There have been people who have personally requested me to continue this
discussion on the list, while the admin forbades such. Since there has
been certain misunderstandings wrto whether the general people should
stick to transputers or move over to NOWs... I post this to the list in
common interest. the other part of the ParProgWar has been taken offline
on my behalf.
Regards
Shourya
It was on Thu, 27 Jan 2000, Shanker R Swaminathan typed in this..
>*>> Shanker :
>*>> >*>Given Transputers which are sufficiently fast- I disagree. A Transputer
>*>> >*>system is closer to a closely coupled system than a NOW. You CANNOT
>*>beat the
>*>> >*>speed of a Closely coupled system with a loosely coupled system( With
>*>> >*>comparable hardware) .
>*>>
>*>> Shourya:
>*>> Logically, you are correct. But only logically. The catch is that it is
>*>> foolish to compare Transputers with modern day COTS hardware..because
>*>> even the fastest of transputers do not match the average
>*>> available hardware.. think of your own machine which I believe is a
>*>> Celeron "overclocked" 333. I do not have the specs right now but shall
>*>> comply on request.So you see, it is not possible to get transputers which
>*>> are "sufficiently" fast.. they have been decommissioned . Try getting
>*>> some to India and you will know exactly how sensitive these imports are.
>*>
>*>Did you see the Line- 'comparable Hardware'?
>*>
You missed the gist of my statements. That was exactly the point. There
is NO comparable h/w wrto transputers. ie Transputers are simply not
produced so that they can be competitive with modern COTS h/w. For eg, a
T818 has a 5Mhz clock with a quadder. how about that !!!!!! You can't say
that given comparable fire power , the IAF would annilhilate the USAF..
we don't _have_ anything beyond Sukhoi's :)
>*>> >*> They can run a far larger number of algorithms
than a >*>> >*>NOW . >*>>
>*>> Huh ?? Do you have any examples in mind ? I am not aware of any .
>*>> Please provide evidence to hold your ground.
>*>
>*>You are more than Welcome- This is one topic i have to give an exam on - If
>*>you drop by I can give you Half my ACA syllabus!
Please comply. At least with three .
>*>
>*>> Infact there are systems like PVM, languages like Paralallaxis which do
>*>> provide scope for 100 % MIMD design.
>*>>
>*>> On the contrary, I could provide you with a 1001 examples as to
>*>> situations which can be handled by NOW but not on Transputers.
>*>
>*>Show that please- My textbooks and teachers tell me the very opposite !
>*>
For starters, ever tried to do a dynamic linked implementation list of
say distributed sorting algorithms on transputers ? Or agents ?
>*>> To top it all, the theoritical model behind transputers message passing
>*>> paradigm is C.A.R. hoare's seminal work on CSP, which is marvellously
>*>> simple (after shedding of the mathematics) as a CONCEPT, and hence the
>*>
>*>Once you remove the mathematics from it- it loses almost everything! Try
>*>showing any report without a solid mathematical base and justification !
When you studied cache memory in BCSE, did you work on the mathematical
analysis of Demming, '68 ? PRobably no, my stand is clear. You can look
at CSP (and many other theories, like the big bang) at a conceptual level
without getting frittered by the underlying maths. This is not to say
that you can get hold of the entire theory or that the eloquence of math
is simply an icing.. but to UNDERSTAND you do not need to get past
elem. maths.
>*>> result, the transputer , too is very simple at the programming level of
>*>> abstraction. You do not need to go into greater details to write parallel
>*>> code.. just like you do not need to know the charge of an elctron which
>*>> brightens up your screen, to write a GUI.
>*>
>*>Believe me- After the analysis of Parellel Algos and their intricacies which
>*>I/we had to _Undergo_ Not one person in my class was in a position to
>*>walk straight!HOW did get the Idea that writing parellel code is simple? You
>*>could have dropped in one of our ACA classes while they were going on-
>*>Parellel Algorithims can look deceptively simple and elegant- their
>*>intricacies are rarely apparent at a cursory look.
>*>
I never said writing parallel code was easy. No self respecting sane
person would say that and that includes me :) All that i said was that
writing parallel code on OCCAM was easy compared to implementations on
NOW, specifically because OCCAM has explicit language constructs to
support parallelism which ANSI C does not have.. even FORTRN 95 does not
do a good job.
>*>> In fact, the gentle learning curve in transputers and the touching
>*>> simplicity of the model is what that still keeps Transputers and OCCAM
>*>> going. So please do not spread FUD about the complexity of transputers,
>*>> just because it is not taught in courses offerred by the roadside
>*>> mushrooming computer institutes.
>*>
>*>On the one hand you are urging everyone to shift to NOWs and on the other
>*>hand you extoll transputers- Make up your mind- You seem confused!
>*>
It is you who looks like clutching to the last straws to hold on, relying
on twisted semantics to defend yourself. The need to shif to NOWs from
transputers is NOT ease of programming on NOWs... or as you would put it
complexity.
The reasons are :
o Cost (the only count on which you were correct )
o Speed
o COTS benefits and the fact that you have US sanctions against
Transputer imports.
o Better programming environments. Try writing
CVS for transputers or play around with idebug.. you'll
know what i am trying to get at.
Least of all , programming ease. OCCAM is much more suited to parallel
stuff than C/Fortran what have you.. In fact we are trying to work on
certain workarounds and tradeoffs to have a haronic combo of NOWs
and Tranps which I cannot diclose right now because I do not have the
auth.
>*>> >*> Given a choice
>*>> >*>between an 8 way system and 8 machines - I would pick the 8 way system
>*>> >*>anyday ( As far as speed goes)
>*>>
>*>> I seriously would reccomend you NOT to.. if the closely coupled system is
>*>> a T8 series . Given my experiences (I have done a modest amount of
>*>> performance evaluation and characteristic determination on Coupled
>*>> systems , both Transp and NOWs), it is far more efficient to stick to a
>*>> NOW , if SPEED is your concern.Otherwise, you will be grossly
>*>> disappointed.
>*>>
>*>Hey- You can't seriously seriously mean what you say - Take a 4 proc system
>*>( a True Multiproc board please with a real interconnection network -a
>*>hypercube/shuffle network etc .) and
>*>take 4 m/c on a network and run - um just any app.( Not just explicitlly
>*>parellel ones) and let me know- ( AGAIN with comparable hardware.). Don't
>*>spread misinformation. ( take a look at Hwang and Briggs- Pg 460).
>*>
I never said I wanted to do a fight with IBM SP2 or Paragon MPP. Stick to
your point.. transputers. And what do you mean when you say any app ??
You want to run a non threaded version of Lotus 1-2-3 1988 on a Origin
2000 ? How do you intend to expolit a NOW situation where you do not have
parallelism in your code , huh :? Come up with proper testing plans..
that is VERY important. Try something simpler like sorting 5 lac numbers
using similar algos on a linear network of NOWs and Transps and get back
with the results. OK, you dont have to do it, I 'll tell you what
happens, the transp loses :( Bad news. Thats why we want to migrate to
NOWs.
Regards
Shourya
Regards
shourya
_______________________________________________________________
Shourya Sarcar <[EMAIL PROTECTED]> <Tel:91-033-4710477>
Department of Computer Science and Engineering Jadavpur University
Calcutta, India 700 032
All the world's a stage..
And I am acting tonight
C - the difference : http://www.eskimo.com/~scs/C-faq/top.html
--
To unsubscribe, send mail to [EMAIL PROTECTED] with the body
"unsubscribe ilug-cal" and an empty subject line.
FAQ: http://www.ilug-cal.org/help/faq_list.html
