I heard someplace that most of the recent probes run VxWorks. I'm pretty sure Mars Pathfinder did. If it's not VxWorks then it's some other similar real time OS.
I think you're missing the point. In code that is controlling guidance on a space ship, you don't *want* lots of features. You want *one* feature - getting from point A to point B on the right route. The more parts and code you throw at that problem, the more likely it is that the guidance will fail. We have bloated software because people want all kinds of features that they will never use. How many features of of Word do you actually *use*. If you stripped it down to that, I'd bet it would crash a whole lot less...
The big difference is service. SGI is not (and I suspect neither are the other vendors you mention) "just doing a wget for the latest beowulf tarball". When you buy a cluster from us, you can get a whole mess of stuff like service contracts. Also, when you take the thing out of the box (or large wooden shipping crate) it will "just work" and have a bunch of software like load balancing and batch scheduling with it. Yeah, this stuff is almost all open source and yeah, you can get it free on the net (probably off our website among other places) and for the majority of the people reading this, that is enough for them to build their own cluster in their spare time. The problem is businesses don't wanna do this. So the point is to add value to Linux/Beowulf/other stuff and resell it and make a proffit. Since one company does it, more follow.
Cray has an incredible reputation in the HPC business so I suspect that some places will buy clusters from them simply because they are Cray and have provided excellent service in the past.
Yes, it is. Which is why it probably won't be connected by ethernet. There are a variety of interconnects such as Myrinet coming on the market now that do something called "OS Bypass". Without too many boring details, basically this allows you to map a page on the local side that "pushes" data to the remote side when you do a write to it. It allows you to send data over the network without doing a call to the OS every time. That is actually what kills you on ethernet, not the fact that ethernet itself is all that slow (especially gigabit).
That said, at least for the time being, a single memory image system like the Cray T3D/T3E or the Origin line from us (SGI) has better latencies by a lot than Myrinet.
The interesting thing is that as these "OS Bypass" interconnects develop, they are going to get more and more like a standard memory interconnect in a single memory image system and we'll come full circle. But I digress.....
Ok, I don't know much about Java so I won't comment on that part of your statement, but I take exception to the comment about Solaris being the best multi-threaded OS out there. (Disclaimer: I work for SGI) The largest system Sun has is the UE10000 and it has at most 64 processors. SGI has built and sold 512 processor systems. Cray has sold 1500 processor systems (though Unicos/mk isn't really multithreaded in the classical sense). So, it seems to me that that title should go to either Irix or Unicos/mk. The Irix kernel is *very* well threaded.
Lastly, to get back to the topic at hand, the SGI compiler comes with a fairly nice multi-threaded debugger, though I haven't used it very much (I work in the kernel where you basically don't get a debugger other than "printf").
What if you wrote a book and I gave it away? Would you take that from me? The point isn't whether money is being made, the point is whether someone is infringing on Apple's intelectual property. It is most certainly not anyone's right to do that. Apple may choose to allow it for PR reasons, but that is all.
Well, the cluster is being installed by IBM, so I suspect they will supply the software shell wants, which is basically (probably) a compiler (probably Fortran) and an MPI or MPICH library. As for the exploration code, I suspect that Shell writes their own. Most really large supercomputer customers run their own code rather than canned software packages.
The Cray vector systems back to the X-MP, I think,
run something called Unicos. Before Unicos, there
was something called COS that I know very little
about. On the scalar machines, the T3D,based on Alpha ran almost no OS at all and required a C-90 as a front-end (that's one heck of a front end!). The T3E (also Alpha) ran Unicos/mk, a micro-kernel based version of Unicos (more like a complete re-write, but...). Basically, every processor ran a microkernel and then certain processors were responsible for taking care of the I/O and interactive jobs. The application processors used RPC's to the I/O processors to get data. The big difference between this and a cluster is that all memory on the system is visible to all the processors. I believe that jobs can migrate between CPU's on the T3E. The T3D and T3E are certainly not "clusters" in the sense of the cluster of individual boxes.
