Slashdot Mirror

How can you say that after all SGI have done for Linux, and open standards in general?

SGI:

- Gave Linux XFS, arguably its fastest and most robust filesystem to date. Far, far more robust than reiser, and quicker than anything else except reiser4 (and then only sometimes), except on deletes where it is slow by design - SGI realised earlier than most that if you need a simple rule, it's pretty safe to assume that people just don't delete files often (excluding short-lived temporary files, which XFS handles _incredibly_ efficiently.) Just check out the low rate at which XFS volumes become fragmented to see how you can take advantage of putting a little thought into deleting files.

- Scaled Linux beyond 32 CPUs for the first time ever. And years later they still hold the record: 1,024 CPUs in the one computer with a single memory space. Nobody else comes close, and I do mean nobody. And this isn't just SGI lab stuff any more - NASA bought 20 of these computers to build the fastest computer on the planet that uses commercial microprocessors.

- Invented OpenGL (hint: what do you think the "Open" in "OpenGL" refers to? bonus marks: compare and contrast OpenGL and DirectX) together with the surrounding (open) glue like GLX. This is pretty much the only reason Linux boxes and Macs have decent 3D, and the only reason you can actually have a decent game of quake even if you're using a dumb terminal. Try playing Quake when connected to a Citrix box. Fun? Didn't think so.

- a bunch of other things I don't know about personally, but here you go anyway.

Anyway, since SGI's main role these days is selling IA64-based supercomputers and workstations, I hope Intel just buy SGI but let them continue to run independently so they can just keep on with all their good work. They provide a useful service to the Linux community, even if you never pay them a cent - this probably has something to do with their current share price (sadly). You might not use OpenGL, Itanium, massive shared memory systems or XFS but the odds are good that at least one of these is helping you, or at least some bugs SGI fixed while getting one of these working.

Mean anything for the STL? I mean, is SGI still working on the STL, and will it continue to keep its excellent documentation publicly (freely) available, etc?

In which way does EMACS copy Microsoft? Which Microsoft operating system runs on 512P system like this?

I'm not picking on you, but how would sgi come out a notebook? With 'their' linux, no less?

They make neat-o UNIX machines, but they're marketed to engineers who want to locate underground petroleum reserves, compute drag coefficients for jumbo jets, or render models of sarcogphagussed mummies from PET/CAT/MRI scans.

None of these jobs are all that portable. Plus, they require more muscle than a single 800 MHz R16000 can deliver, or a 2GHz Opteron, for that matter.

Linux? Sgi doesn't seem to have "their own", though they offer a "Choice of industry-standard Linux operating systems, fully supported by SGI"

PDF

I agree that it would be frickin fantastic to have a portable 'Visualization Workstation' from a company like Silicon Graphics.

It's easy to imagine an incredible screen, eye-catching industrial design, and a hearty, workstation-y robustness packaged up in a lap-sized container. Especially when you compare Apple to SGI as UNIX vendor = UNIX vendor.

You'd probably wind up with a snazzy little box with beauty all over the outside running IRIX or RHE at 800MHz and for $3,500.

We could do better pulling together around Athlon64 mobiles and Ubuntu or something.

C++ and the STL actually overcome both of these limitations. The more flexible of the two C++ sort functions looks like:

template <class Ran, class Cmp>
void sort(Ran first, Ran last, Cmp cmp)

By allowing for template iterators and a templated comparison operator, the compiler can optimize the comparisons and swaps away completely (assuming cmp is either a function object or an inline function).

Here is a detailed reference for std::sort.

Linux does indeed run on "serious" servers, such as this one, which can scale up to 512 processors per *system* (not per cluster). You can check some of its specs here. Relevant quotes:
"Up to 512 processors tightly coupled and operating under a single copy of the O/S"
"All resources in the system are managed through a single instance of the operating system--the industry's largest single system image."

I believe this constitutes "big iron".

Linux does indeed run on "serious" servers, such as this one, which can scale up to 512 processors per *system* (not per cluster). You can check some of its specs here. Relevant quotes:
"Up to 512 processors tightly coupled and operating under a single copy of the O/S"
"All resources in the system are managed through a single instance of the operating system--the industry's largest single system image."

I believe this constitutes "big iron".

You don't seem to be distinguishing between cards that have relatively minor optimisations to increase performance, such as the Intel and Broadcom ones you've mentioned, and the card that is the topic of this article, which runs the whole TCP/IP stack (ie. managing the connections including sequence numbers, round trip times etc. etc. etc. independently of the OS).

This article is about a separate card running the full TCP/IP stack, and it is this card that Phil Karn's criticisms apply to. I don't think can turn that into a criticism of all on card optimisations, such as TCP Segment Offload, which is what I understand is one of the the optimisations that the Intel and Broadcom cards implement. I think you're expanding his opinion wider than it probably is.

As an alternative example of opinions on NIC hardware optimisations, Dave Miller, the main networking subsystem maintainer for Linux, does not like TCP Offload Engines (which is what this article card is) at all, yet he's quite happy to come up with new and better ways to use TCP Segment Offload. Here's an example of his opinion on TOE. Here's his recent post regarding new TSO code he has written and how he is happy with its performance.

It's called "TCP Onloading". Here's an article about it (paid subscription required for the whole article - the abstract is free). Notice that it's coming from Intel research labs.

I could only read the abstract. I really can't see how to interpret it as saying that the whole TCP/IP stack was going to be moved on to the chipset.

You don't seem to be distinguishing between cards that have relatively minor optimisations to increase performance, such as the Intel and Broadcom ones you've mentioned, and the card that is the topic of this article, which runs the whole TCP/IP stack (ie. managing the connections including sequence numbers, round trip times etc. etc. etc. independently of the OS).

This article is about a separate card running the full TCP/IP stack, and it is this card that Phil Karn's criticisms apply to. I don't think can turn that into a criticism of all on card optimisations, such as TCP Segment Offload, which is what I understand is one of the the optimisations that the Intel and Broadcom cards implement. I think you're expanding his opinion wider than it probably is.

As an alternative example of opinions on NIC hardware optimisations, Dave Miller, the main networking subsystem maintainer for Linux, does not like TCP Offload Engines (which is what this article card is) at all, yet he's quite happy to come up with new and better ways to use TCP Segment Offload. Here's an example of his opinion on TOE. Here's his recent post regarding new TSO code he has written and how he is happy with its performance.

It's called "TCP Onloading". Here's an article about it (paid subscription required for the whole article - the abstract is free). Notice that it's coming from Intel research labs.

I could only read the abstract. I really can't see how to interpret it as saying that the whole TCP/IP stack was going to be moved on to the chipset.

Intel currently supports TCP segmentation offload (pushing the packet segmentation task from the TCP stack onto the hardware), and is moving to push the entire TCP stack to a dedicated processor + NIC combo.

I'm curious, where have Intel said they're going to push the whole TCP/IP stack onto a dedicated processor ? I've seen a bit of speculation they're going to do it, but not a statement from Intel themselves. Does one exist ?

I don't necessarily think that just because some customers have bought them, means its a good, well thought out and useful idea. A lot of people buy Britney Spears music, does that make it good or just popular ?

I suggest having a read of any threads on the Linux netdev mailing list, with "TOE" in the title. There has been quite a lot of discussion about the pros and cons, and it seems that the consensis of the main Linux kernel networking hackers is that TOE isn't a good idea.

Here is a quick TOE netdev search link for you.

Allow me to refute your alternative possible explanation with *facts*. :)

[snip netcraft]


Ha ha. Very funny. You know, it has been nearly a decade since you needed a really good OS to serve websites, right?

No, if you want to be a Linux contender, you have to do this, this, this, this, this, this, this, this, etc.

I don't see the word FreeBSD anywhere, do you? Did you think maybe SGI is "obscuring" FreeBSD because it runs on their 512 CPU SSI Servers so much better than Linux? Or the alternative explanation is that FreeBSD (even the current 6 branch) barely scales to 4 CPUs.

Oh, and don't get me started on clueless idiots.

I just installed Irix on a couple of old SGI machines, an Indigo2 Teal, and an Iris Indigo. I could not resist the looks of these boxes, SGI makes the most attractive computers, however I do suffer from Mac Cube lust. I am typing this on the Iris Indigo. Irix was the only system I ever saw that I never had to tell anything about the internet. It figured out my router etc and never asked me anything. Also when I went to set the clock it had figured out the time and date as well. With only 1 gig drives there is some juggling of software to run but I am really impressed with this system. I seem to be stuck with some older browsers for the web but I think these things are cool. I can't wait to get my hands on the 3d glasses to try. I like you, wish SGI had entered the lower cost market with a purple box or somthing. I guess it dates back to when that bonehead swung the company into going NT, then when it started to sink jumped into a cushy job at Mirosoft. A curse upon whats his name. Oh well I am still rooting for SGI they still make great stuff: http://www.sgi.com/

According to http://www.sgi.com/products/remarketed/onyx2/tech_ specs.html and http://graphics.tomshardware.com/graphic/20020625/ index.html let's compare some specs

IR Pixel fill, smooth, Z 224M to 448M

Pixel fill, textured, AA, Z 156M to 312M

Polygons/sec 10.9 M

Memory 64MB + 80-160MB

In 3D Mark 2001 SE (which is *lower* than theoretical numbers) Parhelia scores

FR Smooth 751 MP

FR Textured 2478! MP

Polygons/s 11.1M (with 8 lights!)

Memory 256 MB

not to mention a nice new Athlon will be an order of magnitude faster than whatever MIPS processor is in there. 3-D technology moves very quickly...

Err, what? All graphics cards that implement ARB_color_buffer_float has to implement IEEE 32-bit floats, as stipulated by ARB extension specification. (of course, this is assuming that the scientists are using the color buffer to encode information)
http://oss.sgi.com/projects/ogl-sample/registry/AR B/color_buffer_float.txt
Basically, any up-to-date ATi or NVidia gfx cards are capable of true IEEE 32-bit floating point numbers. What really worries me about the research is that they're not using 64-bit!

Well their new Altix line, which uses Itanium CPUs, is pretty slick. Do you want lots and lots of CPUs and a good NUMA architecture? Then check out what they did with Project Columbia. I'm a lot more impressed with SGI now than I was when they were in the business of making UNIX workstations that ran a really shitty version of UNIX (Irix sucks and blows at the same time) but which had shiny and colorful plastic cases and names such as "Octane", "Fuel" and "Viagra" (oh wait, I made that last one up).

SGI sells single machines with up to 16 GPUs and 96 PCI-X slots. Granted, they are a tad bit more expensive than your average desktop system.

SGI has been delivering workstations with several GPUs for many years. This is the configuration of their Prism Linux workstation:

* 4-16 ATI FireGL GPUs
* 16-256 Intel Itanium 2 CPUs
* Up to 6.1TB with RAM
* Up to 96 PCI/PCI-X slots (91 available)

http://www.sgi.com/products/visualization/prism/

And while everything you've just said is completely true, you've completely ignored the fact that SGI quickly filled in the gap with their first ever Linux workstation. You might like to visit sgi.com

And just a few weeks ago, Fujitsu announced an Itanium chipset, "Pleiades", supporting PCI-Express. Could make for some pretty serious workstations if they deicde to go that way.

In short, I hardly think it's beyond Apple's engineering ability to put together an Itanium workstation, and a fairly nice one at that. It's probably not likely to happen though, for price reasons more than anything else. I just don't see Apple moving into the $10k-$20k workstation market anytime soon.

For that price I'd rather get a used onyx.

He realised that if we made the slightest mistake under the terms of the GPL, even if it was only a perceived mistake, we'd have to spend the next 10 moonths dealing with these people. He canned the project immediately.

Hogwash. Own up to any problems, fix them and appologise. SGI's open letter is a perfect example of how to do things right. They never denied the problem, they quickly worked to verify the extent of the problem, the quickly fixed the problem, and they readily admitted that they were responsible.

Maui X-Stream had done nothing but deny and lie in the face of overwhelming evidence. They could have solved these problems months ago by either ceasing distribution or negotiating a license for the GPL code. Fools like this, who immediately discount the collective voice of free software developers, are getting what they deserve. If you steal GPL code then have the audacity to send out Cease and Desist letters, what exactly are you expecting the response to be?

um, the stats speak for themselves ~ 64-bit scalability to support files to 9 million Tb, filesystems to 18 million Tb ~ Instant data sharing without network mounts or data copies among all major OS: IRIX® Sun(TM) Solaris(TM) IBM® AIX® Windows® (2000/XP) Linux® (32 and 64 bit) Mac OS® X Unix® Flavors # Highly optimized distributed buffering techniques that provide the industry's fastest performance # High availability with automatic failure detection and recovery # Centralized, intuitive Java(TM) language-based management tools # POSIX® compliance that requires no application change So if you really do have large volumesof data and want 99.999% uptime, look at this and tremble in awe! http://www.sgi.com/pdfs/2508.pdf

Ummm... FSN is/was a *file manager*.

Jurassic Park had the most graphical unix system I have ever seen. This was in the early 90s.

It was still a real system; http://www.sgi.com/fun/freeware/3d_navigator.html

Johnny Nnemonic had almost 3D holographic.

Johnny Mnemonic is based on a SF book that takes place in the future. Nothing strange there.

Hacker the movie. The worst portrayal of computers period.

