Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.arch > #3480

Re: 'smttter IBMdroids

Show all headers | View raw

On 9/5/2011 2:16 PM, Robert Myers wrote:
> On 9/5/2011 11:26 AM, Andy "Krazy" Glew wrote:
>
>> Interesting. How do you characterize your category a versus category b?
>>
>> Real-time? Server type workloads?
>
> Those would be category b workloads.
>
>> Pointer chasing over linked data
>> structures?
>>
> Pointer chasing over linked data where what you are likely to encounter
> can be characterized pretty narrowly ahead of time probably deserves a
> category of its own.
>
>> And my understanding is that you have not been terribly happy with that.
>>
> No, I haven't. As you have explained yourself, microarchitecture has
> bent toward that which can be supplied easily and cheaply, and that
> reality favors the sequential access and simd weighting of most of the
> new widgetry. The stuff I want is hard and expensive. I got that.
>
> As I have repeated ad nauseum, the Top500 list values cheap, easy flops
> just as highly as expensive, no-so-easy flops.
>
>> I think that your categorization needs to be tweaked:
>>
>> Category a
>> a.1 sequential and local
>> a.2 non-sequential and non-local
>
> but still predictable in either case.
>
>>
>> Category b
>> Not knowing your definition, I'll suppose it to be RT, server, pointer
>> chasing.
>
> As I said, I think pointer chasing is a special case.
>
> Robert.


On a nearby slightly divergent thread:
On 9/5/2011 11:14 AM, Stephen Fuld wrote:
 >
 > I take Robert's distinction somewhat differently. I may not have it
 > correct, but . .
 >
 > You can think of category A as a program where you know the exact
 > sequence of instructions that will be executed before instruction
 > execution starts. For example, think of a CFD simulation. You could, at
 > least in principle, eliminate (or at least correctly predict) every
 > branch, could prefetch every piece of data that would be needed, etc.
 >
 > Category B would then consist of programs that get unpredictable input
 > after the program starts executing. Yes, this includes servers and RT
 > stuff, but also includes such desktop applications as word processing
 > (i.e. Is the next character the user going to input a command or some
 > input?),
 >
 > As for pointer chasing, it depends on the context. If the pointers are
 > used for a sparse array that is known ahead of time, then it would be
 > category A, but if the pointers were to data items in a database page,
 > then it would be category B.
 >
 > I don't know if this distinction has any consequences for computer
 > architecture, but I think this is what he means. Of course, I welcome
 > any corrections from Robert.



 >>For example, think of a CFD simulation. You could, at
 > least in principle, eliminate (or at least correctly predict) every
 > branch, could prefetch every piece of data that would be needed, etc.

I think that this is NOT true of CFD in general.

I think this is only true of the simplest, regular grid, CFD problems. 
Regular grid, and/or incompressible fluids.

Why I think this: anecdotal interaction with CFD experts.

(1) At UIUC one of the fellows I most valued learning from worked on 
CFD. Eventually he left academia to work on modeling stuff like smoke 
and water for a film company, like ILM (Industrial Light and Magic).

He said that the old, regular, grid type solvers where passe'.  When he 
wanted to work on compressible fluids, gas flow, turbulence, he said 
that it was largely pointer problems - irregular, changing, grids.  And 
because he was interested in turbulence, the structure of the irregular 
grid was not known in advance.

This was the guy who helped me formulate my notion of the ratio of 
compute the pointer chasing.  And while he could not foresee there being 
fewer than circa 100 FLOPS at each grid node, he helped me draw graphs 
showing precipitous decline over time.

(2) Another CFD guy I know of models flows of deformable bags of liquid. 
  Not really  turbulent, but certainly not regular either.




Now, Robert knows more about this field than I do - although I am not 
sure that he knows more about it than the guys I refer to above.

Robert and I share a common interest in bandwidth of non-sequential 
access patterns.  Snarf a few words or a cache line, at a location, and 
then move to a non contiguous location.

FFT style access patterns, such as Robert has talked about, are 
precomputable, prefetchable.

But irregular, dynamic, grid access patterns are much less precomputable.

I wonder how far the commonality between these two workloads can be 
stretched.

