Follow Slashdot blog updates by subscribing to our blog RSS feed

 


Forgot your password?
Close
typodupeerror
×
Hardware

Here come the PowerPC Linux systems 86

Posted by CmdrTaco from the what-the-hell-do-you-do-with-five-cpus dept.
webslacker writes "A company called Total Impact is announcing multiprocessor PowerPC systems using IBM's PowerPC Open Platform (POP) with Linux as its OS of choice. In fact, their short press release states that their base model config will be a 5 processor box, and that they've even gotten a 128 processor system running under the MacOS. " five cpus?! Sweeeet.
This discussion has been archived. No new comments can be posted.

Here come the PowerPC Linux systems More

Here come the PowerPC Linux systems

Comments Filter:
  • Actually, it sounds like their archetecture (4 CPU PCI cards) is more like having 32 4 CPU boxes, using the PCI bus as a high speed, low latency, switched network connection. Plus, I'd bet that the extra CPUs run some really light weight task manager, rather than a complete OS.
  • I don't think this is a complete computer system-just a board. I'm still waiting for something that works right out of the box
  • From the press release :


    The TotalMP Linux system is a module design that will allow users
    to upgrade to more processors as needed. The base model will
    include 5 PowerPC processors and will be scalable to 13
    processors in a single desktop system. Additional processors can
    be added via a passive backplane allowing servers to scale to 100's
    of processors.

    So for those of you wondering how they're doing SMP, that is it : daughter card (PCI ? the article don't tell) that you add to a passive backplane. Not the best way to do MP, IMHO (can you spell bus contention ?).

  • I found that the default one comes with 4 DIMMs, 3 PCI (small by PC standards), a 82559 10/100 Ethernet controller, and ISA Crystal Sound chip (I'd MUCH rather have a Creative ES1373 PCI chip, which is what Intel et. al. use now) Looks like it can only control 4 PCI devices, so a board using a PCI-PCI (or AGP->PCI66) bridge would probably be a Good Thing.

    That's mostly just nit-picking though... I was drooling over the MPC750 specs... 6watts at 466mhz, with copper technology, nifty. Probably would work fine w/o a fan. :) (A 7500 with Altivec would be even cooler, but IBM isn't going to officially mention it in the design, alas.)

  • I don't think MacOS would be able to use those CPU's but since it isn't protected mode you could write multiprocessing support into each of your programs individually. Blegh!
  • "Does anyone (else) remember the DEC Firefly - ISTR it was one of the first implementations of Berkeley cache coherency schemes, and came with *five* CPU's. I guess it was PDP-11 or VAX based, probably running RSX-11 or VMS, but conceivably it could have run Ultrix."

    And how many 5 cpu systems are there compared to 2^x? Hardly any. Theres a reason they like 2^x cpus, or at least even, often faked to 2^x thru bridges a la 6. Thats what I'm getting at. I never said it couldn't have 5, just thats its odd.

    "Principal performance limitations on SMP are (a) memory bandwidth and (b) kernel threading granularity. Cray had the latter well licked in Unicos before they started going to the wall; 32 processor YMP's and x90's work perfectly well, though the main OS intensive bits in supercomputer apps tend to be I/O."

    In the same vein, NEC's SX-5 supercomputers (32 cpus, 128Gb/sec* memory bandwidth) can be clustered, and I think the cluster has some SSI (Single System Image) behaviour.

    "The SGI/Cray Origin 2K (ccNUMA) is supposed to be single system image, and they ship that in a 128 CPU config, using a 2x2 crossbar to join four 32-cpu racks. Anyone played with one?"

    Bet its parallel at heart.

    "I know also that a one-off 128 or 256 proc version of the old Convex S-Class (now HP V-Class / Qdome) was made, I think for CalTech. This would be running SPP-UX I presume, don't know if it's SSI though."

    My guess is parallel, again.

    "If you think 4Gb RAM and 4 CPUs is a big box, you've spent too long listening to the marketing boys in Redmond."

    Did I say anything like this? NO. Thank you for putting words in my mouth. However, any one of those space heaters is well beyond any individual needs of any kind. Someone gets to play with one during a degree lineup or in some company basement, and they think they're god.

