So, what does this say? To me, it says that SMP and mulitcore is here to stay.
I disagree. Well yes, multicores are coming, but for other reasons related to VLSI and microarchitecture scaling.
What this result tells is a little about cpu binning, but mostly about the willingness of the market to pay for the product. So there's lots of wieners out there who are prepared to pay through the nose for bragging rights. Hardly surprising. It's the same logic that sells e.g. sports cars.
Actually, many "normal" architectures have FP MAC units instead of MUL and ADD (at least IA64, PA-RISC, x86 being the major one without it). Many algorithms used in FP calculations can be efficiently implemented in terms of FMAC operations (matrix algebra, FFT etc.). Also, it is more accurate since it avoids the double rounding associated with a separate ADD and MUL, and thus also many "intrinsic" operations such as division and sqrt can be more efficiently implemented.
(Oh, wait! They don't teach Rhetoric in college anymore! Only a few young whipersnappers may know what "ad hominem" is! The rest are playing "Doom" or watching TV.)
Oh, I don't know if playing doom is so bad, considering that the people who claim to know rhetoric and are proud to display it seem to spend all their time on usenet on debates like:
>>>>>>>>>>> [SNIP long winded argument] >>>>>>>>>> Non sequitur. >>>>>>>>> Ah, using fancy latin words instead of offering some counterargument. Could you answer the question? >>>>>>>> Yes, I could. >>>>>>> But you didn't. Seriosuly, get a life dude! >>>>>> Ad hominem. >>>>> No, that wasn't an ad hominem. This however, is: Oh, fuck off! >>>> Ad hominem. >>>Good, you catch on quickly! Now, could you answer my question please? >>Yes, I could! >Well, do it then! Non sequitur.
With the shuttle, if the rocket explodes then everybody dies. With a design which places the crew capsule at the top, and an escape tower attached to it, then when the rocket explodes, everybody gets a really exciting ride and has to abort the mission, but they land safely. There is an enormous difference between low "mission success" reliability and low "the crew survives" reliability. The shuttle had the latter, this design should only have the former.
VoOk: Add to that: One booster fails to fire on launch of the shuttle: The shuttle, in a wide arc, propelled with the remaining thruster, crashes and explodes. One booster fails to fire here: The astronauts use the ladder to get back down because the rocket doesn't move an inch.
And to add further, a capsule heatshield is pretty well protected during launch, as opposed to the shuttle. Also, if it's an ablative shield it can probably take quite a beating compared to the RCC shingles the shuttle uses.
The special thing about InfiniPath is that the adapter is not a PCI-(e|X) card but rather connects directly to the HyperTransport interface on the cpu (requiring a special MB with a "HTX" connector), giving slighlty lower latency than a normal IB adapter.
Well, unlike most cluster interconnects InfiniPath is special in that it connect directly to the HyperTransport interface on the Opterons, without going via PCI(-e|-X). So it won't work on e.g. a Xeon.
But yeah, from a user perspective Infinipath is just like any other cluster interconnect. Perhaps a bit lower latency, but nothing dramatic (as in order(s) of magnitude lower latency).
I've had people contribute code to PMD and say they were only contributing it because they felt the BSD license avoided any possible obligations on their part.
Just like there's plenty of people who only contribute to GPL projects since they don't want "evil corporations" stealing their code.
You can find fanatics driven by ideology rather than common sense in both camps. That's hardly something to cheer about.
I work in a lab at the University of Washington where we are working on _implementing_ a different data flow machine that shares some of fundamentals with the UT machine.
So, which one is better?;-)
For a more serious question, I read the trips overview paper on their site and it all seems to make a lot of sense. So why aren't dataflow machines mainstream? The first papers were published in the early 1980s, not much later than risc started to make some noise.
I don't know if GNU malloc uses mmap() or brk() for its allocation, but in both cases small memory chunk that the user allocates are taken from bigger, contiguous blocks of memory.