Well... I agree. It is silly, but it got some good stuff. The books they mention in it is real classics so they got something right at least :)

The irony is that it actually IS UNIX

I hate to retread a previous comment, but according to the movie "Jurassic Park," it was replaced by UNIX systems.

http://www.sgi.com/fun/freeware/3d_navigator.html

Switch to using FAM (File Activity Monitor) on both systems. It's a daemon implemented on top of kqueue (on *BSD) or dnotify/inotify (on Linux). It talks to instances on other machines to work properly across network filesystems. It also abstracts the underlying API for you. It just gives you a file handle to plug into your monitoring loop, and an API to call when it has input. No need to care if it's using kqueue, dnotify/inotify, or (in the worst-case scenario) a timed loop on the backend.

There's your practical answer. If you'd like to know why you can't directly see when a file handle grows with select/poll, read on.

The thing is that regular files and socket/pipes/character devices are treated in a fundamentally different way on Unix systems. Sockets have nonblocking IO and select/poll. Regular files have much less support - a separate, nasty async API on some systems, and inotify/dnotify on Linux/kqueue on BSD for notifications. (Something on IRIX...they built fam there, after all.) This doesn't make a lot of sense and people like DJB have argued that this doesn't have to be, but...well, that's still how it is.

kqueue is no exception. They've grouped a number of things into the same system call and made it more convenient to safely wait for several types of events at the same time, but you still can't treat them in the same way. On Linux the equivalent of kqueue is accomplished through:

• epoll
• dnotify/inotify signal handlers
• ptrace? I don't remember how you watch processes off-hand.
• more signals for async IO

The biggest pain there is handling signals and epoll stuff simultaneously in a correct manner. If you need to, I urge you to check out the documentation for my sigsafe library. It describes some things not to do plus a couple good ways: the self-pipe trick (a popular way if you're using select/poll/epoll) and my own sigsafe_* signal call wrappers.

What's it like to have a new release of your server operating system that isn't slower?

Standard in the Unix world.

There's no way you can dedicate a CPU to a particular application.. not in any form of pre-emptive OS.

What'd'ya mean "any form of pre-emptive OS"? Just because Microsoft doesn't do it doesn't mean it's not possible. You can certainly do this on Irix, for example. And I haven't looked at the Linux processor set tools, but I assume it's similar.

That 512 CPU setup is a custom fork of Linux made by SGI. Also, measuring scalability in number of CPUs is like measuring speed in MHz.

....and most of it is already in 2.6, and the rest is being integrated, so what's your point?

See all the benchmarks done during the 2.5 development to decide if linux "scales" or not.

Games that use OpenGL Performer can take advantage of multiple CPUs very easily. Some of the ones run by Disney do, for example.

http://www.sgi.com/products/visualization/realityc enter/

Yours for a small fee.

> 192 GB is no longer very large nowadays

Indeed.

SGI Altix 350, thanks to memory only nodes, (IINM, which I could very well be) goes to 384GB per system and only requires a single CPU and single disk. They even have a special on 'remarketed' ones at the moment.

The SGI Altix 350:

Advanced design that scales to 32 processors and 384GB of memory per system.

Actually, it is

irix: http://freeware.sgi.com/ this was provided by sgi via actual sgi paid employees. it's now about a year out of date, but irix is basically end-of-lifed already by SGI (details available to the curious, but more than is relevant here). i like their inst packages because they provide the patches they needed to get the default source to build.

aix: http://www.bullfreeware.com/ this is provided by a third party hardware vendor, and looks fairly up to date. i haven't used it in years but used to about 8 years ago with reasonable success.

and as mentioned above the solaris sites are very up to date.

Rather than "Starter Edition," here's some suggestions, if anyone from Redmond just happens to read this. (I know they won't do it - it's more a mental exercise while I eat)

1. Go download this, and make it natively multi-user if it isn't already. Give it a strong native security model, too...you can get some ideas here, and the best part is, they won't mind you doing that if you don't try and patent said ideas. Also, modularise your GUI, and don't prevent users from accessing the CLI when they want to.

2. Have the CLI composed of this and this for us CLI types.

3. Make the Add/Remove Programs panel essentially a net-aware frontend for either this or this.

4. Use this for hardware detection. Also re drivers, get rid of the suicidal policy of seeing third-party hardware vendors as the enemy, and actually support them...via tools, docs, etc. These people are your friends...they'll help you stay relevant.

5. Download this and use it as your default FS, and then get this and this, (although you already seem to know about this last one) and incorporate both of those into your stock UI. You've essentially got WinFS right there, without all the added complexity you'd no doubt throw into it if you tried to code it from scratch.

