Brian Gupta wrote:
Future of x86
With each extra core we are seeing more wasted heat and energy go into
missed branches of execution.
What is the answer? 1000+ core count massively parallel microprocessors,
that are based on a modified 486 core.
Assuming .65 nm process you could actually get to 4000+ cores on a
single piece of silicon today.
Did you want some cache with that? I think you'll find that the memory
controllers, coherence logic, etc, take a significant amount of space
as well. What kind of memory subsystem would permit you to issue
1000+ simultaneous loads/stores?
Basically taking Niagara, ditch CMT, using the most powerful
non-superscaler x86 core, add 64-bit (particularly memory address
space,) throw in nicely sized shared cache on the chip). I ask because
as far as I know there generally hasn't been a mainstream operating
system that was designed to support this.
Why would you ditch CMT? W/o CMT, you'd need to turn off the
integer pipeline while waiting for memory... Niagara just lets other
threads run if a thread becomes blocked on memory.
Another thing that may happen as a result of this, is that out of order
execution, multiple pipelines, and speculative execution will be moved
into the OS/Compiler. (At least until people grok MasPar coding)
The available instruction level parallelism isn't sufficient to keep
complex chips running. CMT is a easier way of keeping the
integer and fp pipelines performing useful work.
We basically have to take lessons learned from HPC and apply them to a
single system image.
What is the question?
These chips will be here before we can blink an eye. Sun's Solaris has a
long history of SMP designs (Especially after the Cray acquisition), and
have continued to greatly enhance Solaris's NUMA efficiency over the
last decade. What are your thoughts regarding scaling Solaris to
thousands of cores within the next couple of years?
FYI, NUMA support was introduced in Solaris 9 (2002).
Scaling to thousands of cores will take a while for general
purpose workloads, and it's not clear that such machines are
interesting commercially.
- Bart
--
Bart Smaalders Solaris Kernel Performance
[EMAIL PROTECTED] http://blogs.sun.com/barts
_______________________________________________
perf-discuss mailing list
perf-discuss@opensolaris.org