Yeah, NASA is the best at building space stations. That's why we've had an almost continuously manned, American built, space station for 13 years, right? Um, sorry, the Russians are better at it, they just have less money. NASA is very good, mind you, but the Russians have a lot of experience and I don't think I've seen anything about Mir killing anyone.
The only reason the idea of an international space station may be "stupid" is that there are multiple ground control groups with conflicting authority.
Until they boot it on a 1024 processor machine, they haven't supported 1024 processors. They are going to find out the same thing that we found out - large hardware exposes large numbers of bugs in your code. We will be (hopefully, pending software and also pending getting enough ceramic covers for our ASICs...) shipping an actual 1024p system in the next few months. We have already shipped a 512p system and have a bunch more on order (two more shipping next week).
I would disagree with that. At the time that Apple was losing the OS war, they had two platforms - Mac and Apple II. There is almost no way to be more open than the Apple II, yet the Mac certainly won that battle (Steve Jobs had more than a little to do with that, though). I think the real reason Mac "lost" was that it was too expensive. If Apple could have sold the Mac for the same price as a PC clone, I really think they would have won. In the end, the people making the majority of the purchasing decisions (ie non-geeks) don't care about open vs. closed, they care about price first and ease of use a far second. They would probably care about reliability more if they actually understood what it meant...
OS X runs a Mach microkernel at its core. This is not Unix. However, the way Mach works is that you put various "OS servers" on top of Mach. I think originally it was designed as (among other things) a way of running multiple OS "kernels" at once while Mach handled all the device driver interfaces and things like that. Part of OS X (since it was part of NeXTstep) is a BSD OS server (I think this is what BSD Lite is). Along with the OS server is a full BSD userland so you can pop open a window and run your favorite shell. Apple does all they can to mask the fact that deep down, sed/awk/grep and friends can run, but they are indeed there.
I imagine, though I'm not an expert on OS X's kernel architecture, that they will choose to do Cocoa (the official OS X API - or is it Carbon that's the API? Whichever...) as another server directly on top of Mach. So, native OS X aps won't go through the BSD kernel, but native BSD aps will.
I've seen a few prereleases of OS X and I was quite impressed. If I had a modern Mac, I'd definetly be running it. In the mean time, I'll stick with another OS that the Gartner Group thinks is going to go away - Irix (Hey Gartner guys! What OS is coming down the pipe that can do 1024 processors in a single kernel image like Irix and replacing it? What? There isn't one? Didn't think so:)
Re:November 1st, 2000 == September 2000??
on
IBM Releases AFS
·
· Score: 2
Yeah, that's right - every software project always comes out on time so IBM must be acting in bad faith. I'm sure all of *your* projects have been done on time. Oh wait, that's probably true because you've never been part of a large software project. It was late. Deal. A two month delay on a large project is hardly something to whine about.
Also, their lawyers probably had to go through all the code and clear it. That's why it took so long for us (SGI) to open source XFS.
Believe me, sound crews pick on guitarists all the time for stuff like that(*):) We also tend to pick on drummers a lot for being generally deaf...
As for geeks with good lighting, if someone buys me stage lights I'll happily light design my office:) Some of those LED ones with the 16 million colors would be nice. Verilites or Intelibeams would also work quite well.....
(*)The sound crew I was on in college used to have an old rack-mount disco ball controller with a lot of knobs with labels on them of different instruments. When someone would ask us to turn up a given instrument, we'd adjust the knob and people would go away. Of course, it wasn't hooked up to anything:)
Anyway, we now return you to your regular Microsoft flamage.
There are ways of addressing this issue too. The group I work in at SGI is responsible for (among many other things) Irix and Linux RAS features on our hardware. RAS is Reliability, Availability, and Serviceability. One of the things we observed is that the most likely case to get a double bit error (we have ECC memory in all our Origin servers/supercomputers) was grabbing a new page off the freelist and then bzero'ing it.
