jackv@turnkey.gryphon.COM (Jack F. Vogel) (12/17/89)
In article <289@dgsi.UUCP> brian@dgsi.UUCP (Brian Kelley/10000000) writes: [ my opinions on 3090 being a fast unix platform deleted...] >That's not what I heard. At least the word on the street is it's really slow. >Well not *REALLY* slow, but only about twice as fast as a $10,000 DEC Station >3100. I don't know what the source of this opinion was or how the comparison was made. I suspect there are some isolated tests or activities that might produce something like this but I also think it would be misleading. One needs to evaluate performance in a more system-oriented way than how many dhrystones the cpu can crank, at least that's my opinion. Anyone familiar with 370 systems knows that where they shine is batch I/O processing power. I doubt that the Dec Station could match 32 or more channels pumping 5Meg/sec to and from DASD. Also the Dec or other workstation has main storage in the 10's of Megabytes, the 3090 can have 512 Meg (It is fairly common for systems to have 256). The 3090 also has something called extended storage which is slower ram used as a paging device (not that you'd page that often with that much real storage :-}), this can extend into the gigabytes. You can therefore run processes on these beasts that you could never even dream of on your workstation. For example, it is common these days to do the graphic VLSI circuit design on a workstation but this is just the graphic frontend, when it comes to the real processing it is necessary to have the mainframe class system to accomplish it. That's why Intel designed the 486 running AIX on the 3090, it wasn't due to liking Big Blue, they just had to have a system with this kind of power. Well enough of my rambling, you get the idea. > I've also heard that IBM will not let anyone disclose performance >figures for the system. How about some dhrystones? Seriously, are there >industry standard benchmark results available from somewhere? How much >does a typical 3090-600 cost? IBM marketing will, as far as I know, be releasing performance data for AIX/370 when the product becomes generally available. I think its normal procedure for a customer given an early release of a product to be under non-disclosure on things like performance due to the very fact that it is pre-released code, so there is nothing unusual about this. On the question of how much a 3090 costs, well you'll have to talk to your IBM rep about that, although I believe its one of those kinds of products where if you have to ask about the price you can't afford it :-} :-}! Disclaimer: These opinions are mine, not necessarily Locus' or IBM's. -- Jack F. Vogel jackv@seas.ucla.edu AIX Technical Support - or - Locus Computing Corp. jackv@ifs.umich.edu
dboyes@rice.edu (David Boyes) (12/17/89)
In article <6403@turnkey.gryphon.COM> jackv@turnkey.gryphon.COM writes: >In article <289@dgsi.UUCP> brian@dgsi.UUCP (Brian Kelley/10000000) writes: >[ my opinions on 3090 being a fast unix platform deleted...] >>[Jack compares AIX/370 to about double a decstation 3100] >I don't know what the source of this opinion was or how the comparison was >made. I suspect there are some isolated tests or activities that might >produce something like this but I also think it would be misleading. One >needs to evaluate performance in a more system-oriented way than how many >dhrystones the cpu can crank, at least that's my opinion. Anyone familiar >with 370 systems knows that where they shine is batch I/O processing power. >I doubt that the Dec Station could match 32 or more channels pumping 5Meg/sec >to and from DASD. Absolutely. One of the lingering problems with implementing high-performance Unix machines for business applications is their immensely poor disk performance for database-style applications that commonly do block transfers of records from DASD to main storage. IBM can claim transaction rates of over 200/sec (using MVS/CICS) because of the large quantity of hardware support for that type of application. How well AIX/370 can map the traditional Unix character-based I/O system to the IBM architecture remains to be seen, but if it works out at all, you can expect to see Unix finally make a real dent in the business arena, and not just because corporations are having trouble finding programmers with MVS experience. (As a side note, consider IBM's latest DASD announcement. 22.6 gigabyes in a roughly 7 foot square box? That kind of disk storage makes large multiuser Unix systems an interesting possibility again.) >10's of Megabytes, the 3090 can have 512 Meg (It is fairly common for systems >to have 256). Not all that common -- 64M is more common, due to the high cost of memory -- but 256M is not unreasonable. >You can therefore >run processes on these beasts that you could never even dream of on your >workstation. As I understand it, AIX runs under VM/XA. This gives you 990M virtual machines, as well as the ability to support multiple AIX systems on the same iron at the same time, each with the full resources of the machine. >> I've also heard that IBM will not let anyone disclose performance >>figures for the system. How about some dhrystones? Seriously, are there >>industry standard benchmark results available from somewhere? >IBM marketing will, as far as I know, be releasing performance data for >AIX/370 when the product becomes generally available. I think its normal >procedure for a customer given an early release of a product to be under >non-disclosure on things like performance due to the very fact that it is >pre-released code, so there is nothing unusual about this. In most cases, you're not even allowed to ackowledge the *existence* of the product, let alone performance data on it. Anything you read in the trade rags at this point is probably pure speculation until the product hits General Availability. Time will tell. IBM doesn't until they're ready. >> How much >>does a typical 3090-600 cost? A lot. A reasonably configured 3090/600S (with enough DASD, tape drives and memory to be useful) will probably go for more than $5-6 million. That doesn't count air conditioning, chilled water supply, or software licensing. Still, I agree with the Locus folks; if you need that kind of compute power, it's probably a competitive alternative to a small Cray, and you can still run other IBM stuff on it...8-) >Jack F. Vogel jackv@seas.ucla.edu -- David Boyes "... no love was left; All Earth was but one thought - and dboyes@rice.edu that was death Immediate and inglorious; and the pang of of famine fed upon all entrails - men Died and their bones were tombless as their flesh ..." - Lord Byron
schwartz@psuvax1.cs.psu.edu (Scott Schwartz) (12/17/89)
In article <6403@turnkey.gryphon.COM> jackv@turnkey.gryphon.COM writes: >workstation. For example, it is common these days to do the graphic VLSI >circuit design on a workstation but this is just the graphic frontend, when >it comes to the real processing it is necessary to have the mainframe class >system to accomplish it. Around here we do lots of VLSI design on old Sun4s, because it's faster than on our 3090/600E. As you remark, IBM has wonderful thruput and lots of memory, but for scalar arithmatic the "Attack of the Killer Micros" has already been won. >That's why Intel designed the 486 running AIX on >the 3090, it wasn't due to liking Big Blue, they just had to have a system with >this kind of power. Well enough of my rambling, you get the idea. Ask someone at Motorola about how the 88000 was done. -- Scott Schwartz <schwartz@shire.cs.psu.edu> "More mips; cheaper mips; never too many." -- John Mashey
brian@la.locus.com) (12/17/89)
In article <1989Dec17.062913.28580@psuvax1.cs.psu.edu> schwartz@psuvax1.cs.psu.edu (Scott Schwartz) writes:
]Around here we do lots of VLSI design on old Sun4s, because it's faster
]than on our 3090/600E. As you remark, IBM has wonderful thruput and
]lots of memory, but for scalar arithmatic the "Attack of the Killer
]Micros" has already been won.
]
]>That's why Intel designed the 486 running AIX on
]>the 3090, it wasn't due to liking Big Blue, they just had to have a system with
]>this kind of power. Well enough of my rambling, you get the idea.
]
]Ask someone at Motorola about how the 88000 was done.
Actually, it wasn't so much the design of the 486 that was done
on the 3090/600, but rather the simulations. Simulation of VLSI, even
at switch level takes up an enourmous amount of memory (virtual) and
at least as an enourmous amount of CPU. For this the 3090 is good.
Not to mention the huge output files those Intel folks liked to create.
As I understand it, the DESIGN phase WAS done on Sun workstations.lijewski@batcomputer.tn.cornell.edu (Mike Lijewski) (12/20/89)
In article <3819@brazos.Rice.edu> dboyes@brazos.rice.edu (David Boyes) writes: > >As I understand it, AIX runs under VM/XA. This gives you 990M >virtual machines, as well as the ability to support multiple AIX >systems on the same iron at the same time, each with the full >resources of the machine. > AIX/370 does run under VM/XA, but as I understand it the theoretical upper limit to process size is currently 764M or thereabouts. This is still theoretical, since it is my understanding that AIX as a VM does not autolog above 512M or thereabouts. Would one of you developers care to comment on this and/or set me straight as the case may be? -- Mike Lijewski (H)607/277-7623 (W)607/255-0539 (desk)607/255-2960 Cornell National Supercomputer Facility ARPA: mjlx@cornellf.tn.cornell.edu BITNET: mjlx@cornellf.bitnet SMAIL: 1122 Ellis Hollow Rd. Ithaca, NY 14850