    Redmond? Since when did I say anything about MS? You assumed, though it only makes you the ass, contrary to the usual 'u' and 'me'

    "* yes, Gigabytes per second, it's not a misprint"

    Oh gee, like I failed basic college chemistry or something. 'Duh, whats G stand for?' Please.

    You have some sort of superiority complex.

    BTW,

    "Multiple cpus like to be base2 numbers (2^x) or at least even -- don't they?"

    Look up like. Look up have. Have a nice day.



    Anonymous Coward, get it? :)
  • Throwing together 128 boards is not the same as delivering a decent SMP, especially when you look at the economic angle. Hard engineering and economic realities

    - each CPU throws off the thermal equivalent of a warm brick. Unless you've got decent cooling design, expect to waste a lot of floor space and/or exterior cooling

    - handling hardware failure, you need diagnostics and graceful recovery, and with the complexity increasing non-linearly with the number of component combinations, expect some serious downtime

    - balanced systems ... putting together the I/O subsystems + associated maintenance is going to cost an arm and a leg. Remember that raw hardware is only 15-30% of the total cost of any large-scale installation.

    - the cost is in the hardware but the value is in the applications. Unless there are compelling applications or a decent software base, it will be difficult to shift enough applications in time to beat Moore's Law.

    LL
  • besides, seti is searching for life on other planets in space not in vacuum dimwit.
    Somebody needs new batteries for their sarcasm detector.
  • Anyone have any thoughts about how fast these systems could turn out to be? I spent quite a while researching the Alpha, hoping for an SMP alpha box, but realized that the price/performance is completely out of whack unless you're planning on building a rendering farm.
    I was basically planning on waiting for SMP K7 motherboards, but if the G4 is as good (or better) then I'd rather go that route and free myself from x86 tyrany once and for all.
    For example, if the extra 4 CPUs *are* on a PCI card, rather than the motherboard itself, will that cause bandwidth limitations?
    Does Linux even support PPC SMP as of yet?
  • Keep your eyes on The Open PPC Project, currently at http://www.tgeller.com/openppc [tgeller.com]. It's a meeting place for folks interested in building PPC-based boxes, and there's been a lot of talk about G3->G4 upgradeability. Based on the preliminary schematics (available through the site under "Plans"), it won't be that hard.

    The real action's in the mailing list. [phys.sfu.ca] :)

    --Tom

  • That's what this is, a cluster in a box. Each card has four CPU's, and the cards plug into a passive backplane. There is a cable going from the box to plug into a special PCI card in the regular Mac computer, and software that you run to allow mutithreaded/SMP apps (like Photoshop) to be distributed across all of the connected CPU's.

    "Hmm, that rendering job is taking a lot longer than we expected, let's get another couple of these boxes and plug 'em in..."
    (later on)
    "Much better!!"


    Props to these guys. This is some mad hacking.
  • You may be askin' yourself, "Barry and Levon, where'd you get $240 to buy all that puddin' ?"









    Shhh....







    AWWW YEAH!!!!
  • Dumbass, it's not a 128 CPU SMP system, it uses MPI (Message Massing Interface). This is like a beowulf system, except each processor isn't on its own separate computer.

    FWIW, the SP2 supercomputers work in a similar fashion as these, just with a lot more engineering.
  • This _is_ a beowulf, only on one machine, not spread across compute nodes.

  • Not only are PPCs fast... (Score:3)

    by RevDigger ( 4288 ) <haroldp.internal@org> on Saturday September 04, 1999 @05:28AM (#1705142) Homepage
    PPC systems - modest one to four processor systems - would be such sweet servers:
    • Yes, they are indeed very fast
    • G4s are designed with SMP in mind, and future revisions are looking at just putting multiple cores on one chip
    • They are cool...as in, you don't need cryogenic cooling to keep em from melting. A heat sink covers it.
    • they are tiny, so you don't need a sledge hammer to squish them into a 1 or 2 U rack case
    • They are low power
    • The archetecture actually still has a future, it's not crammed with legacy BS
    • Real RISC
    • They are not Intel
    And as a bonus for developing open boards, the BSDs already have ports, and linux already has a port. I am so ready for this.

