in the limit, the packaging, power, network, and thermal costs will drive the individual nodes in any cluster to be 'fatter'. its the same thing as today, but the crossover point will change. there may be some short term bumps due to volume issues, but i think you're wrong. they aren't disjoint architectures but a continuum
the technology and architecture were beautiful. the execution and business planning were poor. because it was such a huge and underfunded effort to get the whole thing (os, compiler, processor, network) brought up from scratch, they lagged current technology at both of the two introductions. stability was a problem.
still, a terrible shame. a testament to the failings of the short-term investment model.
the compiler did automatic parallelization, but only really well for HPC-style loop nests. if you weren't running parallel code, you really suffered, because the individual thread execution rates were so poor, and they ran uncached (one of the nice things about the model is that they used concurrency to hide memory latency, but if you didn't have it to exploit...)
yeah. but they never got a very good set of kernels. because of the nature of the flaw(s), the psfs varied across the image and didn't do so continuously. they got part of the way there, but it was never even close. if only there had been a decent test pattern set lying around in space.
actually it wasn't just NASA engineers. they had an open call for help, and alot of people worked on the problem. which was very cool.
but those programmers need managers. and those managers need managers, and the managers managers need secretaries. and you need legal and financial people to deal with a government contract of that size. and its likely you are going to license some of the final product, and subcontract some other parts, each of which has their own overhead (including margins).
at least half of those programmers are going to be completely useless. more than 1/4 barely functional.
and those facilities you are paying for (including IT, sorry slashdot) are as likely to be a drag on progress rather than of any benefit
and all of those managers are so busy lying and playing politics and jockeying for position that no one above them can really tell what is going on, and so critical pieces just dont happen and the whole house of cards falls apart.
large scale software development is heinously inefficient and risky. you can tell i'm not bitter.
not to excuse the $100M, but the model is flawed, and they insist on pursuing it because its the only thing they know.
thats kind of a silly comment. the kinesis helps alot by chording off of the thumb, alowing the hand to stretch across the palm, which is a much easier motion than stretching out the fingers and trying to chord that way.
i've been using one for >10 years; they help alot.
since tridge was a contractor, and working on this in his own time, in what way is he bound by general agreements that osdl may have made? in what way is linux or osdl responsible for what he might choose to do?
funky? torus? torroidal meshes are as old as the hills. meshes dont have any very interesting properties, except that they have a straightforward mapping into large machines laid out in a grid.
hypercubes are nice, but the wiring gets messy at scale
clos gets you log hops for arbitrary base, and simplified routing.
yes. exactly. the real problem is the state of basic research and funding.
having darpa, onr, doe, and that whole crew fund research has always warped things. while it did mean that more money was going into cs, the military spin and the very loose kind of peer review resulted in alot of that money going to stupid things. and most of the good basic work had to be disguised by putting pictures of tanks and helicopters on slides and using the word 'warfighter' alot.
actually, this is a really interesting contradiction. its exactly this real world slop that i would want ga/gp to adapt to handling, not the kind of closed and simple world that i might be able to program to in the normal fragile way.
not that i dont agree with you about numbers of generations and population size.
umm. like other secret key technologies, its probably quite useful as a bulk data encryption after a session key has been negotiated using public/private
because of the massive cost of spinning a custom architecture, with all of the associated software, and the quite stunning investment by the commodity processor industry, this game is very much over.
unless you have a completely novel computing architecture and someone with alot of money and a specialized need willing to pay for it.
i agree with you completely. but when i taught, all the students bitched that i didn't have slides on the web. and all the faculty gave me dirty looks, as if i was being lazy.
i have to strongly agree. i've been doing languages and operating systems, networks, graphics, essentially everything, for a long time now. except for databases. i've always had scorn for databases. until i needed one that was fault tolerant and scaled and had certain distribution properties. and now i'm several months in, and its by far the most interesting and difficult distributed systems problem i've ever looked at. i have alot of catching up to do with the state of the art. and i'm regretting not having at least mastered the basics.
actually given the state of things i think Aunt Sally would blame the site for not working with her 'perfectly normal' computer. it seems thats how we got here in the first place.
does anyone know what one uses to program this thing? i found some references to ibm journal publications that are accepted, but not in print yet, and a low level (but nicely simple) message passing library. is it just nasty MPI?
yes. exactly, its just SMP on a single die. nothing to see.
except that its a very concrete example of the trend that started when intel decided not to chase after 4GHz.
this year, its duals. but next year it will be quads, and number-of-cores is going to be the dominant growth mode for commodity parts.
which means that performance sensitive applications are finally going to have to start exploiting concurrency. which is a big change indeed.
as pointed out elsewhere, it also means that vendors and customers are going to see a much longer lifetime for their system designs. which is going to have all sorts of interesting second order effects. cheap, no margin systems. truly modular architectures (ht, pci-e)
exactly what criteria are used to distinguish between a 'legitimate business' and someone who is going to use the information to steal my identity. or someone who, inadvertently or not, will pass the data to someone who is. the whole model is flawed.
