprietary information. If you are not the intended recipient, please contact the sender by reply e-mail and destroy all copies of the original message. HealthMarkets(r) is the brand name for products underwritten and issued by the insurance subsidiaries of HealthMarkets, Inc. -The Chesapeake Life Insurance Company(r), Mid-West National Life Insurance Company of TennesseeSM and The MEGA Life and Health Insurance Company.SM -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
In a6b9336cdb62bb46b9f8708e686a7ea0115baa1...@nrhmms8p02.uicnrh.dom, on 09/05/2012 at 11:45 AM, McKown, John john.mck...@healthmarkets.com said: If it is because the z architecture is not good at numeric computation, The z architecture is fine for numeric computations. The problem is that the implementation is competing with processors manufactured in bulk. If IBM could sell millions of z boxen then they'd be able to cut the price dramatically. I've always wondered what would have happened had IBM used a 370 instruction set on the PC instead of Intel. -- Shmuel (Seymour J.) Metz, SysProg and JOAT Atid/2http://patriot.net/~shmuel We don't care. We don't have to care, we're Congress. (S877: The Shut up and Eat Your spam act of 2003) -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
On Wed, Sep 5, 2012 at 2:28 PM, Shmuel Metz (Seymour J.) shmuel+...@patriot.net wrote: In a6b9336cdb62bb46b9f8708e686a7ea0115baa1...@nrhmms8p02.uicnrh.dom, on 09/05/2012 at 11:45 AM, McKown, John john.mck...@healthmarkets.com said: If it is because the z architecture is not good at numeric computation, The z architecture is fine for numeric computations. The problem is that the implementation is competing with processors manufactured in bulk. If IBM could sell millions of z boxen then they'd be able to cut the price dramatically. I've always wondered what would have happened had IBM used a 370 instruction set on the PC instead of Intel. 16MB ought to be enough for anybody? :-) Since IBM wasn't manufacturing the chips, of course that wasn't even on the table, but it's still a VERY interesting Gedankenexperiment... -- zMan -- I've got a mainframe and I'm not afraid to use it -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
With IBM's acquisition of SPSS several years ago the recent acquisition of Netezza (for use as an attached processor for computational workloads on zSeries), IBM's z/Series intentions seem to have changed. After the AS (Application System) disaster (early eighties, great demo, not scalable, ADRS based if I recall), I hope the performance concerns are addressed. Even the DB2 folk no longer accept a performance hit with a new release (more code features take more resources was a mantra at IDUG for years, finally falling flat with V8.) In particular, with the minimization of locking, data above the bar, increased use of zIIP general performance improvements, analytics with DB2 on zSeries might be cost effective for big data in a shared workload environment. See (unfortunately marketing oriented): http://www.clabbyanalytics.com/uploads/zBAfinalfinalfinal.pdf and http://berniespang.com/2012/06/08/clients-chose-ibm-system-z-for-analytics-over-teradata-and-oracle-exadata/ It would have been interesting if they had put something like this together for the 2010 census data in the way SAS did for the 1980 data, but there's plenty more data sources against which these marketing claims will soon be tested. --Original Message- From: IBM Mainframe Discussion List [mailto:IBM-MAIN@LISTSERV.UA.EDU] On Behalf Of Mark Post Sent: Wednesday, September 05, 2012 1:07 PM To: IBM-MAIN@LISTSERV.UA.EDU Subject: Re: zEC12, and previous generations, why? type question - GPU computing. On 9/5/2012 at 12:45 PM, McKown, John john.mck...@healthmarkets.com wrote: I guess that I should preface this with another question. Does anybody use a z for heavy numeric computation anymore? Or has that all gone to Intel and Power boxes? Why is that? If it is because the z architecture is not good at numeric computation, I have a question. As has been pointed out in another thread here, the dollar cost per instruction is much higher on System z than other architectures. So for purely computational workloads, although System z may have a faster CPU than the other architectures, it costs more for the same amount of computation. A lot of high performance computing is restartable in that if a computation node fails, starting that piece of work over from the beginning isn't hard. Most of the qualities that are built into System z aren't needed for that type of work, so no need to spend the big bucks for it. Mark Post -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
On 5 September 2012 17:17, zMan zedgarhoo...@gmail.com wrote: On Wed, Sep 5, 2012 at 2:28 PM, Shmuel Metz (Seymour J.) shmuel+...@patriot.net wrote: The z architecture is fine for numeric computations. The problem is that the implementation is competing with processors manufactured in bulk. If IBM could sell millions of z boxen then they'd be able to cut the price dramatically. I've always wondered what would have happened had IBM used a 370 instruction set on the PC instead of Intel. 16MB ought to be enough for anybody? :-) Since IBM wasn't manufacturing the chips, of course that wasn't even on the table, but it's still a VERY interesting Gedankenexperiment... There *was* a single-chip 370 produced by someone in the late 70s - a 168i. I think it was a university or research institute, but not IBM. I'm not finding anything on Google with a casual search, but things like this are easily overwhelmed. Tony H. -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