The theory on why this occurs is that the memory on the freelist isn't being accessed (well, ok, we have some bugs occasionally, but...:) and it degrades because of this. Since you don't care about the data on the page, it kinda sucks to panic during the bzero. So, Irix, starting with 6.5.7, knows how to "nofault" this bzero operation and if it fails, it grabs a new page off the freelist and discards the bad page. This is a feature we are thinking of adding to Linux. Other types of pages for which this same recovery can work are mapped files (ie, program and library text/read only data) and clean user data (ie, just swapped in and not yet modified). This is about the best way to solve the intermittant failure problem.
One other interesting thing we've noticed with RAM is that failure rates over time stay pretty much constant since, while manufacturing techniques are getting much better, the memories are getting larger. This causes failure rates to stay nearly flat.
Re:Job's business strategy - Focus on Customer
on
X On OSX Now Free
·
· Score: 2
I've seen the comment that Apple charged customers for a beta a lot. At least on the high end, the people who want early access to software end up paying more since they end up with more support. I still think you shouldn't charge for beta's, but it is hardly unusual.
Uh, don't most PC's go "beep" on startup? In fact, I've yet to meet one that didn't (well, I've met several that haven't but they had dead components). Of course, someone is probably going to tell me how to turn those off?:) (No points for the "disconnect the speaker answer") I don't see how the PC beep is better than the Mac chime.
I suppose on the PC the sound off the sound card is different than the sound off the mobo so you can seperate the two.
Well, one really good reason is so if we *find* Martian life, we'll be pretty sure that it is from Mars, or at least got there from ancient impacts with Earth and didn't tag along on the last space flight. This is why NASA doesn't want Gallileo to crash into Europa (or whichever moon they think has a chance of life).
So Quake isn't the big issue here. Oracle is the big issue. As is Sybase and DB2, etc. The problem is, at what point will ISV's say "this isn't Linux anymore"? The whole reason that large companies like us (SGI) and IBM, et. al. are going to Linux is to get more applications. If we issue the SGI patch for moster systems, we could do all kinds of things like rearrange and add locks, add kernel threading types, and make the kernel preemptible. Is that really still the Linux kernel in the eyes of Oracle? Probably not. Then we lose 'cause customers aren't going to buy from us to run Oracle if they can't get support from Oracle (whether they will buy from us to run Oracle anyway is another question).
The other reason that we are scared of the monster systems patch is the number of Linux kernels that come out. How often do we recheck the patch? Which kernels do we release the patch officially for? How do we decide? There are no really good answers to any of those questions which is why the big patch is to be avoided if at all possible.
The point of the hearings on the AOL/Time Warner deal is to ensure that the combined company, which will be *huge*, doesn't end up with market control of *everything*. It's like the deal that happened awhile back where two phone companies merged and one of them had to sell their cellular business.
At the moment, they aren't doing horrendously nasty things with AIM but what would you say if the following happened: AOL installs filters on IM (since they controll it) to block all messages that criticize every movie released by Warner Bros.? How many people would notice if they did that? Now you could argue that it's their service, they should be able to block whatever they want, but if there is no other service and no way for anyone to see how IM worked, what would prevent the censorship? Things like this are why the FCC gets involved.
To add a bit to this, Mach is a microkernel. That means that by itself, it isn't intended to be the sole kernel on the machine, but just provides the basic support for the hardware. It also provides very nice interprocess communication mechanisms.
On top of that, you write a kernel "server" that provides whatever kernel API you want. The BSD portion of MacOS X (and, of course, NeXTstep) runs as a kernel server on top of Mach. I think, though I'm not certain, that the stuff like Cocoa sit directly on Mach and not on top of the BSD layer (otherwise, there'd be little point to using Mach instead of a monolithic BSD kernel).
OSX/NeXTstep is not the only OS that uses Mach as a foundation. mkLinux is a Linux kernel server for the Mach microkernel. That's what the "mk" stands for. Cray apparently considered Mach as the microkernel foundation for Unicos/mk (the Cray T3E operating system) but went with another microkernel called Chorus instead.
This is the same kind of comment that pops up every time there's an article talking about real supercomputers. The big difference is that you will *never* see 3 gigaflops on a G4 outside of some specialized benchmark. When you're on a real Cray, you're usually striding through so much data that the processor cache is totally useless (vector Crays before the SV1 [or maybe the J90] didn't have a cache). That means that you fall back on the bandwidth of the memory bus, which was not designed for memory references on almost every clock cycle.
In addition, the Cray is going to push a lot more stuff between disk and memory than that G4. Not important? Well, try dealing with several hundred gigabyte files constantly and disk performance becomes an issue.
Basically, the Cray is designed to run at or near peak performance constantly. The G4 (and all other personal computers) are not.
Also, it takes 8 G4's to equal a Cray from 1991. That's pretty good for an almost 10 year old machine...
I am a kernel hacker, though mostly on Irix (just switching to Linux). A "race" is actually something slightly different (and is a bug 99% of the time): A race is when two processes both grab the same resource without a lock. In general, the system will die in eternal flame (or some other such bad thing) since the two processes will conflict with each other while they are both using a resource that they thing they own but really don't.
I think the improvements he is talking about are to make finer grained locks (ie, remove as many of the "lock_kernel()"'s as possible and make other locks lock off smaller areas). Another thing you can do to improve performance (in some cases) is to replace "spinlocks" (locks where the process sits and waits for the resource to become available not allowing anything else to happen on the processor) with "mutexes" (locks where the process will yield the processor if it is waiting for a resource).
In addition to that stuff, I think there have been general performance improvements like enhancements to the TCP/IP stack and the scheduler and such.
Uh, *every* public company exists solely to gain value for its shareholders. IBM's board of directors wouldn't be doing their jobs if they weren't making money. The same goes for Slashdot, which is part of a public company. What's the profit to VA to run a discussion site where they are occasionally flamed? To get their name in front of a lot of potential customers - that's why. Companies open source software only because they think that in some way they will profit from it whether it is by having other people fix bugs or from PR.
As for the difference between IBM and M$, IBM (at least now) makes money leagally. With M$ it's questionable (and if the current ruling stands, then they *are* making money illegally).
So the bottom line is, just because they do it only for money doesn't mean that it's not good for the people receiving it.
Remember that a number of media outlets sign NDA's to look at stuff also. I don't know for a fact that this is the case here, but what if signed an NDA to see the Cube before it was released so that they could get a story out about the release very quickly? Then, they certainly don't have the right to run a story on it that they agreed not to run. Now if Apple is suing other random news sites just for covering the issue, that's a whole other story.
One final point/question: Is running a picture of the cube different from running only a story (with no picture) that says "Apple is going to release a cube" different from running the picture? Couldn't you be violating Apple's trademarks/service marks/copyrights/patents/whatever protects product designs by using the image of their product without their permission? (This last bit is a question, not a troll) In my opinion, the two are different. But then again, I'm hardly an expert on any area of law:)
Slashdot Mirror
I heard someplace that most of the recent probes run VxWorks. I'm pretty sure Mars Pathfinder did. If it's not VxWorks then it's some other similar real time OS.
I think you're missing the point. In code that is controlling guidance on a space ship, you don't *want* lots of features. You want *one* feature - getting from point A to point B on the right route. The more parts and code you throw at that problem, the more likely it is that the guidance will fail. We have bloated software because people want all kinds of features that they will never use. How many features of of Word do you actually *use*. If you stripped it down to that, I'd bet it would crash a whole lot less...
Cray has an incredible reputation in the HPC business so I suspect that some places will buy clusters from them simply because they are Cray and have provided excellent service in the past.
That said, at least for the time being, a single memory image system like the Cray T3D/T3E or the Origin line from us (SGI) has better latencies by a lot than Myrinet.
The interesting thing is that as these "OS Bypass" interconnects develop, they are going to get more and more like a standard memory interconnect in a single memory image system and we'll come full circle. But I digress.....
