Slashdot Mirror


Silicon Graphics Will Put Linux On Origin

deran9ed writes: "Silicon Graphics plans to introduce a version of its Origin 3000 series computer built around Intel's 64-bit IA-64 Itanium processor running Linux, according to SGI Chairman and CEO Robert Bishop. The current Origin 3000 computers from SGI are built around processors from MIPS Technologies and run SGI's proprietary Irix operating system. SGI has not decided as yet on the name for the new product line. Infoworld article."

3 of 102 comments (clear)

  1. Re:Let's all hope that.. by stripes · · Score: 3
    SGI has been working on NUMA support for Linux for quite a while now. They've been the ones doing the discontiguous memory patches and a bunch of other related things.

    Rock on! That's cool.

    It is missing things needed to make NUMA systems useful (as opposed to "can boot and run"). For example:

    • Phsycally coping shared pages. 500ns is a long time to wait for memory, it is common for NUMA kernels to be able to copy pages from one memory board to another. Very useful for code, also useful for read-only data, sometimes useful for infrequently-written data.
    • Physacally moving pages from one memory board to another. If one thread is doing a lot of writes to a page, it can be very useful to move the data page closer.
    • User level hints to the kernel about when to do these things.
    • A scheme to let sysadmins partition up a big machine (letting a compony/university pool money from multiple departments), allowing unused boards to be used by non-owers is good, or at least allowing something like a cron job to change it...
    • Support for draining and pulling boards, as well as plugging in new ones to a running system.

    I have no doubt SGI can add those things to Linux, IRIX does (almost) all of that allready.

    What seems odd to me is that it is pretty clear that Itanic will not be cost nor performance competitive when it finally ships - all the other big boys have said they aren't going to bother with Itanic for anything but 1-4 way type boxes. McKinley (the successor to Itanic) is looking pretty good, recent reports say that it will debut at 1.4GHz around the end of the year (whereas Itanic can barely do 800MHz today).

    Normally when being told one CPU only runs at 800Mhz and another runs at 1.4Ghz, so the 800Mhz one is crap, I have a lot of objections. Like "the 800Mhz one may do a lot more work per cycle", or "they could be designed for diffrent markets". However in this case both are for the same market, and McKinley is likely to do more work per cycle. The Itanic is a shammbling disaster, if anyone but Intel was behind it, they probbably would be bankrupt by now.

    That said, there is a good chance that both IA64 systems have the same memory interface, making it a useful test run to design multi-way systems around this. It also has probbably been in the works a lot longer then it has been known that the Itanic is a dog.

    There is also the chance that Intel is subsidising (or outright funding!) this thing. That makes it less of a risk to build.

  2. Re:Let's all hope that.. by stripes · · Score: 5
    [Let's all hope that..] they do a better job with these then with the Intel based workstations they sell now.

    Well the current Intel boxes don't have any more CPUs then you can get elsewhere (they do have more memory bandwidth, by a factor of around 3). The O3000 supports 100s of CPUs, if the O3000 IA64 does as well they will at least have a nice story to sell people on. I don't know if it something people are willing to pay a lot for, but time will tell that.

    It is interesting that they will port Linux to it. As far as I know Linux isn't tuned to work in a large NUMA (Non-Uniform Memory Access -- local memory in maybe 5ns, memory from a few racks away at 500ns) with 100s of CPUs. It will be interesting to see what they change to make it happy. Esp if they don't go the simple route (treating it as a bunch of total different machines with a fast network).

    What I would really like to see is a port of Linux to the MIPS based machines though. If SGI tweaked it all out that would be sweet.

    I don't really see why they don't do that either. It isn't like there is anything special about the IA64 or MIPS that makes Linux or IRIX better on one or the other. Even if they think putting Linux on the MIPS boxes will scare off existing IRIX users, won't porting Linux to the IA64 O3000, but not porting IRIX to it will be even worse?

    SGI does lots of good and interesting stuff, but they really are inscrutable sometimes.

  3. Re:Why not for MIPS? by InsaneGeek · · Score: 4

    I asked this of our rep a few months ago, asking about the port work of Linux to Mips. What he said was that Irix has a buch of special magic going on in conjunction with the Mips CPU. i.e. make sure code gets to the cpu through at the same time, special read ahead things, tweaks with the bus and cpu, etc. to put them into a Linux system would fundamentaly change how things are currently done in most of the kernel . Think how hard they've been trying just to get big memory support into the kernel, could you imaging trying to get a completely different core change that would only help Mips cpu's.

    They are doing porting work, and Linux is running on Mips currently (so is NetBSD), but the special performance enhancements aren't in the kernel. Currently running Irix on a Mips Origin will be much faster than with Linux, but on the smaller workstations O2, Indigo II, etc. Linux will be faster (for non gui work) because it's so light-weight.

    Don't plan on getting a speedy Linux on your Onyx, Origin systems anytime soon, unless you want to take a big performance cut.