t...@harminc.net (Tony Harminc) writes: There *was* a single-chip 370 produced by someone in the late 70s - a 168i. I think it was a university or research institute, but not IBM. I'm not finding anything on Google with a casual search, but things like this are easily overwhelmed. SLAC did 168E ... basically could run problem state fortran at 168 speed ... for data collection/reduction along the accelerator line ... long ways from single chip. 168 had been four circuits per chip, 3033 initially was 168 logic initially layed out on something like 40circuits per chip ... but just using 4 circuits in each chip ... getting 20% chip improvement. during development there was some rework of part of the 168 logic to make better use of higher chip density ... get 3033 up to 50% faster than 168. http://www.jfsowa.com/computer/memo125.htm i've frequently claimed that John Cocke's 801/risc was reaction to horrible complexity of (failed) FS effort ... initially simplified (aka reduced instruction set) for single chip implementation ... and then later simplified instructions that were all single machine cycle. 360s, 370s, etc ... have been microcode implemented on variety of other kinds of engines. circa 1980 there was an effort to replace the wide variety of internet microprocessors used for controllers, lowmid range 370s, the planned as/400 replacement for s/38, etc ... all with 801/risc Iliad chips. For various reasons the efforts floundered and they went back to doing custom processor implementations. It took another couple decades ... but lots of stuff is now risc in one way or another (as previously mentioned the past couple generations of i86 are risc processors with hardware layer translating i86 to risc micro-ops). In the mid-80s, there was a 3-chip 370 that ran at 168 speed from Boeblingen called ROMAN. I had a project/proposal to pack arbitrary mix of large numbers of ROMAN and Iliad chips in the same rack ... with large number of racks (sort of precursor to latest rack announcement). a couple old email refs: http://www.garlic.com/~lynn/2011b.html#email850314 http://www.garlic.com/~lynn/2007d.html#email850315 I had also been working with NSF on what was to become NSFNET backbone (i.e. tcp/ip is the technology basis for the modern internet, NSFNET backbone was the operational basis for the modern internet, and CIX was business basis for the modern internet). Above refs. that I had to find standin for me doing presentation to head of NSF ... because of rack cluster effort meeting. Lots of low mid-range clone 370 vendors were starting to spring up all over the place. Somebody in Siemens germany had somehow acquired a proprietary ROMAN document ... and was trying to get it returned to IBM with all fingerprints removed. He sent it to somebody at Amdahl in silicon valley ... who arranged to hand it over to me. Other trivia ... SLAC had hosted the monthly IBM user group meetings (BAYBUNCH) and was also the first webserver outside of europe. http://www.slac.stanford.edu/history/earlyweb/history.shtml and from long ago and far away ... mentions slac/cern 168E ... having become 3081E (3mips to 14mips). Date: Fri, 7 Jul 89 10:52:39 CDT From: wheeler Subject: requirements task force Note that both DEC and Apollo (along with hp) are heavily into distributed environment, heterogeneous network/system/enterprise management, and networks. Note that heterogenous means more than OSI, TCP/IP, UNIX, etc ... it means interoperability between all of them along with DECNET, VMS, XNS, etc. Apollo's FDDI group is heavily involved in XTP and the former manager of the Apollo FDDI group (they've been active for some time and spending lot of time optimizing performance of high thruput adapters) left Apollo and formed synernetics (he was involved with XTP at Apollo and synernetics is a XTP/TAB member). They are working on initial cut of FDDI station management (SMT ... and have been out talking to a number of groups, including IBM ... I also believe he has even had contacts with Andy). Distributed 370s have a hard time keeping up. Way back in history (someplace), I spent a lot of time up at SLAC (there is tight coupling between SLAC and CERN). At that time SLAC was doing the 168E, a bit-slice processor that would run standard 370 Fortran programs. They technology has been improved and now CERN SLAC are calling it a 3081E (i.e. the processing power of 3081). The design was to have one of these processors at each of the (large number of) data collection points. Something that will be competing with this will be the 370 simulator that xx has done for the SUN4. He currently has VM/370 running at about 168 thruput on the old SUN4 (big register memory are super for large integrated applications but state switch overhead can be heavy/horrible ... something that we are currently grappling with ... some possibilities exist for pipelining/overlapping state switch like yy is doing with Vector Buffer architecture).
Re: zEC12, and previous generations, why? type question - GPU computing.
