Slashdot Mirror
New Mega Alphas
GoNINzo writes: "Compaq has just announced the new Alpha servers. The have between 8 and 32 CPUs, run with a 64-bit 731 MHz Alpha chip, and current are distributed with Digital Unix or VMS. How soon before these machines are shipping with Linux preinstalled?"
The natural result of that is that hacking on SMP stuff is not of top priority to the average person using Linux let alone those that are actively "kernel hacking."
That's a mouthful that effectively says "few people truly care about SMP."
Of those that do have SMP hardware, how many have more than 2 CPUs? My SMP motherboard has only slots for 2. The Slashdot What's a good motherboard for SMP Linux? discussion mostly found 2-way and 4-way SMP hardware.
Recent pricing at PriceWatch indicates that quad Xeon mobos start around $2500, and that ignores CPUs.
Certainly consistent with these being very expensive puppies that there is, resultantly, relatively little experience with.
Soup it up to 8-way SMP, and the pricing obviously heads into the stratosphere, thus further discouraging the wide deployment that allows the "open source" principle that
If it costs $20K for the motherboard, and $100K for the system, that rather diminishes the number of "eyeballs."
I think I have to agree that Linux is unlikely to be as ready to take advantage of high end SMP hardware as VMS, "whatever they want to call Ultrix today," Irix, or Solaris.
It only will get much better when there's a goodly population of kernel hackers with 16-way Alpha boxes :-). (Drool...)
Alternatively, it would be rather cool if there was some platform where we could get "massively SMP" motherboards where the CPUs didn't have to be "massively expensive." I dunno what, exactly. StrongARM would be interesting, but I gather that it is not terribly supportive of SMP. MIPS looks like the architecture for which there exist both cheap CPUs and massively parallel systems ( e.g. - the SGI/Origin "Cray" boxes).
If you're not part of the solution, you're part of the precipitate.
Analysts estimate the GS80 will cost around $100,000, the GS160 around $500,000 and the GS320 more than $1 million. The high-end servers initially will come with OpenVMS or Tru64 Unix; Linux will be available later.
:P
--- snip ---
the '80 is the 8 processor, '160 is 16, i'll leave it as an excersize for the reader to figure out the '320.
--
blue
i browse at -1 because they're funnier than you are.
Compaq says they expect $1 billion in revenue from these boxes over the next four years. They haven't named a price range, but given the market they intend to compete in, I'm guessing these are a very pricey solution for dot com web service machines... better to just get the bandwidth and a bunch of low end boxes in a cluster, I'll bet.
-- Still waiting for the Nike endorsement
It has been slow since the maintainer of the project (Tim Leary) passed away a few years ago... I seen one running at a comdex a few years ago, the machine keep getting a "God Complex" whenever the self-awareness patch was loaded into kernel space.... The machine had a huge ego and starting sending tcp/ip packets to Solaris machine, reportly the packets contained things like "So, check me out, Linux stud, that's right, you want to get with this baby" and "Check out my uptime chic, my uptime is huge, and that isn't the only thing huge, my other huge thing has a high uptime to, want to get with this baby?"
Microsoft did intergrate a self-awareness patch into Windows98, but everytime they installed it, it was slow, sluggish and crashed a lot. Most thought it was just windows being windows, but some experts believe that the machine was depressed and tried to commit sucide once it was aware of it's true nature...
Microsoft advised users to install the "heavy drinking" patch, to "easy the pain" of the Windows self-awareness patch.
"`Ford, you're turning into a penguin. Stop it.'" -THHGTTG
It was, unfortunately, a fairly typical corporate presentation, unencumbered by any real hard data.
They simulcast the CEO's announcement from NY, and we heard from another high-level manager, some dude from Oracle and a pre-recorded "howdy" from Larry Ellison. This was followed by a lunch break, some local presentations, a brief panel discussion and the requisite giveaways. Linux *did* get a mention in a couple of slides.
A great deal of time was spent talking about how great the Compaq/Oracle partnership is. I asked about Informix benchmarking, but didn't get a whole lot of response beyond "we expect the results to be good."
Whatever. There wasn't much in the way of new information. Lots of processors, Tru64 Unix scalable, manageable, really fast, etc, etc.
A brief list of buzzwords:
functionality
knowledge management
megatrend
the "edge of the web"
clickstream
"go to market"
Zero-latency-enterprise
And there you have it.
As for Beowulf, in fact the current Linux port sort of does a "beowulf-in-a-box". We support SMP up to four processors. Above that and you run multiple instances of the OS in different "partitions" of CPU-sets. Again, if a business need arose to require supporting a single instance of Linux on a 32-processor system, we could probably make that happen ("Given enough money, all problems are shallow?" 8-) )
I helped install 2 sites with GS160's in the past 6 months. There isn't a lot exciting about them. These machines are fast, because they have the new EV7 Alpha chips, but they really aren't anything groundbreaking. The same basic hardware has had 3 different names in the last 5 years TurboLaser 7000 a.k.a. Alpha 8400 a.k.a. GS160. It is a shared bus architecture that can support mulitple PCI backplanes. The thing to be on the lookout for from Compaq is called "WildFire." That will be a true distributed hardware architecture. This means Compaq will finally catch up with Sun... The operating system has a similar history (OSF1 a.k.a. Digital Unix a.k.a. Tru64 Unix.) Whenever Intel releases Merced, they have promised that the OS will support it. Keep an eye on this verison of Unix. They just added clustered filesystems, which, if it works, will make this unix ideal for distributed high-availability applications.
Call it "The operating system formerly known as the operating system formerly knows as OSF/1"
Okay, these were supposed to ship early last year, but compaq was getting lousy yields on the custom asics used in the box.
What's cool about these is they've got a crossbar switched architecture, so they scale better than a bus or omega switched network. What sucks is that the alpha architecture hasn't kept up in the clock speed race. Ghz alphas were supposed to be out by now too. Historically, the alpha has had the highest clock speed of any chip on the market. No longer. This is a bummer since they don't do as much per clock as, say, a PA-Risc system.
Still, these are sweet boxen and should deliver truly obscene tpc-c and tpc-h results. Drop a couple of storageworks coffins into the server room and all your data warehousing needs are met.
--shoeboy
(former microserf)
Right now Linux has shown only mediocre or below performance _relative_ to other, SMP optimized OS's when there are more than 4 CPU's. So you may get Linux only if you pound on the table when you ask for it, and the tech's will probably be scratching their heads in wonder that you'd spend the money for a system that Linux can't take advantage of nearly as well as DEC Unix or VMS.
Do we need to re-examine the entire approach to the kernel to really take advantage of 8+ CPU's? Why does a Linux kernel which works well with many CPU's have to be the same kernel for systems with only 1 or 2 CPU's?