    - H

  • It is well known that Linux doesn't yet scale very well beyond an SMP of two processors.
    It's commonly stated that OS Foo doesn't scale as well to multiple processors as OS Bar. How about some numbers, people?
  • I think you nailed the biggest problem with this setup: the extra processors have to use the PCI bus. Which will certainly increase the use of the bus a lot, getting worse the more processors there are. And since the PCI bus is so slow (compared to the system bus the first CPU is using) it will definitely impact performance of the additional processors, as well as likely causing the CPU card and other PCI cards to have to fight for limited PCI bandwith.

    This is what the RC5 numbers (that show a near-perfect scaling of performance as the additional CPUs are added) don't tell you. RC5 doesn't require much I/O to/from the CPU. It's mostly just internal number crunching, and at the end spits out the results. Most other apps aren't like this and won't see much (if any) benefit from the setup (in fact, it could decrease overall performance if the CPU card hogs the PCI bus too much).
  • you can have your voodoodoo while I drool at those geforce 256's...yum
  • Quoted from the website:

    "Total Impact's transparent software interface, Total Freedom, eliminates complicated and expensive re-coding of software applications developed for the Macintosh in order to offer an acceleration solution. Virtually any software application that runs on a Power Macintosh can easily be modified to take advantage of the Total Power MP accelerator boards, whether it is written in C, C++, FORTRAN or PASCAL."

    So that's different from existing Daystar systems. The current (or upcoming) version of MacOS also has SMP built-in if I am not mistaken.

    So they did it with G3 CPUs under MacOS. If they can get that running, they shouldn't have too hard a time getting it going under Linux...

    Rendering... Compiling... *drool*

    barbabob
  • It's on an expansion chassis, isn't it? Only needs one of the machine's slots.
  • I can see it now....
    Enterprise in the middle of a battle....
    Screen suddenly goes blue and computerized voice says.....
    "You have performed an illegal function... shall I Re-Boot?"
  • >likely causing the CPU card and other PCI cards >to have to fight for limited PCI bandwith.

    The two G4 boxes (from apple) that use the new Sawtooth motherboard (the medium and high-end models) have 2X PCI. ie PCI bus runs at 66 MHZ instead of 33. I know that 66 MHZ isn't _that_ fast, but it is an improvement...

    I also remember hearing something about a 64 bit wide PCI bus. Anybody know about that?
  • IBM have ideed promised to make G4-chips, but not with the AltiVec-units as Motorola does.

    - Henrik
  • >each CPU throws off the thermal equivalent of a
    >warm brick. Unless you've got decent cooling
    >design, expect to waste a lot of floor space and/
    >or exterior cooling

    Not quite.. I'm sitting on a PowerMac G3/333 and the CPU is 28 centigrades hot.. If i put a pen in my fan, it'll rise to 40C..

    I don't think the cooling will be any problem at all.

    - Henrik
  • this using that new "open" IBM powerpc spec mothabored? That seems incredibly quick. Maybe I'm confused.

    128 processers under MacOS!?!?! I'm skeptical about all this...


  • "We have the multiprocessor expertise and are currently running a 128 processor system under the MacOS"

    Read carefully - apparenty their 128 processor system is not using that new architecture. It seems today they have nothing but the idea to try and build something.

  • Re:Mmmmmmmmm...new toy. BUT... (Score:2)

    by Anonymous Coward
    You don't need 2^x number of processors for SMP like you don't need 2^x number of cards for PCI. Basically, all you need is a clean shot to the bus arbiter, an electrically stable bus, and you should be good to go. Assuming the PowerPC has the same "tag" based scheme as the high-end POWER chips, that is.
  • Well... We can be pretty sure it's eather 604's or 7500's (G4). Since at least 750 (G3) is a VERY BAD MP chip. My guess is 7500's tho. Since it's by far the best MP chip of the PPC's
  • They claim to have a 5 processor box, but (as an Apple user) I'm annoyed that they don't mention what PowerPC chips they are using.

    For all I know, they're using 604e's!!

    You can have MP under MacOS (look at Apple's own, or the DayStar Genesis), but the OS itself doesn't take advantage of them, only programs written with MP instructions (like After Effects and PhotoShop)

    However, MP Linux, ooh baby that'd be sweet.

    Pope
  • FIVE cpus? Not 4, 8 or even 6? Huh.

    Multiple cpus like to be base2 numbers (2^x) or at least even (I recall 6 ppros via way of two 'four cpu' hubs that used the 'fourth' cpu spot to fake out and talk to the other hub, and the same for the other hub, hence 6 total) don't they?