It uses mmap() for big allocs (IIRC the threshold is 4 MB) and brk() for smaller ones.
There would be one solution, and that's using different arenas, or memory regions for allocation. For instance every window might have its own allocation region, so when you close the window/document, the memory BLOCK is freed.
Something like memory pools, used e.g. by apache portable runtime, gcc, and whatnot. Or the hierarchical allocator used by samba..
I really wish Java apps, Cocoa apps, and other (Mozilla) would do this, as they seem to suffer from this fragmentation problem, only increasing their memory usage, even after closing all documents/windows etc.
Actually, Java being garbage collected and not having "raw" pointers can avoid fragmentation by repacking heap memory. Of course the JVM must still use malloc() or mmap()/brk() directly to get the memory from the OS.
The upcoming GCC 4.1 release will include a stack protector. Basically it's a reimplementation of the old propolice patch.
Hopefully mainstream distros that have been wary of propolice will start using this new feature. And perhaps glibc malloc will borrow a few tricks from this new openbsd malloc too.
There is no solid definition of what differentiates a cluster
Exactly my point.
but it's a bit misleading and confusing to refer to an architecture that doesn't have any of the characteristics of a traditional cluster as a cluster.
But then, what are the traditional characteristics of a cluster? And why haven't those become the normal definition then?
Most clusters have a star or bus network topology of Infiniband, Myrinet, or Ethernet. An X1 uses a 3D torus of redundant, high-bandwidth links.
I don't think network topology is what defines a cluster. Consider the Scali Dolphin, which at least a few years ago was a relatively popular cluster interconnect. Depending on the model, it was a 1d, 2d or 3d torus. Or the SX-6/8, which uses a star topology with a crossbar switch.
Each node is a board in a backplane, not just a rackmount system.
Fair enough. So I agree the X1(E) is not a cluster then, perhaps "vector MPP" is more appropriate. But again, what about the SX-6/8? Each 8 CPU node is a separate cabinet, connected via the network switch. Starts to sound like a cluster to me.
While defining a cluster as a machine where each node has its own OS is a convienient definition, that also includes things which clearly aren't true clusters like networks of workstations and many MPPs.
Indeed. Given the lack of a solid definition, then perhaps one shouldn't go and label other people as idiots just because they use a slightly different definition of cluster?;-)
This part of the article makes it plain that the author really doesn't know much about HPC systems. A vector machine uses vector processors. That's it. It says nothing about whether the machine is a cluster or MPP and what kinds of programming models it supports (distributed memory, shared memory, etc.).
The SX-6/8 and X1/X1E are not clusters.
Now that's a bit inconsistent. First you say that there's no agreed upon taxonomy of HPC systems, and then you blast the parent poster for not using YOUR definition of a cluster.
Another AFAIK common definition of a cluster is a set of independent computers (each with its own memory, OS image etc.) connected via some network. By that definition, almost every machine on the top500 is a cluster, including the NEC and (I think) Cray vectors.
They have enormous memory bandwidth (a necessity on vector machines) and are distributed-memory machines. UPC and Co-Array Fortran run very well on them because the address reach of any one processor covers the whole machine and maps extremely well onto the GAS programming model.
BTW, IIRC according to NEC documents the recommended programming model for the SX-6 and SX-8 is a combination of OpenMP and MPI. In that sense it's no different than a normal "cluster of SMP:s". Don't know about the X1, though Cray's been making noise about CAF, for good reason I think.
Nah, it's all about using the right tools for the job.
Clusters are a good thing, as they provide a very cost-effective platform for running codes with modest communication requirements. Just like running a communication-intensive code on a cluster will limit performance, running a code with little communication on a "real" supercomputer is a waste of money.
The sad thing about the current "HPC crisis" is not the rise of clusters, but the use of clusters for tasks which they are ill suited for (typically, "grand challenge" tasks).
The crucial thing is to have an appropriate balance, so that the tasks which really need a "real" supercomputer can get one, and other tasks can run on clusters.