Lastly, to get back to the topic at hand, the SGI compiler comes with a fairly nice multi-threaded debugger, though I haven't used it very much (I work in the kernel where you basically don't get a debugger other than "printf").
What if you wrote a book and I gave it away? Would you take that from me? The point isn't whether money is being made, the point is whether someone is infringing on Apple's intelectual property. It is most certainly not anyone's right to do that. Apple may choose to allow it for PR reasons, but that is all.
Well, the cluster is being installed by IBM, so I suspect they will supply the software shell wants, which is basically (probably) a compiler (probably Fortran) and an MPI or MPICH library. As for the exploration code, I suspect that Shell writes their own. Most really large supercomputer customers run their own code rather than canned software packages.
The Cray vector systems back to the X-MP, I think, run something called Unicos. Before Unicos, there was something called COS that I know very little about. On the scalar machines, the T3D,based on Alpha ran almost no OS at all and required a C-90 as a front-end (that's one heck of a front end!). The T3E (also Alpha) ran Unicos/mk, a micro-kernel based version of Unicos (more like a complete re-write, but...). Basically, every processor ran a microkernel and then certain processors were responsible for taking care of the I/O and interactive jobs. The application processors used RPC's to the I/O processors to get data. The big difference between this and a cluster is that all memory on the system is visible to all the processors. I believe that jobs can migrate between CPU's on the T3E. The T3D and T3E are certainly not "clusters" in the sense of the cluster of individual boxes.
The only reason the idea of an international space station may be "stupid" is that there are multiple ground control groups with conflicting authority.
Until they boot it on a 1024 processor machine, they haven't supported 1024 processors. They are going to find out the same thing that we found out - large hardware exposes large numbers of bugs in your code. We will be (hopefully, pending software and also pending getting enough ceramic covers for our ASICs...) shipping an actual 1024p system in the next few months. We have already shipped a 512p system and have a bunch more on order (two more shipping next week).
I would disagree with that. At the time that Apple was losing the OS war, they had two platforms - Mac and Apple II. There is almost no way to be more open than the Apple II, yet the Mac certainly won that battle (Steve Jobs had more than a little to do with that, though). I think the real reason Mac "lost" was that it was too expensive. If Apple could have sold the Mac for the same price as a PC clone, I really think they would have won. In the end, the people making the majority of the purchasing decisions (ie non-geeks) don't care about open vs. closed, they care about price first and ease of use a far second. They would probably care about reliability more if they actually understood what it meant...
I imagine, though I'm not an expert on OS X's kernel architecture, that they will choose to do Cocoa (the official OS X API - or is it Carbon that's the API? Whichever...) as another server directly on top of Mach. So, native OS X aps won't go through the BSD kernel, but native BSD aps will.
I've seen a few prereleases of OS X and I was quite impressed. If I had a modern Mac, I'd definetly be running it. In the mean time, I'll stick with another OS that the Gartner Group thinks is going to go away - Irix (Hey Gartner guys! What OS is coming down the pipe that can do 1024 processors in a single kernel image like Irix and replacing it? What? There isn't one? Didn't think so :)
Also, their lawyers probably had to go through all the code and clear it. That's why it took so long for us (SGI) to open source XFS.
As for geeks with good lighting, if someone buys me stage lights I'll happily light design my office :) Some of those LED ones with the 16 million colors would be nice. Verilites or Intelibeams would also work quite well.....
(*)The sound crew I was on in college used to have an old rack-mount disco ball controller with a lot of knobs with labels on them of different instruments. When someone would ask us to turn up a given instrument, we'd adjust the knob and people would go away. Of course, it wasn't hooked up to anything :)
Anyway, we now return you to your regular Microsoft flamage.
The theory on why this occurs is that the memory on the freelist isn't being accessed (well, ok, we have some bugs occasionally, but... :) and it degrades because of this. Since you don't care about the data on the page, it kinda sucks to panic during the bzero. So, Irix, starting with 6.5.7, knows how to "nofault" this bzero operation and if it fails, it grabs a new page off the freelist and discards the bad page. This is a feature we are thinking of adding to Linux. Other types of pages for which this same recovery can work are mapped files (ie, program and library text/read only data) and clean user data (ie, just swapped in and not yet modified). This is about the best way to solve the intermittant failure problem.