Yes, there are organizations that use zEnterprise servers for heavy numeric computation. Like decimal floating point. Cryptography is another excellent example. And you can buy optional CryptoExpress adapters if you want to augment the excellent capabilities found in every machine. You can also buy the optional zBladeCenter Extension (zBX) if you want to add DataPower accelerators, Power blades, and/or X86 blades. You can also add an optional IBM DB2 Analytics Accelerator, to boost many types of DB2 queries. So we're way ahead of you, John. ;-) I think the simple answer is that it depends what you optimize for in designing a server processor (or complex). But IBM has broken a lot of rules already about which server should do what, and I predict more rules will be broken. With respect to the 370-on-a-chip, IBM sort of did that with the 1975 introduction of the IBM 5100 Portable Computer starting at $8,975 (1975 dollars), although it was for a relatively narrow initial purpose (to get APL running). The 5100 sold reasonably well from what I've read, but I think there were three basic problems which prevented it from becoming a blockbuster: 1. The price was not low enough for mass market appeal. (Apple had a similar problem with the Lisa in the early 1980s.) 2. The software selection didn't exactly hit the mark, although it was a good try for the time. (IBM learned the value of software somewhat later in its evolution but not in time for the 1981 IBM PC.) 3. It probably didn't have the right third party marketing and distribution channels. With some very notable exceptions, like typewriters, at that time IBM would have had some challenges with this type of product. Keep in mind that for 1975 this was absolutely amazing technology, but amazing technology required some expense. Being early is pricey. If the 5100 debuted in, say, 1977 or 1978, it would have still been well timed but could have dramatically reduced the chip and board count. I also think the small built-in monitor could have been sacrified (at least as an option) in favor of a display port of some kind -- ideally RF for TV hookup. And IBM might have gone with a diskette drive for storage -- the 5100 was too early for the 5.25 inch drive, which debuted in 1976. Finally, if IBM had provided a little more guidance on the 370 subset instruction set they implemented, software developers could have taken over from there. So I think the 5100 could have been a nice 5110 by tweaking the recipe a bit. But history didn't happen that way. IBM had some success with the System/4 Pi avionics processors which are descended from System/360. Timothy Sipples Consulting Enterprise IT Architect (Based in Singapore) E-Mail: sipp...@sg.ibm.com -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
Re: zEC12, and previous generations, why? type question - GPU computing.
I believe IBM produced a pc with a 370 to run VM on a PC. Merrill Lynch had one. Somewhere in the late 80's I believe. On Thu, Sep 6, 2012 at 1:52 AM, Timothy Sipples1 sipp...@sg.ibm.com wrote: Yes, there are organizations that use zEnterprise servers for heavy numeric computation. Like decimal floating point. Cryptography is another excellent example. And you can buy optional CryptoExpress adapters if you want to augment the excellent capabilities found in every machine. You can also buy the optional zBladeCenter Extension (zBX) if you want to add DataPower accelerators, Power blades, and/or X86 blades. You can also add an optional IBM DB2 Analytics Accelerator, to boost many types of DB2 queries. So we're way ahead of you, John. ;-) I think the simple answer is that it depends what you optimize for in designing a server processor (or complex). But IBM has broken a lot of rules already about which server should do what, and I predict more rules will be broken. With respect to the 370-on-a-chip, IBM sort of did that with the 1975 introduction of the IBM 5100 Portable Computer starting at $8,975 (1975 dollars), although it was for a relatively narrow initial purpose (to get APL running). The 5100 sold reasonably well from what I've read, but I think there were three basic problems which prevented it from becoming a blockbuster: 1. The price was not low enough for mass market appeal. (Apple had a similar problem with the Lisa in the early 1980s.) 2. The software selection didn't exactly hit the mark, although it was a good try for the time. (IBM learned the value of software somewhat later in its evolution but not in time for the 1981 IBM PC.) 3. It probably didn't have the right third party marketing and distribution channels. With some very notable exceptions, like typewriters, at that time IBM would have had some challenges with this type of product. Keep in mind that for 1975 this was absolutely amazing technology, but amazing technology required some expense. Being early is pricey. If the 5100 debuted in, say, 1977 or 1978, it would have still been well timed but could have dramatically reduced the chip and board count. I also think the small built-in monitor could have been sacrified (at least as an option) in favor of a display port of some kind -- ideally RF for TV hookup. And IBM might have gone with a diskette drive for storage -- the 5100 was too early for the 5.25 inch drive, which debuted in 1976. Finally, if IBM had provided a little more guidance on the 370 subset instruction set they implemented, software developers could have taken over from there. So I think the 5100 could have been a nice 5110 by tweaking the recipe a bit. But history didn't happen that way. IBM had some success with the System/4 Pi avionics processors which are descended from System/360. Timothy Sipples Consulting Enterprise IT Architect (Based in Singapore) E-Mail: sipp...@sg.ibm.com -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN -- George Henke (C) 845 401 5614 -- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN