As the next round of NSF machines come online, this list will quickly be outdated. The least of these machines will have at least 1PB of *scratch* space. Tape libraries are going to be completely necessary rather quickly. SDSC's 19PB might seem like a lot, but we're constantly expanding that to deal with how quickly users dump data into it.
No, and I don't know why everyone thinks this will be a BlueGene/P. NCSA bid PERCS on this. It's fairly safe to bet such a machine would run AIX (you certainly don't want to waste a petaflop on linux.
I honestly haven't heard of anyone crying foul over NCSA getting their cake. Yes, it really really sucks that PSC and SDSC didn't (we were quite shocked), but it's that ORNL got track2 that digs you to the bone. Do you honestly think ORNL thought they'd get it? I bet there were as shocked as any of us about that. Yes, I'm whining about greed, the DOE has the deepest pockets, there's no need to go after NSF money. My sincere congratulations to NCSA, any number of other sites could have bid PERCS, but they didn't. It also helps to have state money backing you (this seems to be a point of contention, but Meecham wants his moneys worth, fair enough).
You're missing the mark completely here. The problem is that the DOE doesn't allow their machines to be open to public researchers. You don't get time on a DOE machine unless you're doing atomic research, etc. The NSF is inherently seperated from the DOE because of this fact. That's why we have BG/L at livermore, it's for the nuke guys. The public sector isnt' getting as many big machines anymore, just the military. So all those astrophysicists and earth sciences people won't get access period. This isn't something to be played off as a bunch of people whining, this is a serious change in the way the NSF does business. More to the point, it approaches a travesty for the scientific computing community at large.
This is a HUGE deal and, frankly, rather selfish of a DOE lab. There are a handful of non-DOE national labs out there that are struggling (I mean, very very close to closing doors) that needed this round of funding. Places like NCSA, PSC, SDSC. They're all shit outta luck now, cept for those NCSA jerks. Funny how NCSA suddenly got interested in very big iron when they've been keeping the mid-sized machines for the past decade..
it's an MPP arch, it's not about aggregate mem, it's about mem per core. We're talking about a class of machine that often sees multiple gigabytes per core. These aren't (efficient) shared memory machines.
Yep, check out the SCEC [http://www.scec.org/] project, for example. There are some fluid dynamics and astrophysics code as well. And, well, naturally this was all designed with nuke "preservation" code in mind...
rrr, to be out of NDA would be a wonderful thing but let's just say Sun needs to follow up on promises it made to customers... but a note on the availability of BG racks. You can buy even just 1 rack (we ran a rack of BGL for about 2 years(first installation outside of LLNL) before getting 2 more). but yes, you're completely correct, the problem is finding code that will both port and scale.
there's just a bit more memory, it's actually 512 or 1GB in BG/L. I imagine it'll be a bit more in P. We actually have researchers who cram their code that does need more memory into our 3 rack setup and understand that they just want the first few steps to run and don't care if they run out of memory as long as those first few steps run:)
Each of the 1024 2-way nodes on a single rack (2048 procs) is a powerpc440d (a cut down 440 w/ an extra FPU, the unfortunately named 'double-hummer'. Nodes are loaded onto a node board (16 nodes per board + 1 or 2 IO nodes, 16 node boards per midplane) that slides into the mid-plane (2 midplanes per rack). There are 3 networks, a mesh network (like noted by spatialguy) where you have a connection to each "nearest neighbor" node surrounding you, a 3d torus network (don't ask me, i just know you specify the dimensions 4x4x3 or something, I'm just an admin:) ) A torus can span multiple racks and I believe can even encompass the whole machine. There is also a vanilla gigabit ethernet network for doing things like NFS/GPFS/Lustre mounts (remember, no local disk).
Ah, so the heart burn is a side effect. I was put one straterra just a week or two ago. I tended to get horribly depressed and had what has been described as a "flattened affect" as well as the common jitters on most stimulants (i've been on most stimulant ADD medications out there). Straterra's been a fairly un-obtrusive medication that seems to help with concentration and not have as many side effects as other medications.
What you have to remember is that reactions to medications vary from case to case. ritalin was praised when it first came out as a cure-all, but often the side effects of it really ended up hurting the patient more than the benifits (again, depression, appetite suppression, flattened affect). my advise, find a competent nuero-psychologist that has a reputation (i have some suggestions if you're in the boston area).
you may be thinking of flourinert. This is actually used. I remember seeing a setup down at the San Diego Supercomputer Center that cooled itself by litterally passing flourinert over certain parts. iirc they called it, simply, a waterfall... and was on one of the Cray's
I had used this same basic idea in a project about 2 years ago in a computer project. I've still got the 3d models for it too. My partner for the project and myself were told to design a computer for a specific market and come up with design, info, and a marketing pitch. Let's see if i can find the model.... ha! found it! http://lenin.nu/~jwhite/graphics/gallery/comp. basically we had coolant being pumped through a specialized heatsync, through a set up copper coils with fans next to each coil drawing the heat away and pumped back through the heatsync. I always wanted to see it actually implemented.
actually, there was a difference in the implementations of RenderMan that Pixar uses and the one used in Final Fantasy. The major difference between the two was how the skin materials were handled (Final Fantasy actually introduced intentional "blemishes" into the skin textures in an attempt to make them more realistic (and pulled it off quite nicely).
This thing is hideously power hungry. Running winamp under WINE offers far better performance... how'd they manage to skrew this up so horribly? there's been a beta mac version for a while now that hasn't really seemed to be in development lately (it crashes when you try to access the menu).
*or* they know exactly what they're doing and you're bein an idiot and assuming the guys a fsckin moron. People dont go out and buy 10 node clusters of 2 way machines on a whim.
Slashdot Mirror
You'd be surprised how on the button you might be. I'm always hearing proposals that include "the first 3-10k processors off the line".
As the next round of NSF machines come online, this list will quickly be outdated. The least of these machines will have at least 1PB of *scratch* space. Tape libraries are going to be completely necessary rather quickly. SDSC's 19PB might seem like a lot, but we're constantly expanding that to deal with how quickly users dump data into it.
No, and I don't know why everyone thinks this will be a BlueGene/P. NCSA bid PERCS on this. It's fairly safe to bet such a machine would run AIX (you certainly don't want to waste a petaflop on linux.
I honestly haven't heard of anyone crying foul over NCSA getting their cake. Yes, it really really sucks that PSC and SDSC didn't (we were quite shocked), but it's that ORNL got track2 that digs you to the bone. Do you honestly think ORNL thought they'd get it? I bet there were as shocked as any of us about that. Yes, I'm whining about greed, the DOE has the deepest pockets, there's no need to go after NSF money. My sincere congratulations to NCSA, any number of other sites could have bid PERCS, but they didn't. It also helps to have state money backing you (this seems to be a point of contention, but Meecham wants his moneys worth, fair enough).
You're missing the mark completely here. The problem is that the DOE doesn't allow their machines to be open to public researchers. You don't get time on a DOE machine unless you're doing atomic research, etc. The NSF is inherently seperated from the DOE because of this fact. That's why we have BG/L at livermore, it's for the nuke guys. The public sector isnt' getting as many big machines anymore, just the military. So all those astrophysicists and earth sciences people won't get access period. This isn't something to be played off as a bunch of people whining, this is a serious change in the way the NSF does business. More to the point, it approaches a travesty for the scientific computing community at large.
This is a HUGE deal and, frankly, rather selfish of a DOE lab. There are a handful of non-DOE national labs out there that are struggling (I mean, very very close to closing doors) that needed this round of funding. Places like NCSA, PSC, SDSC. They're all shit outta luck now, cept for those NCSA jerks. Funny how NCSA suddenly got interested in very big iron when they've been keeping the mid-sized machines for the past decade..
I haven't heard yet, but will compute nodes and IO nodes still be basically the same sans/plus gigE? ie: will the IO nodes be 4 way 2/4G nodes?
it's an MPP arch, it's not about aggregate mem, it's about mem per core. We're talking about a class of machine that often sees multiple gigabytes per core. These aren't (efficient) shared memory machines.
Yep, check out the SCEC [http://www.scec.org/] project, for example. There are some fluid dynamics and astrophysics code as well. And, well, naturally this was all designed with nuke "preservation" code in mind...
rrr, to be out of NDA would be a wonderful thing but let's just say Sun needs to follow up on promises it made to customers... but a note on the availability of BG racks. You can buy even just 1 rack (we ran a rack of BGL for about 2 years(first installation outside of LLNL) before getting 2 more). but yes, you're completely correct, the problem is finding code that will both port and scale.
there's just a bit more memory, it's actually 512 or 1GB in BG/L. I imagine it'll be a bit more in P. We actually have researchers who cram their code that does need more memory into our 3 rack setup and understand that they just want the first few steps to run and don't care if they run out of memory as long as those first few steps run :)
Each of the 1024 2-way nodes on a single rack (2048 procs) is a powerpc440d (a cut down 440 w/ an extra FPU, the unfortunately named 'double-hummer'. Nodes are loaded onto a node board (16 nodes per board + 1 or 2 IO nodes, 16 node boards per midplane) that slides into the mid-plane (2 midplanes per rack). There are 3 networks, a mesh network (like noted by spatialguy) where you have a connection to each "nearest neighbor" node surrounding you, a 3d torus network (don't ask me, i just know you specify the dimensions 4x4x3 or something, I'm just an admin :) ) A torus can span multiple racks and I believe can even encompass the whole machine. There is also a vanilla gigabit ethernet network for doing things like NFS/GPFS/Lustre mounts (remember, no local disk).
it's not a button, it's a switch. and they now have a lil plastic sliding door that you have to flip open before you can hit the kill switch
There's been a opensource, cross-platform solution for a number of years (shameless plug), http://www.osmq.org/index.html
um, i may be mistaken, but this description sounds awfully similar to the backlit keyboards on the newest line of powerbooks.
appetite supression is bad when you're dealing with a 100lb kid w/ a fast metabolism.
Ah, so the heart burn is a side effect. I was put one straterra just a week or two ago. I tended to get horribly depressed and had what has been described as a "flattened affect" as well as the common jitters on most stimulants (i've been on most stimulant ADD medications out there). Straterra's been a fairly un-obtrusive medication that seems to help with concentration and not have as many side effects as other medications.
What you have to remember is that reactions to medications vary from case to case. ritalin was praised when it first came out as a cure-all, but often the side effects of it really ended up hurting the patient more than the benifits (again, depression, appetite suppression, flattened affect). my advise, find a competent nuero-psychologist that has a reputation (i have some suggestions if you're in the boston area).
yeah, cray has their "waterfall" system. Basically just letting liquid flow over the processor. And they look insanely cool.
you may be thinking of flourinert. This is actually used. I remember seeing a setup down at the San Diego Supercomputer Center that cooled itself by litterally passing flourinert over certain parts. iirc they called it, simply, a waterfall... and was on one of the Cray's
I had used this same basic idea in a project about 2 years ago in a computer project. I've still got the 3d models for it too. My partner for the project and myself were told to design a computer for a specific market and come up with design, info, and a marketing pitch. Let's see if i can find the model.... ha! found it! http://lenin.nu/~jwhite/graphics/gallery/comp. basically we had coolant being pumped through a specialized heatsync, through a set up copper coils with fans next to each coil drawing the heat away and pumped back through the heatsync. I always wanted to see it actually implemented.
actually, there was a difference in the implementations of RenderMan that Pixar uses and the one used in Final Fantasy. The major difference between the two was how the skin materials were handled (Final Fantasy actually introduced intentional "blemishes" into the skin textures in an attempt to make them more realistic (and pulled it off quite nicely).
This thing is hideously power hungry. Running winamp under WINE offers far better performance... how'd they manage to skrew this up so horribly? there's been a beta mac version for a while now that hasn't really seemed to be in development lately (it crashes when you try to access the menu).
*or* they know exactly what they're doing and you're bein an idiot and assuming the guys a fsckin moron. People dont go out and buy 10 node clusters of 2 way machines on a whim.
yellow dog is very closely based on RH... if i recall, to the point that if you look at some of the conf files, they're ripped right from RH
"Just what we need! A Druish princess!"