C506634@UMCVMB.MISSOURI.EDU ("Eric Edwards") (03/21/90)
Large, shared, systems seem to be good for 1) large, compute intensive tasks 2) large, IO intensive tasks 3) tasks in which data must be shared with other users. And are poor at 1) User interaction 2) small jobs that need to have a predictable completion time. Wheras, single user systems excel at the second group and come up short on the first. So..... Why not combine the two. Use smaller, cheaper, and less capable micros to provide the user interface and run small jobs. When the going gets tough they could send the job to central supercomputer. A smaller central machine would be needed then would be required with a strictly centralized system. Not only do the actual cycles not goint to supporting the user interface but also becuase the system could run close to capasity. Response time is not much of an issue if the jobs are slow anyway. You still waste mips on the micros but since they are smaller machines and probably not state of the art machines the mips they waste cost less. Also, the communication links don't need as much capicity. |****/ Eric Edwards (prefered) | Of course ada is good for large ^****| |***/ c506634@umcvmb.missouri.edu | programs. All Ada programs are ^***| |***^ c506634@umcvmb.bitnet | large! /***| |****^ csh> ls -l Ada/* => -rw-r--r-- 1 c170234 85992 Ada/HelloWorld /****|
xxremak@csduts1.lerc.nasa.gov (David A. Remaklus) (03/21/90)
In article <9003210649.AA06750@jade.berkeley.edu> C506634@UMCVMB.MISSOURI.EDU ("Eric Edwards") writes: >Large, shared, systems seem to be good for >1) large, compute intensive tasks >2) large, IO intensive tasks >3) tasks in which data must be shared with other users. >And are poor at >1) User interaction >2) small jobs that need to have a predictable completion time. > >Wheras, single user systems excel at the second group and come up short on the >first. So..... Why not combine the two. > Checkout TCF (Transparent Computing Facility) proposed by IBM for the AIX platforms. While I don't have the specifics myself, they seem to have created a "seamless" path from the PS/2 through the 3090. Essentially, they create shadowed files where multiple execution images appear under the same name. To illustrate, consider the program /bin/sh. With TCF, there is a common NFS like file system between the various platforms (PS/2, RT, RS600, 3090, etc). Through a normal directory list, a user would only see a single /bin/sh, when actually there could be an executable for each. Which ever platform you were running on is the executable chosen. The really neat feature is the case where an executable does not exist for your platform. In that case, an executable for a different platform is chosen and automatically executed there without you having logged into the machine. This is best seen in the case of a.out files where you may be logged into a PS/2 where you edit, compile and link, but what you created was perhaps a 3090 executable. When you run it, it executes on the 3090 without your being aware that it did. IBM did a song and dance on this thing at a recent SHARE meeting that I think was last year (89). The presentation material ought to be in the proceedings. The really sad thing is, that true to form, IBM doesn't appear willing to release the specs for TCF into the public domain. Its a real shame as I have never seen anything like it since. Dave R. -- David A. Remaklus NASA Lewis Research Center Cleveland, Ohio xxremak@csduts1.lerc.nasa.gov
merriman@ccavax.camb.com (03/22/90)
In article <9003210649.AA06750@jade.berkeley.edu>, C506634@UMCVMB.MISSOURI.EDU ("Eric Edwards") writes: > . . . > > Use smaller, cheaper, and less capable micros to provide the user interface and > run small jobs. When the going gets tough they could send the job to central > supercomputer. A smaller central machine would be needed then would be > required with a strictly centralized system. Not only do the actual cycles not > goint to supporting the user interface but also becuase the system could run > close to capasity. Response time is not much of an issue if the jobs are slow > anyway. You still waste mips on the micros but since they are smaller machines > and probably not state of the art machines the mips they waste cost less. Sounds like a Local Area Vax Cluster with DECnet/DOS for the PC-type machines. > Also, the communication links don't need as much capicity. It turns out that anything less that Ethernet is a drag because of file transfer times.
usenet@nlm-mcs.arpa (usenet news poster) (03/22/90)
What you want on your desk is enough computing to support your environment (windows, editors etc.) because that is where keystroke response time matters. It is also the computing you use in a continuous mode (as opposed to bursts). What you want the net for is access to specialized processors (database servers, floating point engines, parallel nets etc.) because these are the sorts of things you use in bursts and can't afford to put on your desk anyway. There are lots of systems going in this direction (Vax clusters, IBM's trans- parent computing interface, diskless workstation local networks, +-X-terminal local nets). The problem is standardization. You don't get the benefit of the net if your local processor can't talk effectively with the resource you want. David States (standard discalimer, views are my own only)
Ewing@yale.edu (Martin Ewing) (03/22/90)
References:<9003210649.AA06750@jade.berkeley.edu> <1990Mar21.142014.20419@eagle.lerc.nasa.gov> Organization: Yale Computing & Information Systems In article <1990Mar21.142014.20419@eagle.lerc.nasa.gov> xxremak@csduts1.lerc.nasa.gov (David A. Remaklus) writes: > Checkout TCF (Transparent Computing Facility) proposed by IBM for the > AIX platforms. While I don't have the specifics myself, they seem to > have created a "seamless" path from the PS/2 through the 3090. I am not sure whether we need/want "seamless". A hierarchical situation is attractive, however. Central/decentral doesn't have to be either-or. I really don't want to run Excel or Lotus 1-2-3 (/M notwithstanding) on a central machine. Responsiveness and privacy favor my local Mac. For a big simulation program, or the company payroll, you go to the mainframe. Another factor is cost and flexibility of software. I can go out and buy package X for my little system for $100, while for a big system you're talking $10K and a lot of persuasion. ------------------------------------------ Martin Ewing Yale C&IS Science and Engineering Computing Facility 203-432-4243, Ewing@Yale.Edu
jkrueger@dgis.dtic.dla.mil (Jon) (03/23/90)
xxremak@csduts1.lerc.nasa.gov (David A. Remaklus) writes: >Checkout TCF (Transparent Computing Facility) proposed by IBM for the >AIX platforms. Oh, I've proposed it for all computers from the eniac to the MIPS R3000. Since neither I nor IBM has shipped any product, let's propose 'till the cows come home. -- Jon -- Jonathan Krueger jkrueger@dtic.dla.mil uunet!dgis!jkrueger The Philip Morris Companies, Inc: without question the strongest and best argument for an anti-flag-waving amendment.