Slashdot Mirror
SGI NUMAflex Linux System On Display @ SC2002
jarrod.smith writes "
According to SGI will unveil
its Intel® Itanium® 2
NUMAflex shared-memory supercomputer architecture (which runs Linux as its OS) at Supercomputing 2002 which runs this week in Baltimore, MD.
The link at SGI says the system will be on display at the show. The exhibit floor opens this evening. Unfortunately I did not go this year. Can those lucky enough to be at the meeting scope it out and post comments?"
1: Fine-tune Linux to fit the platform
2: Design the platform to utilize linux
So it sounds fair to me.. Consider installing some propriatory OS instead.. where they cannot play around, change kernel design, drivers, VM or whatever they fancy. Would not that be a greater drawback ?
Probable impossibilities are to be preferred to improbable possibilities.
Aristotele
The Los Alamos National Laboratory is building a supercomputer based on a Beowulf Cluster with 1024 nodes (2 processor in each node). You can read the story here or in this Slashdot thread.
SGI has been working on this machine for a while now, and you can be sure the linux it runs has been tweaked for performance.
This machine is Intel based. SGI had to choose to port IRIX or make Linux run well. IRIX isn't going on their IA64 computers, period.
SGI has always focused on high speed internal communication in its machines, and SGI has tended to lag behind the curve in the MIPS processor's raw MHZ. The IA64 chips are much faster than any MIPS processor out there. This machine has some amazing performance--I think everyone will be surprised at what SGI has done. If you need a high end single image shared memory Linux or Intel solution, SGI has filled that role.
To be fair, if you don't need this machine.. don't buy it or think about it. SGI is a dying company and in the long term you'll get better support elsewhere. Sun and HP will have competing machines out, but they won't perform as well. I trust HP and Sun to be around longer than SGI though, and they won't fuck you over like SGI will.
What else can I say? Oh, the one I saw was painted in penguin colors.
Well, damn as nearly. Linux is only in catchup when the manufacturor will not release spec on how to use thier hardware.
When it comes to NUMA machines, Linux is up there. It may not excel at everything (yet - I'm sure that it will get there if it's not already). I'm mostly talking about the 2.5 kernel series.
From the status list
New scheduler for improved scalability (Ingo Molnar)
Support for Next Generation POSIX Threading (NGPT team)
Syscall interface for CPU task affinity (Robert Love)
Hotplug CPU support (Rusty Russell)
NUMA topology support (Matt Dobson)
Per-cpu hot & cold page lists (Andrew Morton, Martin Bligh)
NUMA aware scheduler extensions (Erich Focht, Michael Hohnbaum)
The biggest performance changes in 2.5 seem to be in the many thread and many CPU region, including NUMA.
I'd trust it. (Yes, I do do scientific supercomputing).
...okay so Linux is being applied to all these terrific projects of scale both large and small. Is it because it's an open system with seemingly hyperactive development or is it because it's simply better than anything else out there?
I'm trying my best to maintain a level of respect for the MS operating system product so I'd like to know if anyone knows of any amazing projects MS OSs are being used for. For that matter, what about other OSs in general?
I think it's terrific that Linux is used this way but I wonder if it's because of its availability or because of its technology. I tend to think it's for its availability but I'm no expert. I think answers and other points of perspective from others in the Slashdot community would help to show some objectivity here.
My problem with one of these is that Itanium2s are so damn hot, I wonder about the density of computational power. Granted, an individual Itanium2 smokes the hell out of an R14k, but since our 600MHz R14ks use on the order of 15W and I think an Itanium2 is up around 150W, we can get many times the computrons (R14ks aren't fast, but they aren't *that* slow) per Watt. Cooling is kind of limiting factor, because these machines go in labs, not prepared computer rooms.
Right now we have a couple of 8-way Onyx2s and we're in the process of building an 8-way Origin 300. For the kind of work we do (realtime simulations), where latency is king, we prefer to put each part of a task on its own CPU, so having 8 processors is nice, whereas having an 8-way Itanium2 would be prohibitively costly to cool, although it'd be nice for when we just crank: see 483/499 SpecInt/FP base for 600MHz R14k vs. 810/1350 SpecInt/FP base for 1GHz Itanium2.
I *would* like to move to Linux from IRIX, I think. I really like the IRIX realtime support (all the REACT stuff), but I am tired of poor tool support and limited lack of updates, etc. I think $500k worth of machines (in *that* lab) would warrant a better resolution of some issues we've had.
Finally, I very much anticipate the day that these Linux scalability improvements filter down into something like a 4-way Clawhammer system. That system could very likely do a lot of our work (at what, maybe $20k for a system?) that we now drop $50k on for SGIs.
SGI has been coding for many years (3-4 at least). It released many code in the public domain such as the (defunct) Apache acceleration program which was simply amazing. Don't forget XFS, one of the best filesystem out there. I feel that for marketing reason, they will not release any "performance tuning" code anymore and keep it for themselve. That's good for customers but not so good for us. In a certain way, such code is probably proprietary to their architecture.
-- Leeeter than leet