    As for 128 LOL there's hardly any OS that would have a clue. Only parallel setups achieve that. Can you say idle time? :) "Hi, meet my new space heater. Doubles as a computer, only cost $100,000"



    Anonymous Coward, get it? :)
  • MacOS 7,8,9... no, they don't properly use multiple CPU's. But don't forget: MacOS X is more similar to BSD. More importantly, a chain of events is likely to make a many-CPU PowerPC box even more viable: (1) SGI is donating much of their source to Linux (2) they're likely to donate their multiprocessor code (3) Linux on PPC could use this in Linux to scale to hundreds of CPU's. Oh, yeah: (4) G4 chips are awesome, and will give a wonderful return on the investment in such a machine.
  • Hmmm...I wonder how long it will be till VA Research starts offering these kinds of boxes?
  • Is it just me or do the numbers they list on there key processing rates for RC5 seem a little suspect. For example:

    3,093 Kkeys/sec- 4 CPU's
    6,177 Kkeys/sec- 8 CPU's

    Now, if you do the math, that's 99.9% efficiency at scaling from 4 to 8 processors. My understanding is that getting multi-processor systems to scale linearly is not a simple task. Especially considering that they seem to be using a hack of a way to do it (transmitting though the PCI bus). Anybody here with more technical expertise care to comment on this?

    ---

  • Just think,if you had this and added the LBC 2000
    barcalounger for geeks,the INAC(imac toilet)and a
    minifridge,you could successfully be assimilated
    by the network.Ha! I'd never get up again :)
  • If these machines are priced similarly, to Intel's offerings, then I would buy one....

    Another benefit is that you get a Factory assembled Machine without paying the MS Tax.

    looks promising; I'll be watching....
  • This company has produced PCI cards with multiple PPC chips on them for a while. It used to be 604s, now it'll be G3s or G4s. Assuming they're using a Daystar Clone, with 6 PCI slots, combined with several of those external PCI chassies, they could easily have a system with 128 PPC chips, although you'd have to write programs using their API to take advantage of them.
  • And just wait until there's multi-CPU G4 boards. Oh, baby!

    We know the G4 already can run linuxppc. now we just need the boxes. :)

  • Total Impact have been at this for a while (Score:1)

    by Anonymous Coward
    They have marketed PowerPC's on PCI cards for quite some time. There have been drivers written for a lot of different OS's and they work in x86 machines as well as Apple Mac's.

    Don't think of Symmetric Multiprocessing, as this ain't it. Think of Asymmetric Multiprocessing where one CPU controls the work of the others. This means that parallel tasks are the ONLY tasks that will receive a speed bump. RC5 is a GREAT example, they are probably running an instance on each CPU. For those of you who can't resist the urge, yes this is kind of like a Beowulf in a box.

    HPCF

  • well, yes and no (Score:2)

    by Anonymous Coward
    RC5 cracking scales really well. Just ask distributed.net. RC5 key processing is not IO-bound.

    So, because of that, it is still misleading for them to advertise RC5 key processing as an indicator of multiprocessor performance. Real world multiprocessing tasks will generate far more bus activity, and thus will not scale nearly as well.
  • I second this. I already *know* that my next computer won't be running an Intel processor. Whether it's an Athlon, an Alpha, or a PowerPC, I'm not sure. All I want is to have some *choice*! Manufacturers, take IBM's specs on the PowerPC motherboard and *MAKE* *THEM*! I know I'd buy one.
  • Will these run NetBSD? I've been playing with BSD on PPC chips and it a damn sight better than everything else I've ever touched (except DEC UNIX). I'd seriously look into one of these if it ran NetBSD (or any other varient).
  • The parent post is utter crud. Is it me, or has the average level of computer science knowledge on /. dropped drastically of late?

    Does anyone (else) remember the DEC Firefly - ISTR it was one of the first implementations of Berkeley cache coherency schemes, and came with *five* CPU's. I guess it was PDP-11 or VAX based, probably running RSX-11 or VMS, but conceivably it could have run Ultrix.