Slashdot Mirror
in the limit, the packaging, power, network, and thermal costs will drive the individual nodes in any cluster to be 'fatter'. its the same thing as today, but the crossover point will change. there may be some short term bumps due to volume issues, but i think you're wrong. they aren't disjoint architectures but a continuum
the technology and architecture were beautiful. the execution and business planning were poor. because it was such a huge and underfunded effort to get the whole thing (os, compiler, processor, network) brought up from scratch, they lagged current technology at both of the two introductions. stability was a problem.
still, a terrible shame. a testament to the failings of the short-term investment model.
the compiler did automatic parallelization, but only really well for HPC-style loop nests. if you weren't running parallel code, you really suffered, because the individual thread execution rates were so poor, and they ran uncached (one of the nice things about the model is that they used concurrency to hide memory latency, but if you didn't have it to exploit...)
i'm pretty sure you can evaluate conservative fixed point analysis in parallel if you have a fine-grained machine (like an smt one)
yeah. but they never got a very good set of kernels. because of the nature of the flaw(s), the psfs varied across the image and didn't do so continuously. they got part of the way there, but it was never even close. if only there had been a decent test pattern set lying around in space.
actually it wasn't just NASA engineers. they had an open call for help, and alot of people worked on the problem. which was very cool.
but those programmers need managers. and those managers need managers, and the managers managers need secretaries. and you need legal and financial people to deal with a government contract of that size. and its likely you are going to license some of the final product, and subcontract some other parts, each of which has their own overhead (including margins).
at least half of those programmers are going to be completely useless. more than 1/4 barely functional.
and those facilities you are paying for (including IT, sorry slashdot) are as likely to be a drag on progress rather than of any benefit
and all of those managers are so busy lying and playing politics and jockeying for position that no one above them can really tell what is going on, and so critical pieces just dont happen and the whole house of cards falls apart.
large scale software development is heinously inefficient and risky. you can tell i'm not bitter.
not to excuse the $100M, but the model is flawed, and they insist on pursuing it because its the only thing they know.
thats kind of a silly comment. the kinesis helps alot by chording off of the thumb, alowing the hand to stretch across the palm, which is a much easier motion than stretching out the fingers and trying to chord that way.
i've been using one for >10 years; they help alot.
how is this iWarp-based? the iwarp did register-to-register 'systolic' style communication, not async bulk transfer.
even cm wouldn't make it 'high resolution'. mm anyone?
you mean like latex could do 20 years ago
since tridge was a contractor, and working on this in his own time, in what way is he bound by general agreements that osdl may have made? in what way is linux or osdl responsible for what he might choose to do?
actually its obscenely expensive. it was meant to be "best-of-breed". it just failed to be so.
funky? torus? torroidal meshes are as old as the hills. meshes dont have any very interesting properties, except that they have a straightforward mapping into large machines laid out in a grid.
hypercubes are nice, but the wiring gets messy at scale
clos gets you log hops for arbitrary base, and simplified routing.
yes. exactly. the real problem is the state of basic research and funding.
having darpa, onr, doe, and that whole crew fund research has always warped things. while it did mean that more money was going into cs, the military spin and the very loose kind of peer review resulted in alot of that money going to stupid things. and most of the good basic work had to be disguised by putting pictures of tanks and helicopters on slides and using the word 'warfighter' alot.
actually, this is a really interesting contradiction. its exactly this real world slop that i would want ga/gp to adapt to handling, not the kind of closed and simple world that i might be able to program to in the normal fragile way.
not that i dont agree with you about numbers of generations and population size.
umm. like other secret key technologies, its probably quite useful as a bulk data encryption after a session key has been negotiated using public/private
because of the massive cost of spinning a custom architecture, with all of the associated software, and the quite stunning investment by the commodity processor industry, this game is very much over.
unless you have a completely novel computing architecture and someone with alot of money and a specialized need willing to pay for it.
i agree with you completely. but when i taught, all the students bitched that i didn't have slides on the web. and all the faculty gave me dirty looks, as if i was being lazy.
computers in companies make people more productive?
i have to strongly agree. i've been doing languages and operating systems, networks, graphics, essentially everything, for a long time now. except for databases.
i've always had scorn for databases. until i needed one that was fault tolerant and scaled and had certain distribution properties. and now i'm several months in, and its by far the most interesting and difficult distributed systems problem i've ever looked at. i have alot of catching up to do with the state of the art. and i'm regretting not having at least mastered the basics.
actually given the state of things i think Aunt Sally would blame the site for not working with
her 'perfectly normal' computer. it seems thats how we got here in the first place.
no, intel wasn't. a better candidate would be the tera MTA processor. but it wasn't the first either.
does anyone know what one uses to program this thing? i found some references to ibm journal publications that are accepted, but not in print yet, and a low level (but nicely simple) message passing library. is it just nasty MPI?
the precursor to vrml was actually called weboogl,
which just rolls off the tongue
yes. exactly, its just SMP on a single die. nothing to see.
except that its a very concrete example of the trend that started when intel decided not to chase after 4GHz.
this year, its duals. but next year it will be quads, and number-of-cores is going to be the dominant growth mode for commodity parts.
which means that performance sensitive applications are finally going to have to start exploiting concurrency. which is a big change indeed.
as pointed out elsewhere, it also means that vendors and customers are going to see a much longer lifetime for their system designs. which is going to have all sorts of interesting second order effects. cheap, no margin systems. truly modular architectures (ht, pci-e)
exactly what criteria are used to distinguish between a 'legitimate business' and someone who is going to use the information to steal my identity. or someone who, inadvertently or not, will pass the data to someone who is. the whole model is flawed.