[nylug-talk] After meeting tonight, anyone want to discuss XEN stuff? (Also a bunch of misc thoughts and notes).

[Bryan J. Smith]

On Thu, 2008-03-27 at 10:48 -0400, Sunny Dubey wrote:
> closed source drivers of any kind on a server = epic fail

How so?

I load all sorts of binary objects in my Linux kernel on servers every
day.  To use such a blanket, generic statement not only cross a broad
set of contexts, but has 0 technical depth.

Binary objects are a reality of many things, from intelligent hardware
and related firmware, to specific engines for storage off-load and
network processing elements (NPEs) to even a few regulation mandated
binary objects (e.g., WLAN PHY/radio aspects -- i.e., I'm decade-long
colleagues of the Mathews').

The problem here isn't binary objects.  The problem here is the reason
why they are required for GPU performance on the traditional Intel
system-peripheral interconnect.  Intel needs a GPU to act like a CPU,
including all the coherency garbage that comes with it.  But instead of
renovating their platform, they used the peripheral bus.  So now they
require software hacks for coherency.

Those hacks cross the IP of many, including Intel itself that protects
it.  Most people don't realize that Direct in-Memory Execution (D[i]ME)
on AGP was an Intel "trade secret" for a long-long time (even though AGP
itself was merely a higher clocked, DDR/QDR, bridged/segmented, 32-bit
PCI bus).  Although PCIe is a full PCI SIG standard, there are many
aspects with GPUs that are still just as proprietary when it comes to
the kernel-memory interface to guarantee GPU-CPU coherency.

That's what you're talking about.  If you don't use a "software
coherency" scheme, and put it in to hardware, you drastically cut down
on the GPL issues, and probably eliminate any need for a binary object.
Unfortunately such designs are not easy, and AMD is still evolving
theirs after all these years -- let alone couldn't get a single CPU
vendor to create an HTX GPU design (hence why they bought ATI).

> If RoR was built by "real Unix people" (tm), why would the platform 
> matter ?

My thought exactly.  It's distro pissing IMPO.

Although I'll re-iterate that distribution via source-build lends itself
to much more free-form development than a packaged distribution.  Hence
why I will give Gentoo kudos there, because it's different.  But as far
as package distro v. package distro, it's all just distro pissing to me.

> Web developer groupthink mentality scares me.

It doesn't scare me.  What scares me is when people try to apply
"Internet technology" configuration management practices to non and
vice-versa.  Companies that are on the edge of Internet software end up
having to up-end package distribution, removing any
integration/regression testing done by the package distribution, so it
really just "gets in the way."

Gentoo is great there.  Same for any other leading edge development, or
even embedded.

Going the opposite direction, a lot of my fellow Gentoo advocates really
undercut the benefits of Gentoo by selling it wholly ignorant of
engineering lifecycle like integration and regression testing.  They
arrogantly think more traditional firms with 5+ year vertical
applications that can't change and need the ultimate in ABI/API
compatibility are "stupid."  I just shake my head.

There's a reason people pay $3K Enterprise Linux distros that have 7+
years of maintenance, backports and, generally, "are less fun."  ;)



-- 
Bryan J  Smith              Professional, Technical Annoyance
mailto:b.j.smith at ieee.org  http://www.linkedin.com/in/bjsmith
-------------------------------------------------------------
           Fission Power:  An Inconvenient Solution