Re: New architecture ppc64el on Debian
On Tue, Sep 24, 2013 at 09:08:48AM +0900, Joel Rees wrote:
> Ergo, there hasn't been sufficient support for least significant byte
> first mode to demonstrate that the CPU can actually do it. Makes more
> sense when I think about it that way.
Certainly running Linux this way is pretty new and unproven.
> And naive guys like me would assume that developers would check
> something more fundamental than the name of the processor when trying
> to decide how to fight with the automatic optimizations.
> I tend to forget I'm not the only impatient. non-omniscient programmer
> in the world. :-/
Yeah, let us know when you find any of those.
> Which I tend to do a lot more than some people think I should.
> Although I am quite sanguine to the engineering management issues of
> putting a too-powerful CPU in the hands of the engineers and letting
> that fact slip to the guys in charge of adding features to the design.
How could you ever have too much?
> It occurs to me that the reason a least-significant-byte-first PPC
> port has failed to get traction until now is that no sensible engineer
> has wanted to waste his own time working on it. (... management always
> being too focused on immediate profits.) So the "news" in IBM's
> announcement is that they are deciding to put company money behind it.
That is probably true.
> If I were they, I think I would be more focused on fixing the
> management issues that allowed the cell version of the PPC to go to
> production as a clot of customer (Sony) requested features instead of
> an engineered product. It seems to me that they've now kicked out the
> wrong managers and are repeating the mistakes. (Standard Operating
> Procedure, as they say.)
I am not sure the cell was really a good idea. It was an interesting
idea though, but pretty difficult to program for, although people seem
willing to deal with GPUs that need similar treatment. Of course
people are hooking GPUs up to powerful CPUs that can help manage things.
The powerpc part of the cell was rather wimpy and perhaps not up to the
task it needed to do.
> ppc-le or whatever it ends up being called shouldn't take a huge
> announcement and tons of money, even at this point. Three extra
> servers to dedicate to little-endian is at most, what, USD 30,000?
> (Actual cost to IBM would be more like $3,000, I think, but they can
> probably expense the full advertised price.) The engineering support
> is more expensive, but from Breno's posts it looks like they don't
> understand the process here. Does the Fedora community give IBM some
> kind of special fast-track support that they've gotten spoiled by?
It does seem like a pretty minor thing to announce, but perhaps the
implications (at least to IBM) are a lot bigger.
> The fact that an unofficial port doesn't yet exist points to a big
> part of the problem, too.
Well it only got kernel support last month.
> And I suppose this entire post is a little off-topic.
Only a little. Debating the point of yet another Debian port is always