One other interesting thing we've noticed with RAM is that failure rates over time stay pretty much constant since, while manufacturing techniques are getting much better, the memories are getting larger. This causes failure rates to stay nearly flat.
I've seen the comment that Apple charged customers for a beta a lot. At least on the high end, the people who want early access to software end up paying more since they end up with more support. I still think you shouldn't charge for beta's, but it is hardly unusual.
Uh, don't most PC's go "beep" on startup? In fact, I've yet to meet one that didn't (well, I've met several that haven't but they had dead components). Of course, someone is probably going to tell me how to turn those off? :) (No points for the "disconnect the speaker answer") I don't see how the PC beep is better than the Mac chime.
I suppose on the PC the sound off the sound card is different than the sound off the mobo so you can seperate the two.
Well, one really good reason is so if we *find* Martian life, we'll be pretty sure that it is from Mars, or at least got there from ancient impacts with Earth and didn't tag along on the last space flight. This is why NASA doesn't want Gallileo to crash into Europa (or whichever moon they think has a chance of life).
The other reason that we are scared of the monster systems patch is the number of Linux kernels that come out. How often do we recheck the patch? Which kernels do we release the patch officially for? How do we decide? There are no really good answers to any of those questions which is why the big patch is to be avoided if at all possible.
At the moment, they aren't doing horrendously nasty things with AIM but what would you say if the following happened: AOL installs filters on IM (since they controll it) to block all messages that criticize every movie released by Warner Bros.? How many people would notice if they did that? Now you could argue that it's their service, they should be able to block whatever they want, but if there is no other service and no way for anyone to see how IM worked, what would prevent the censorship? Things like this are why the FCC gets involved.
On top of that, you write a kernel "server" that provides whatever kernel API you want. The BSD portion of MacOS X (and, of course, NeXTstep) runs as a kernel server on top of Mach. I think, though I'm not certain, that the stuff like Cocoa sit directly on Mach and not on top of the BSD layer (otherwise, there'd be little point to using Mach instead of a monolithic BSD kernel).
OSX/NeXTstep is not the only OS that uses Mach as a foundation. mkLinux is a Linux kernel server for the Mach microkernel. That's what the "mk" stands for. Cray apparently considered Mach as the microkernel foundation for Unicos/mk (the Cray T3E operating system) but went with another microkernel called Chorus instead.
In addition, the Cray is going to push a lot more stuff between disk and memory than that G4. Not important? Well, try dealing with several hundred gigabyte files constantly and disk performance becomes an issue.
Basically, the Cray is designed to run at or near peak performance constantly. The G4 (and all other personal computers) are not.
Also, it takes 8 G4's to equal a Cray from 1991. That's pretty good for an almost 10 year old machine...
I think the improvements he is talking about are to make finer grained locks (ie, remove as many of the "lock_kernel()"'s as possible and make other locks lock off smaller areas). Another thing you can do to improve performance (in some cases) is to replace "spinlocks" (locks where the process sits and waits for the resource to become available not allowing anything else to happen on the processor) with "mutexes" (locks where the process will yield the processor if it is waiting for a resource).
In addition to that stuff, I think there have been general performance improvements like enhancements to the TCP/IP stack and the scheduler and such.
As for the difference between IBM and M$, IBM (at least now) makes money leagally. With M$ it's questionable (and if the current ruling stands, then they *are* making money illegally).
So the bottom line is, just because they do it only for money doesn't mean that it's not good for the people receiving it.
One final point/question: Is running a picture of the cube different from running only a story (with no picture) that says "Apple is going to release a cube" different from running the picture? Couldn't you be violating Apple's trademarks/service marks/copyrights/patents/whatever protects product designs by using the image of their product without their permission? (This last bit is a question, not a troll) In my opinion, the two are different. But then again, I'm hardly an expert on any area of law :)