Back to comp.arch | Previous | Next — Previous in thread | Next in thread | Find similar | Unroll thread

Thread

'smttter IBMdroids Robert Myers <rbmyersusa@gmail.com> - 2011-08-31 22:15 -0400
  Re: 'smttter IBMdroids Robert Myers <rbmyersusa@gmail.com> - 2011-08-31 22:20 -0400
    Re: 'smttter IBMdroids Anne & Lynn Wheeler <lynn@garlic.com> - 2011-08-31 22:31 -0400
      Re: 'smttter IBMdroids Robert Myers <rbmyersusa@gmail.com> - 2011-09-01 07:30 -0700
  Re: 'smttter IBMdroids Quadibloc <jsavard@ecn.ab.ca> - 2011-09-01 09:48 -0700
    Re: 'smttter IBMdroids nmm1@cam.ac.uk - 2011-09-01 17:58 +0100
    Re: 'smttter IBMdroids Stephen Fuld <SFuld@alumni.cmu.edu.invalid> - 2011-09-01 10:58 -0700
      Re: 'smttter IBMdroids nmm1@cam.ac.uk - 2011-09-01 18:51 +0100
    Re: 'smttter IBMdroids Anne & Lynn Wheeler <lynn@garlic.com> - 2011-09-01 15:42 -0400
    Re: 'smttter IBMdroids Robert Wessel <robertwessel2@yahoo.com> - 2011-09-01 18:19 -0500
      Re: 'smttter IBMdroids Anne & Lynn Wheeler <lynn@garlic.com> - 2011-09-01 20:57 -0400
      Re: 'smttter IBMdroids nmm1@cam.ac.uk - 2011-09-02 09:17 +0100
      Re: 'smttter IBMdroids nmm1@cam.ac.uk - 2011-09-02 09:19 +0100
        Re: 'smttter IBMdroids Robert Wessel <robertwessel2@yahoo.com> - 2011-09-02 04:44 -0500
          Re: 'smttter IBMdroids nmm1@cam.ac.uk - 2011-09-02 11:39 +0100
            Re: 'smttter IBMdroids Robert Wessel <robertwessel2@yahoo.com> - 2011-09-02 11:14 -0500
        Re: 'smttter IBMdroids Robert Wessel <robertwessel2@yahoo.com> - 2011-09-02 04:47 -0500
          Re: 'smttter IBMdroids nmm1@cam.ac.uk - 2011-09-02 11:26 +0100
          Re: 'smttter IBMdroids Anne & Lynn Wheeler <lynn@garlic.com> - 2011-09-02 08:48 -0400
          Re: 'smttter IBMdroids Anne & Lynn Wheeler <lynn@garlic.com> - 2011-09-02 09:29 -0400
  Re: 'smttter IBMdroids Del Cecchi <delcecchi@gmail.com> - 2011-09-01 17:15 -0500
    Re: 'smttter IBMdroids Robert Myers <rbmyersusa@gmail.com> - 2011-09-02 10:10 -0700
      Re: innovation from Armonk, was 'smttter IBMdroids John Levine <johnl@iecc.com> - 2011-09-02 21:21 +0000
        Re: innovation from Armonk, was 'smttter IBMdroids nmm1@cam.ac.uk - 2011-09-02 21:56 +0100
        Re: innovation from Armonk, was 'smttter IBMdroids Anne & Lynn Wheeler <lynn@garlic.com> - 2011-09-02 17:49 -0400
      Re: 'smttter IBMdroids Del Cecchi <delcecchi@gmail.com> - 2011-09-02 18:04 -0500
        Re: 'smttter IBMdroids Thomas Womack <twomack@chiark.greenend.org.uk> - 2011-09-03 00:16 +0100
          Re: 'smttter IBMdroids Robert Myers <rbmyersusa@gmail.com> - 2011-09-03 14:25 -0400
            Re: 'smttter IBMdroids Thomas Womack <twomack@chiark.greenend.org.uk> - 2011-09-03 20:09 +0100
        Re: 'smttter IBMdroids Robert Myers <rbmyersusa@gmail.com> - 2011-09-03 22:02 -0400
          Re: 'smttter IBMdroids Del Cecchi <delcecchi@gmail.com> - 2011-09-03 21:50 -0500
            Re: 'smttter IBMdroids Robert Myers <rbmyersusa@gmail.com> - 2011-09-03 23:52 -0400
      Re: 'smttter IBMdroids "Andy \"Krazy\" Glew" <andy@SPAM.comp-arch.net> - 2011-09-05 08:26 -0700
        Re: 'smttter IBMdroids Stephen Fuld <SFuld@alumni.cmu.edu.invalid> - 2011-09-05 11:14 -0700
          Re: 'smttter IBMdroids Terje Mathisen <"terje.mathisen at tmsw.no"> - 2011-09-05 22:05 +0200
          Re: 'smttter IBMdroids Robert Myers <rbmyersusa@gmail.com> - 2011-09-05 17:27 -0400
        Re: 'smttter IBMdroids Robert Myers <rbmyersusa@gmail.com> - 2011-09-05 17:16 -0400
          Re: 'smttter IBMdroids "Andy \"Krazy\" Glew" <andy@SPAM.comp-arch.net> - 2011-09-05 14:40 -0700
            Re: 'smttter IBMdroids Robert Myers <rbmyersusa@gmail.com> - 2011-09-05 15:45 -0700
        Re: 'smttter IBMdroids Del Cecchi <delcecchi@gmail.com> - 2011-09-06 21:47 -0500

csiph-web