🚀 go-pugleaf

In article <PCG.92Dec23150744@decb.aber.ac.uk>, pcg@aber.ac.uk (Piercarlo Grandi) writes:
|> On 21 Dec 92 13:33:18 GMT, solman@athena.mit.edu (Jason W Solinsky) said:
|> 
|> solman> If you define the codes which we are concerned with to be codes
|> solman> which can only exploit ILP, then of course the level of
|> solman> parallelism is limited, but you are not dealing with general
|> solman> purpose computing anymore.
|> 
|> Ah, this discussion was indeed mostly about ILP/superscalarity/VLIW.

This is where our misunderstanding lies. It is my argument that hyperscalar
chips will be able to exploit macro level parallelism. A hyperscalar processor
would be able to issue eight instructions from a single thread simultaneously,
but its primary purpose would be to issue eight instructions from several
different threads simultaneously. This is where the talk of getting rid of
"imaginary lines" comes from. The processor of the future must be able to
excecute one highly parallelizable thread very quickly (which as you note
could be easily done with vector processors) AND execute several different
threads simultaneously.

|> pcg> Indeed pipeline designs with more than a few stages of pipelining
|> pcg> run into huge problems, and are worth doing only if a significant
|> pcg> proportion of SIMD-like operation is expected. Pipeline bubbles
|> pcg> start to become a significant problem beyond 4 pipeline stages on
|> pcg> general purpose codes, even on non superscalar architectures.
|> 
|> solman> This can be taken care of by interleaving different threads in
|> solman> the software, or using hardware which will take care of the
|> solman> interleaving on its own. The above statement is only true when
|> solman> the compiler is too dumb to notice higher level parallelism.
|> 
|> Well, here you are saying, if I read you right, that if an application
|> is suited to MIMD style computing then multithreading is the answer.
|> This in itself is nearly a tautology. If you are also implying that many
|> more applications are suited to *massive* MIMD style (macro) parallelism
|> then is commonly believed, I would ber skeptical; what I believe is that
|> quite a few *important* applications can be successfully (macro)
|> parallelized, MIMD style. For example compiling GNU Emacs; each
|> compilation of each source module can proceed in parallel, and indeed
|> one can spawn a separate thread for each function in each source module.

In the near future, we are not going to have the capability to put "massive"
numbers of execution units on a chip. I would be surprised if we saw chips
with more than 50 execution units in the near future. What I am saying is that
this level of MIMD (A macro parallelism, ILP product of a little over 100)
can be supported efficiently by enough applications to call it general purpose.
I think both the example given here and the paper that has been cited support
this. More importantly, the tendancy of software with greater levels of
abstraction to exhibit more macro-parallelism, suggests that parallelizability
of software will continue to outpace that of the hardware.

|> solman> The key question in chosing how large register files and caches
|> solman> should be, is "How large a {register file or cache} do I need
|> solman> for `good' performance on the algorithms I want to run?"
|> solman> Invariably, the size chosen is too small some of the time, while
|> solman> much of it is left unused at other times. In the multiple CPU
|> solman> version, this still happens. In the hyperscalar version,
|> solman> however, some of the execution units and threads will need a
|> solman> larger {cache or reg file} and some will be unable to utilize
|> solman> the existing space, but because they can share the same caches
|> solman> and register files, it is far less likely for performance to be
|> solman> limited by cache or register file size.
|> 
|> I agree with the sentiment here; partitioning resources can indeed lead
|> to starvation in one place but to overcommitment in another.
|> 
|> The problem is that sharing resources can have tremendous costs; sharing
|> register files between multiple parallel functional units looks nice
|> until one figures out the cost in terms of multiporting.

One thing is for sure, we can't solve the problems of resource allocation
simply by extending current methods. Just adding ports to register files
would be a horrendously expensive idea. Instead, just spread several files
with a couple of ports around and tag the data to signify where it came
from and where its going. There are many different ways to do it. Just
extending the current trend of multi porting isn't one of them though.
Besides, It will take too many cycles to wait for data to cross the entire
length of the chip. This should only be necessary when the local register
file is full.

|> It's obvious you know this, but maybe you are underestimating this cost;
|> moreover avoiding partitioning will really solve problems only when
|> there are huge swings in resource allocations by different threads; if
|> these are fairly predictable, then static partitioning does not look so
|> bad. I reckon that this is the case in many instances.

I'm sure it is, but the optimal partitioning changes from user to user. If you
want to use a single design as a general purpose uP, then this no longer looks
like a good idea. This is especially the case when you consider that most users
will be multi-tasking several programs at once and these are unlikely to be
very similar in resource consumption.

Jason W. Solinsky