At this stage of the game, what is it that we can do on Mars with a manned mission that we cannot accomplish better, cheaper, and safer, with a robotic mission?
Capture the imagination of the public, and thus keep the pork flowing to a few huge contractors.
Scientifically, there's little that a manned mission can do which an unmanned can't, just like you say.
Apollo was able to put a couple of guys on the moon and get them back. In contrast, the shuttle is incapable of going beyond LEO. It's not exactly breaking new ground in terms of space exploration.
After all, thanks to the science the manned space program has done, we now know a little about how ant colonies behave in zero gravity. Surely that will bring new advances in biology as well as help solve the problems the world faces today. Those $250 billion were well spent. Compared to, say, $10 billion for ITER, and other big science projects.
Oh, wait, that ant experiment crashed with Columbia. Guess we won't know how they behave in zero G then. What a loss to humanity!
I mean, the space shuttle was an experiment in how to screw it up. Expensive, unsafe, you name it.
When we have non-chemical propulsion and somewhat fancier materials it might be time to take another look at the reusable space plane approach. Until then, the basic multi-stage capsule on top design seems to be the best choice in terms of operational cost and safety.
I read that the Russians and the EU are collaborating on some kind of shuttle. Well, not a shuttle as in the space shuttle but more like one of those lifting body CEV designs.
Still, ironic that the US is ditching the shuttle and going back to capsules as the Russians are going the opposite way.;-)
My generalization, huh? I wasn't the one saying that beowulf clusters run MPI.
Including this post, I have posted only two comments to this thread. And the other comment I posted was the comment saying PVM is dying. Well, whatever, it's not like it matters anyway.;-)
Slashdot Mirror
Actually, the system is provided with the IBM XL family of compilers.
So, what does this say? To me, it says that SMP and mulitcore is here to stay.
I disagree. Well yes, multicores are coming, but for other reasons related to VLSI and microarchitecture scaling.
What this result tells is a little about cpu binning, but mostly about the willingness of the market to pay for the product. So there's lots of wieners out there who are prepared to pay through the nose for bragging rights. Hardly surprising. It's the same logic that sells e.g. sports cars.
Actually, many "normal" architectures have FP MAC units instead of MUL and ADD (at least IA64, PA-RISC, x86 being the major one without it). Many algorithms used in FP calculations can be efficiently implemented in terms of FMAC operations (matrix algebra, FFT etc.). Also, it is more accurate since it avoids the double rounding associated with a separate ADD and MUL, and thus also many "intrinsic" operations such as division and sqrt can be more efficiently implemented.
Considering the sort of movies they dl and watch at those places I guess heat won't be a problem
(Oh, wait! They don't teach Rhetoric in college anymore! Only a few young whipersnappers may know what "ad hominem" is! The rest are playing "Doom" or watching TV.)
Oh, I don't know if playing doom is so bad, considering that the people who claim to know rhetoric and are proud to display it seem to spend all their time on usenet on debates like:
>>>>>>>>>>> [SNIP long winded argument]
>>>>>>>>>> Non sequitur.
>>>>>>>>> Ah, using fancy latin words instead of offering some counterargument. Could you answer the question?
>>>>>>>> Yes, I could.
>>>>>>> But you didn't. Seriosuly, get a life dude!
>>>>>> Ad hominem.
>>>>> No, that wasn't an ad hominem. This however, is: Oh, fuck off!
>>>> Ad hominem.
>>>Good, you catch on quickly! Now, could you answer my question please?
>>Yes, I could!
>Well, do it then!
Non sequitur.
etc. etc. etc.
AC writes:
With the shuttle, if the rocket explodes then everybody dies. With a design which places the crew capsule at the top, and an escape tower attached to it, then when the rocket explodes, everybody gets a really exciting ride and has to abort the mission, but they land safely. There is an enormous difference between low "mission success" reliability and low "the crew survives" reliability. The shuttle had the latter, this design should only have the former.
VoOk:
Add to that: One booster fails to fire on launch of the shuttle: The shuttle, in a wide arc, propelled with the remaining thruster, crashes and explodes. One booster fails to fire here: The astronauts use the ladder to get back down because the rocket doesn't move an inch.
And to add further, a capsule heatshield is pretty well protected during launch, as opposed to the shuttle. Also, if it's an ablative shield it can probably take quite a beating compared to the RCC shingles the shuttle uses.
The special thing about InfiniPath is that the adapter is not a PCI-(e|X) card but rather connects directly to the HyperTransport interface on the cpu (requiring a special MB with a "HTX" connector), giving slighlty lower latency than a normal IB adapter.
Well, unlike most cluster interconnects InfiniPath is special in that it connect directly to the HyperTransport interface on the Opterons, without going via PCI(-e|-X). So it won't work on e.g. a Xeon.
But yeah, from a user perspective Infinipath is just like any other cluster interconnect. Perhaps a bit lower latency, but nothing dramatic (as in order(s) of magnitude lower latency).
I've had people contribute code to PMD and say they were only contributing it because they felt the BSD license avoided any possible obligations on their part.
Just like there's plenty of people who only contribute to GPL projects since they don't want "evil corporations" stealing their code.
You can find fanatics driven by ideology rather than common sense in both camps. That's hardly something to cheer about.
I work in a lab at the University of Washington where we are working on _implementing_ a different data flow machine that shares some of fundamentals with the UT machine.
So, which one is better?
For a more serious question, I read the trips overview paper on their site and it all seems to make a lot of sense. So why aren't dataflow machines mainstream? The first papers were published in the early 1980s, not much later than risc started to make some noise.
I don't know if GNU malloc uses mmap() or brk() for its allocation, but in both cases small memory chunk that the user allocates are taken from bigger, contiguous blocks of memory.
It uses mmap() for big allocs (IIRC the threshold is 4 MB) and brk() for smaller ones.
There would be one solution, and that's using different arenas, or memory regions for allocation. For instance every window might have its own allocation region, so when you close the window/document, the memory BLOCK is freed.
Something like memory pools, used e.g. by apache portable runtime, gcc, and whatnot. Or the hierarchical allocator used by samba..
I really wish Java apps, Cocoa apps, and other (Mozilla) would do this, as they seem to suffer from this fragmentation problem, only increasing their memory usage, even after closing all documents/windows etc.
Actually, Java being garbage collected and not having "raw" pointers can avoid fragmentation by repacking heap memory. Of course the JVM must still use malloc() or mmap()/brk() directly to get the memory from the OS.
The upcoming GCC 4.1 release will include a stack protector. Basically it's a reimplementation of the old propolice patch.
Hopefully mainstream distros that have been wary of propolice will start using this new feature. And perhaps glibc malloc will borrow a few tricks from this new openbsd malloc too.
So, uh, what can this "OpenDocument" thing do that latex couldn't do 20 years ago? Why should I switch?
There is no solid definition of what differentiates a cluster
Exactly my point.
but it's a bit misleading and confusing to refer to an architecture that doesn't have any of the characteristics of a traditional cluster as a cluster.
But then, what are the traditional characteristics of a cluster? And why haven't those become the normal definition then?
Most clusters have a star or bus network topology of Infiniband, Myrinet, or Ethernet. An X1 uses a 3D torus of redundant, high-bandwidth links.
I don't think network topology is what defines a cluster. Consider the Scali Dolphin, which at least a few years ago was a relatively popular cluster interconnect. Depending on the model, it was a 1d, 2d or 3d torus. Or the SX-6/8, which uses a star topology with a crossbar switch.
Each node is a board in a backplane, not just a rackmount system.
Fair enough. So I agree the X1(E) is not a cluster then, perhaps "vector MPP" is more appropriate. But again, what about the SX-6/8? Each 8 CPU node is a separate cabinet, connected via the network switch. Starts to sound like a cluster to me.
