performance of hardware dynamic scheduling
|
From: Bill Todd on 7 Apr 2008 07:14 Ken Hagan wrote: > On Sat, 05 Apr 2008 19:17:01 +0100, Bill Todd <billtodd(a)metrocast.net> > wrote: > >> Ken Hagan wrote: >>> but ia64 wasn't just a technical failure. >> >> And I never claimed that: I said "it was all about failure to execute >> technically (or perhaps more specifically failure to be realistic in >> their technical - and schedule - projections)". > > The reason for "unrealism" is significant here. Yes, it is. They failed to deliver. > If they were surprised by that, then that sounds like guessing wrong on > technicalities. They did, both in terms of schedule and in terms of what they could deliver regardless of schedule: McKinley may have come somewhere near delivering what was promised (though at least 4 years later than it was promised), but the promises were made for Merced, which came nowhere near delivering (despite the fact that it was at least 3 years late). If they weren't surprised by that, but bashed on anyway, > then presumably they had marketing (or someone) telling them that it didn't > matter, which sounds like guessing wrong on what the market would tolerate. Not at all: it was, as I already observed, a gamble that they could scare away the competition and corner the market before the deficiencies of the product became evident. They clearly knew what the market *wanted* (the basis for this discussion - remember?) because that's what they promised in 1994 -1995, and what they carefully refrained from correcting as the completely untenable nature of those promises started to become evident internally (i.e., they *still* knew what the market wanted and were careful not to disillusion it any earlier than necessary). > >> Second, from what I've heard, many in Intel actually *believed* their >> own Itanic hype early on, and even after the magnitude of the >> short-fall must have become evident to them instead of taking the >> common-sense approach of cutting their losses they instead took the >> gamble that they had sufficient monopoly power to *control* the market >> (which is quite different from trying to create what the market wanted). > > That would be both modes of failure then. (In deciding what projects to > start, you have to guess what you can finish and what you can sell. I'm > calling "could not finish" a technical failure and "could not sell" a > marketing failure Then you're confused about the basis of this discussion: it's not at all about inability to *market* a (sub-standard) product - it's about understanding what the market *wants*. - bill
From: Jason Riedy on 8 Apr 2008 01:36 And Nick Maclaren writes: > At one stage, many people were calling the Tera SoEMT :-( It certainly didn't help that the MTA-1 did not behave quite like the documentation claimed. The number of cycles between successive instructions in a single streams had a huge variance and varied roughly linearly with the number of streams. That's evidence that the hardware scheduler wasn't completely round- robin. Perhaps it didn't switch on every *outgoing* event, but *incoming* memory events definitely affected the scheduler. Also, the lock time for the full/empty bits had odd timing overhead in some situations, hinting at attempts of heuristic optimizations somewhere in the system. Again, perhaps not a pure "SoEMT" scheduler, but some events likely influenced the scheduling. The memory latency did show some cache behavior, too, thanks to the "hot-spot" caches per memory unit. Making microbenchmarks for the MTA-1 was fun. It was not so much fun dealing with the random, O(5 minute) uptimes combined with the O(5 minute/kB) file transfers... Jason
From: Nick Maclaren on 8 Apr 2008 03:31 In article <87myo4kiwd.fsf_-_(a)sparse.dyndns.org>, Jason Riedy <jason(a)acm.org> writes: |> |> > At one stage, many people were calling the Tera SoEMT :-( |> |> It certainly didn't help that the MTA-1 did not behave quite like |> the documentation claimed. ... Indeed. And that is true of almost all modern designs - the general and advance descriptions are almost always misleading, and the 'precise' ones are usually incomplete and do not define the terms used in the more general descriptions. That is one of the reasons that the terms change meaning so badly over time and beteween 'domains' - there isn't ANYTHING in writing to stop them drifting! God help us, IT abuses even well-defined terms badly enough :-( Regards, Nick Maclaren.
From: Ken Hagan on 8 Apr 2008 06:53 On Mon, 07 Apr 2008 12:14:46 +0100, Bill Todd <billtodd(a)metrocast.net> wrote: > Not at all: it was, as I already observed, a gamble that they could > scare away the competition and corner the market before the deficiencies > of the product became evident. That's still a statement about what the market wants, if you allow for negative values of wanting. (It's also rather odd to talk about "before the deficiencies became evident" when the very first chips exhibited the deficiencies. It's not like Merced was great but Itanium stank.) > They clearly knew what the market > *wanted* (the basis for this discussion - remember?) because that's what > they promised in 1994 -1995, But it is pointless to say that the market wants something really fast with loads for potential for getting even faster even after x86 hits a ceiling. If I ask "What does the market want?", I expect my marketing department to choose between *available* options. Since that list of options is necessarily guesswork (provided by the engineers), the marketing folks have to apply a little caution. If the engineers deliver only part of the promise, is the product still saleable? It's all a huge matrix of trade-offs, but the data for that matrix comes from both engineering and marketing departments. I think Intel missed that fact that their market predictions were highly dependent on the engineers delivering at the high end of the envelope. Either they didn't make predictions for a lower-performing product (which is negligent) or they got them wrong (believing they could sell it anyway). > and what they carefully refrained from > correcting as the completely untenable nature of those promises started > to become evident internally (i.e., they *still* knew what the market > wanted and were careful not to disillusion it any earlier than > necessary). So they *also* failed to update the matrix data as the years rolled by. *That's* equivalent to them saying that the product *as it turned out* was what the market wanted.
From: Nick Maclaren on 8 Apr 2008 07:08
In article <op.t9agjwyuss38k4(a)khagan.ttx>, "Ken Hagan" <K.Hagan(a)thermoteknix.com> writes: |> |> If I ask "What does the market want?", I expect my marketing department |> to choose between *available* options. Since that list of options is |> necessarily guesswork (provided by the engineers), the marketing folks |> have to apply a little caution. If the engineers deliver only part of |> the promise, is the product still saleable? It's all a huge matrix of |> trade-offs, but the data for that matrix comes from both engineering |> and marketing departments. Yes. What I don't know is whether the engineers stated that it couldn't be done and were ignored / overridden by marketing and executives, or they failed to warn the management in appropriately strong terms. Or, more likely, it wasn't a simple black-and-white situation, but a compendium of mistakes, miscommunication, confusion, self-delusion and so on. Doubtless some people from HP know - but none of them have gone public with that information :-) Regards, Nick Maclaren. |