    Principal performance limitations on SMP are (a) memory bandwidth and (b) kernel threading granularity. Cray had the latter well licked in Unicos before they started going to the wall; 32 processor YMP's and x90's work perfectly well, though the main OS intensive bits in supercomputer apps tend to be I/O.

    In the same vein, NEC's SX-5 supercomputers (32 cpus, 128Gb/sec* memory bandwidth) can be clustered, and I think the cluster has some SSI (Single System Image) behaviour.

    The SGI/Cray Origin 2K (ccNUMA) is supposed to be single system image, and they ship that in a 128 CPU config, using a 2x2 crossbar to join four 32-cpu racks. Anyone played with one?

    I know also that a one-off 128 or 256 proc version of the old Convex S-Class (now HP V-Class / Qdome) was made, I think for CalTech. This would be running SPP-UX I presume, don't know if it's SSI though.

    If you think 4Gb RAM and 4 CPUs is a big box, you've spent too long listening to the marketing boys in Redmond.

    Dave

    * yes, Gigabytes per second, it's not a misprint

  • What's really bad about x86 (Score:3)

    by RallyDriver ( 49641 ) on Saturday September 04, 1999 @08:02AM (#1705180) Homepage
    The problem with the x86 architecture is not being big-endian (several RISC chips do both) it's the fact that's it's a crappy asymmetric*, CISC, strictly 32-bit and below design. Current Pentiums and friends use a 32-bit RISC core, with microcode to implement the x86 instruction set, bringing them full circle to the early 1970's and things like IBM 370 and VAX. Yuck.

    The only reason self-respecting Linux / *BSD / Be fans dirty their desks with such badly architected technology is that the Windows PC revolution has made the price/performance ratio of the hardware second to none. Architecturally nice though UltraSparc might be, I'm fscked if I'll pay US$3.5K for a 350MHz Ultra 5 with IDE, SDRAM and other PC-component technology when I can get a 500MHz Pentium for $1.2K (and falling) that is more or less the same performance. It's all about mass production.

    Yes, Sun have missed the point with their latest workstation offerings.

    Despite the design effort being poured in by Intel, AMD and others, I expect the gap between Intel x86 and RISC to widen as x86 becomes increasingly crippled by backward compatibility.

    Panic not though - Linux is already up and running on all the major RISC architectures, including the much delayed IA-64. With multi-platform use well established, it is well poised to take advantage of any new CPU that comes along.

    I too would like to see an increase in the availability of RISC cpu's on PC ATX form factor motherboards, with PCI and AGP, etc.

    * in the use of registers - each of the 16 registers in the x86 works with a different subset of the instructions, in contrast to typical modern CPU's where there are typically 32 registers, all fully interchangeable. This feature makes it easier to write good optimising compilers.
  • That's strange, I don't believe I've ever seen
    a blue screen (or a crash) while using BeOS...

    Or did you blindly assume that because I prefer
    an intuitive GUI, I must be a windoze user?

    -WW

    --
    Why are there so many Unix-using Star Trek fans?
    When was the last time Picard said, "Computer, bring
  • All I want is a G3 on a ATX motherboard. Where can I get this?
  • They have something shipping. It's an external PCI-based "PowerBox [totalimpact.com]" that you load up with their CPU cards. I've heard of these guys before. The seem to do some crazy hardware hacking.

    --
  • They use G3s. I'm sure they'll switch to G4 now that it's out, but they've had MP products for Macs for quite some time, and they're G3 based.

    --
  • FIVE cpus? Not 4, 8 or even 6? Huh

    This is just a guess, but it looks like the main thing they make is a 4 CPU PCI expansion board. So, if you have a single CPU mac with a PCI bus and you add one card you have 1 + 4 cpus. This also makes sense as they say at one place in their page that you can have 13 CPUs (1 + 3*4).

    Here [totalimpact.com] is a pic from their page which lead me to this assumption.

  • would it be better to have 1 computer with 128 processors or 8 16 processor boxes? as i understood it beowulf was better suited for parrallel processing than just having one computer with a mass of processors?
    char *stupidsig = "this is my dumb sig";
  • Remember that IBM have no yet agreed to make G4's, as of yet only motorola are making them.
  • Oh, just shut up.
    Beowulf this and beowulf that. Beowulf is just PVM and MPI on top of linux. It i`s only usefull on "embarassingly parallelizable" computations. It is completely of no use to linux users, whos main problems are solved with apache, emacs, and wordperfect. If you don't already own an SP1, challenge, or origin, you probably have no use of a beowulf cluster.

