Crisis in HPC Discussion - Nick Maclaren, University of Cambridge

From: nmm1@cus.cam.ac.uk (Nick Maclaren)
Newsgroups: uk.org.epsrc.hpc.discussion
Subject: Re: Superscalar considered harmful
Date: 7 Oct 1995 17:32:55 GMT
Organization: University of Cambridge, England
Distribution: uk

In article 43 Peter Welch  wrote:
>
>Maybe there is something wrong with the concept of superscalar
>superpipelined processors?  What they do is extract a very limited
>amount of concurrency at the instruction level (2 to 5 instructions
>per cycle, if you are lucky with their interdependencies).  It's
>difficult to find sufficient inter-instruction indpendencies to
>keep the piepline full a decent percentage of the time and as soon
>as something goes wrong (like a cache miss), the thing stalls and
>you have to flush the pipeline and start again.  ...

[To people who have heard this before, please accept my apologies.]

One really interesting thing about this is that it has happened before! IBM built the 360/91 and 370/195 to execute up to a dozen instructions in a tight pipeline (effectively superscalar). They had the devil's own job with the compilers etc. keeping them up to speed, and they never did work entirely satisfactorily. When IBM designed the 308x series, they went back to much shorter pipelines :-)

Yes, compiler technology has moved on, but not that much. Instruction rescheduling was there as OPTIMIZE(4) in Fortran H Extended Enhanced back in the 1970s, for the 370/195 people only. And as for the 360/91 pipeline drain instruction needed to ensure the run-time system worked, words fail me :-)

I cannot see that simple 'superscalar' architectures will work now, or ever, while they are expected to perform efficiently on a single stream of 'third generation' code. I don't know what the solution should be, but I do know that it has to be fairly radical.

Nick Maclaren,
University of Cambridge Computer Laboratory,
New Museums Site, Pembroke Street, Cambridge CB2 3QG, England.
Email: nmm1@cam.ac.uk
Tel.: +44 1223 334761 Fax: +44 1223 334679