Check out Jonathan’s piece today on the Rock. Sometimes he makes me nervous... but hey, let's have some fun with it. The first three people who can decipher his artful clues and guess (by commenting here) what the Big Cool Nifty Rock Secret is will win a round of blog applause from me and also a neat Sun sweatshirt or hoodie (my own, unopened conference loot sitting at the back of a closet, this is not a Sun contest etc etc).
Obviously, I can’t announce the winners until we announce the device.
Contributions
Comment feed for ongoing:
From: Clayton Wheeler (Apr 10 2007, at 23:30)
It was apparent from the Register's <a
href="http://www.theregister.co.uk/2007/04/10/sun_fujitsu_misformainframe/">scoop</a>
that something was up with memory sizing, since the biggest Rock
machine is allegedly going to have a staggering 512 DIMM slots (512
slots!) yet Jonathan referred to 256 TB OS instances. Well, I don't
think I missed the announcement of half-terabyte DIMMs, so I think I'm
going to go back and read a great deal more into Jonathan's use of
"single software domain" instead of, say, "box" or "machine."
In fact, let's do the arithmetic. Last I checked out the x64 boxes
have 4 GB DIMMs as the highest-density option; that gives 2 TB of
memory per alleged high-end M9000, so this would seem to mean that Sun
is introducing a way of yoking together up to 128 of these monsters.
I have from time to time read about Sun's development efforts around
Infiniband, which was originally designed to break the tight linkage
between CPUs, memory, and I/O, instead making them peers meshed
together via a real network. In fact, I was just today reading about
how the default transport for NFS in Solaris is now not UDP, nor TCP,
but RDMA over Infiniband. Really? The default? Interesting.
And what's more, when I read the Register's reference to an I/O
expansion box, I thought that was odd, but didn't give it too much
thought.
So my money is on the Big Secret being that the Rock systems are the
first realization of the original Infiniband vision, or something near
to it. We now get something resembling the SGI Origin 2000 systems,
where we have (to begin with) CPU+memory and I/O bricks connected
together into big single-image NUMA systems with a big old
interconnect. But it's all Infiniband, and the Solaris team has a
gigantic Infiniband putback waiting in the wings.
Likely enough, each Rock chip sees Opterons their integrated memory
controller and raises them an integrated Infiniband interface
(controller? dunno the terminology). They're made for incrementally
scaling up as well as out, all with a single super-fast interconnect
scheme (and a pair of expensive Infiniband switches, I
suppose). Result? Build a single system with a pair of Rock boxes, and
you can finally lose a single physical box and laugh it off, without
all the insanely fragile complexity-quadrupling clustering nonsense
we've been stuck with for ages. Just let Dynamic Reconfiguration and
the friendly local field service engineer deal with the dead box.
If I'm right, I like the idea. I like it a great deal.
[link]
From: Fazal Majid (Apr 10 2007, at 23:50)
It's an asynchronous chip using Ivar Sutherland's research for much less power consumption and potentially better performance?
[link]
From: Micah Dubinko (Apr 11 2007, at 00:18)
On-chip map/reduce? -m
[link]
From: David Wragg (Apr 11 2007, at 00:45)
Well, the hardware scout thread stuff isn't mentioned in Jonathon's blog piece, so that would be my first guess. But that has been widely disclosed already, so a real Big Secret could be something else.
[link]
From: Beat Bolli (Apr 11 2007, at 00:50)
64-Thread CPU (16 cores were mentioned in the post) with thread migration?
[link]
From: Dan Creswell (Apr 11 2007, at 01:39)
Hmmm, fault-tolerance, minting not fab'ing? Is it FPGA based?
[link]
From: Taras Tielkes (Apr 11 2007, at 01:50)
My guess would be hardware STM.
[link]
From: Jonno Downes (Apr 11 2007, at 03:50)
My guess - it runs java opcodes natively.
[link]
From: Antone Roundy (Apr 11 2007, at 06:31)
Here's my guess: it's build-to-order (ie. you get to mix and match 16 cores of your choice, or at least you can choose some of them--some may be mandatory) and/or an FPGA. To address 256 TB would require 48 bits, so it's got to be at least 64 bit...I'll stick with that for my guess. It will come in any color you want, as long as it's gray.
[link]
From: Ryan Cousineau (Apr 11 2007, at 08:53)
My guess is that commenter "number9" in Jon's blog has given the game away: it's a clockless chip.
Asynchronicity!
[link]
From: Paul Morriss (Apr 12 2007, at 00:22)
No-one's had a guess, so I'll have to start the ball rolling:
Some sort of RAID controller-like CPU minder that watches what's going on and does load balancing, and switches threads off locked CPUs?
[link]
From: Kemp Watson (Apr 12 2007, at 06:20)
I think Clayton is close, but I disagree with Infiniband. Sun has just released Neptune, and indicated future boxes will have 10GbE onboard. I think 'dual 10GbE on-chip with a distributed memory controller' would be closer for Rock. Jonathan doesn't say 256 TB in a box, just in a 'coherent system'; he mentions bringing 'fault tolerance to the masses'. I think Rock will be a building block for simple, cheap(ish) 10GbE-based grid computing. Perhaps Sun will fill the void from SGI's painful departure from the SSI (single system image) world by making a system that combines the best of shared-memory systems with distributed computing.
[link]