  • Solaris supposedly does 128 processors cleanly (If it doesn't do that, then it isn't good for *anything* =P )

  • So if i get this right, you have one PPC in the mainboard and a 4 cpu node on the pci bus with its own memory subsystem and a kernel running. So you will actually run two instances of the linux kernel and have a 4-way node with no I/O apart from PCI and a 1-way node (the POP board) controlling the system.

    So basically what they are saying is that their 4-way PCI cards can be plugged into the PCI bus of the POP board. Hooray they can plug a PCI card into a PCI slot. Woooohaaa!!

    I announce that I can make a 5-way PPC box. I know how to plug a PCI card into a PCI bus!!

    Anyway the systems will be cool for cluster-friendly applications, as the PCI bus is better than most networks both with respect to latencies and bandwidth (considering that the traffic is passing the PCI bus anyway it IS better).

    But then again - networks can be switched, and the length of the PCI buys is limited, so a real network is probably better for many nodes. They should switch the PCI bus, but then they would be making a RS/6000 SP like system (at an astronomical cost)

    The concept is cool. But there is nothing new about this. The 4-way boards have been aroud for a while. OK, so maybe the news is that they will be providing Linux software for the board (A PCI interface to BEOWULF would be cool, since some programs exist and a lot of people are working with BEOWULF).

    ________________
    May the source be with you...
  • That's dickish of you since Apple has contributed money and development resources to the G4 development effort. Motorola wouldn't even be making desktop PPCs if it wasn't for the Apple market and IBM CERTAINLY wouldn't be.. they have their own fish to fry. Yeah, they're being "stupid" or "nasty" with the G3 upgrade block but in my opinion there's more to that story than we're hearing.. judgement withheld on that one.
  • PPC systems *DO* have PCI slots, you know. :)

    I'm happily using a 3dfx branded Voodoo3 card right now, with generic drivers released last month. Very cool.

    Also, in the new G4 systems being released, all but the very bottom-end machine include an AGP slot in addition to PCI slots, so rest assured that PCI and ASP will be supported.

    Daniel
  • Didn't you go to the site and read it?! It's not SMP. They've just got cpu's on pci cards - max 133 MB/s. IMHO, linux would not support it. They could port their library and get software to use the extra cpus I suppose (see their rc5 benchmark) The altivec technology in G4's makes that rather second rate tho.
  • The PowerPC 750 lacks something that I think is called "cache coherence" circuitry, due to it's heritage from the PowerPC 603 & 603e... Apple has in the past said that providers could still create G3 Multi-Processing cards if they created a software patch to accomodate the lack of circuitry... It's proven that it's possible by the existance of the BeBox, a 603 based multiprocessing machine...
  • Gee... I can just imagine my VESA bootup screen, full of 128 pictures of Tux.
  • But if they're G4s, at a gigaflop each (for moderately SIMDable computations)...

    MAN what a box that could be! (Then think about a cluster of them...)

    Can you imagine the consternation among the export controllers, when a high-end home-computer can finally be used to predict the weather several days out or model nuclear devices?

    The Millenium is upon us, and the Singularity is not far beyond.

  • >128 processers under MacOS!?!?! I'm skeptical about all this.

    It sounds like most of their products are like a "rendering farm in a box"... MacOS runs on a standard Mac (with 1 or 2 CPUs), and then batch jobs written using their MPI libraries use the rest of the CPUs.

  • It actually seems to be above 100% scalability... the jump from 8 to 6 CPU's slightly more than doubles it's throughput... Higher up the scale, it seems to peter of a little, though
  • Well, their current cards seem to have memory on board, then use MPI over the PCI bus--ie, not SMP, but a bunch of processors used for batch rendering, etc.
  • Back during DESCHALL, I remember someone noticing that a 2 CPU 604e was more than twice as fast a 1 CPU, because it turns out that the overhead of the MacOS interupt handlers is pretty high. Since all interupts were handled on cpu #0, you could watch the cracking rates of the two instances and get really depressed about your OS... (though the same guy said the "super efficient" BeOS was even worse)

Slashdot Top Deals

BLISS is ignorance.

Close