6. For the Agent angle, incorporate the last point, as well as putting help/docs in a non-binary format, making them searchable with this, converting said search results for use with this, and then use the AIML output as input for something like this. Also, instead of making the agent a tightly anthropomorphic personality, make it more generic, and more as though it's simply "the operating system" communicating with a user, rather than that dog or Clippit instead.

7. Give Outlook a major overhaul. This and this are examples of directions it IMHO should go in.

Just some random ideas, anywayz. Dreaming's fun. ;) I'll probably get modded Offtopic, but it was worth it.

>or 8way and bigger, go ahead and TRY to find >opensource unix (I am a big freebsd fan, btw) that >scales in cpu's as well as a sun or sgi or similar >box. SGI Altix and SGI Prism (admittedly both based on itanic), run Gnu/Linux, and since NASA's 10240 cpu machine is a cluster of SGI Altix nodes, I would say that is scaling well. http://www.sgi.com/products/servers/altix/ http://www.sgi.com/products/servers/prism/ I'm sure there's other manufacurers which scale almost as well but I can't be bothered going and digging them out.

>or 8way and bigger, go ahead and TRY to find >opensource unix (I am a big freebsd fan, btw) that >scales in cpu's as well as a sun or sgi or similar >box. SGI Altix and SGI Prism (admittedly both based on itanic), run Gnu/Linux, and since NASA's 10240 cpu machine is a cluster of SGI Altix nodes, I would say that is scaling well. http://www.sgi.com/products/servers/altix/ http://www.sgi.com/products/servers/prism/ I'm sure there's other manufacurers which scale almost as well but I can't be bothered going and digging them out.

And indeed it was. An SGI Irix system running 3D File System Navigator.

omicronish: "For years I thought all UNIX systems had cool graphical UIs like [in Jurrasic Park], and then I tried a real one and was disappointed by these crazy things called "characters"."

Queer Boy: "In the late 90s IRIX did have a graphical menu that was similar to that."

That would be the "3D File System Navigator", or FSN. It's still around, as is SGI, at least for now. This page tells about it and has some screenshots:

http://www.sgi.com/fun/freeware/3d_navigator.html

I played with it on the Ingdio 2 we had in the lab at Unnamed Univerity. Pretty useless, but fun for a few minutes.

"This is Unix. I know this!"

My roommate has an Indigo2 sitting next to his desk... We enjoy the high tech Jurassic Park graphics.

They say linux doesn't scale well. SGI has Linux systems with 256 cpus in a node. http://www.sgi.com/products/servers/altix/ Microsoft is only now getting a cluster version of their OS http://developers.slashdot.org/article.pl?sid=05/0 3/04/2134229&tid=201&tid=231&tid=156

Uhhm, you might want to visit this web page. Or maybe this one.

"Scalable 64-Bit Production-quality Linux Platform" is about right. 128 simultaneous processors? They've been selling boxes that do that for a couple of years now. They run Linux. All with the same kernel. Not a cluster. One big computer.

More recently, they've been selling 256-way systems and made-to-order 512-way systems. One kernel.

Uhhm, you might want to visit this web page. Or maybe this one.

"Scalable 64-Bit Production-quality Linux Platform" is about right. 128 simultaneous processors? They've been selling boxes that do that for a couple of years now. They run Linux. All with the same kernel. Not a cluster. One big computer.

More recently, they've been selling 256-way systems and made-to-order 512-way systems. One kernel.

This problem has already been solved by OpenGL Performer

Applications, even 'games', written using Performer, will immediately benefit from multiple CPUs.

Altix 350

Altix 3000

The Altix 350 can take something like 384 GB of RAM, and the 3000 can take 24 TB! That should hold you for a while. Maybe. ;)

Altix 350

Altix 3000

The Altix 350 can take something like 384 GB of RAM, and the 3000 can take 24 TB! That should hold you for a while. Maybe. ;)

FreeBSD can run on enterprise class hardware if you feel like buying enterprise class hardware.

No, actually it can't.

This, this, this, this, this, this, even this, and of course this is enterprise class hardware. FreeBSD runs none of them, so what enterprise class hardware *does* it run on?

FreeBSD doesn't even understand NUMA, which is basically indispensable for Opteron, POWER, or any serious Intel IA64 or x86 platforms.

FreeBSD can run on enterprise class hardware if you feel like buying enterprise class hardware.

No, actually it can't.

This, this, this, this, this, this, even this, and of course this is enterprise class hardware. FreeBSD runs none of them, so what enterprise class hardware *does* it run on?

FreeBSD doesn't even understand NUMA, which is basically indispensable for Opteron, POWER, or any serious Intel IA64 or x86 platforms.