While defining a cluster as a machine where each node has its own OS is a convienient definition, that also includes things which clearly aren't true clusters like networks of workstations and many MPPs.
Indeed. Given the lack of a solid definition, then perhaps one shouldn't go and label other people as idiots just because they use a slightly different definition of cluster?
This part of the article makes it plain that the author really doesn't know much about HPC systems. A vector machine uses vector processors. That's it. It says nothing about whether the machine is a cluster or MPP and what kinds of programming models it supports (distributed memory, shared memory, etc.).
The SX-6/8 and X1/X1E are not clusters.
Now that's a bit inconsistent. First you say that there's no agreed upon taxonomy of HPC systems, and then you blast the parent poster for not using YOUR definition of a cluster.
Another AFAIK common definition of a cluster is a set of independent computers (each with its own memory, OS image etc.) connected via some network. By that definition, almost every machine on the top500 is a cluster, including the NEC and (I think) Cray vectors.
They have enormous memory bandwidth (a necessity on vector machines) and are distributed-memory machines. UPC and Co-Array Fortran run very well on them because the address reach of any one processor covers the whole machine and maps extremely well onto the GAS programming model.
BTW, IIRC according to NEC documents the recommended programming model for the SX-6 and SX-8 is a combination of OpenMP and MPI. In that sense it's no different than a normal "cluster of SMP:s". Don't know about the X1, though Cray's been making noise about CAF, for good reason I think.
Nah, it's all about using the right tools for the job.
Clusters are a good thing, as they provide a very cost-effective platform for running codes with modest communication requirements. Just like running a communication-intensive code on a cluster will limit performance, running a code with little communication on a "real" supercomputer is a waste of money.
The sad thing about the current "HPC crisis" is not the rise of clusters, but the use of clusters for tasks which they are ill suited for (typically, "grand challenge" tasks).
The crucial thing is to have an appropriate balance, so that the tasks which really need a "real" supercomputer can get one, and other tasks can run on clusters.
Capture the imagination of the public, and thus keep the pork flowing to a few huge contractors.
Scientifically, there's little that a manned mission can do which an unmanned can't, just like you say.
Apollo was able to put a couple of guys on the moon and get them back. In contrast, the shuttle is incapable of going beyond LEO. It's not exactly breaking new ground in terms of space exploration.
Hey, don't be such a spoilsport.
After all, thanks to the science the manned space program has done, we now know a little about how ant colonies behave in zero gravity. Surely that will bring new advances in biology as well as help solve the problems the world faces today. Those $250 billion were well spent. Compared to, say, $10 billion for ITER, and other big science projects.
Oh, wait, that ant experiment crashed with Columbia. Guess we won't know how they behave in zero G then. What a loss to humanity!
Ahem: HOWTO Spot a Wannabe Web Standards Advocate.
Ladies and Gentlemen, we have a winner!
I have to say, I find it hard to imagine a market for a mainframe laptop. ;-)
I mean, the space shuttle was an experiment in how to screw it up. Expensive, unsafe, you name it.
When we have non-chemical propulsion and somewhat fancier materials it might be time to take another look at the reusable space plane approach. Until then, the basic multi-stage capsule on top design seems to be the best choice in terms of operational cost and safety.
I read that the Russians and the EU are collaborating on some kind of shuttle. Well, not a shuttle as in the space shuttle but more like one of those lifting body CEV designs.
;-)
Still, ironic that the US is ditching the shuttle and going back to capsules as the Russians are going the opposite way.
My generalization, huh? I wasn't the one saying that beowulf clusters run MPI.
;-)
Including this post, I have posted only two comments to this thread. And the other comment I posted was the comment saying PVM is dying. Well, whatever, it's not like it matters anyway.
They could be utilizing PVM... PVM: Parallel Virtual Machine
In principle yes. Practically speaking, PVM is dead, or at the very least dying fast. Except for running some legacy code, nobody is using it anymore.