4 ms·

"Supercomputers, the fastest of the fast, run Linux almost exclusively." And it's so disappointing. Consider the Blue Gene series. Great new architecture, cust

by jff 15y ago

"Supercomputers, the fastest of the fast, run Linux almost exclusively." And it's so disappointing. Consider the Blue Gene series. Great new architecture, custom networks, custom hardware. Initially, they put Linux on the I/O nodes and a custom kernel on the compute nodes, because you want a minimum of OS noise (see the FTQ benchmark for more info on this). But everyone clamored for Linux on compute nodes--so now we boot this enormous Linux kernel on each compute node. Plan 9 was ported (I did some of it), but nobody wanted it because it wasn't Linux. I'm of the opinion that no group in the world is as set-in-their-ways and reactionary as the supercomputing application developers. If it can't run Linux and Fortran 77, they don't want it.

mebesilly 15y ago
The only real competitors to Fortran on supercomputers are Chapel http://chapel.cray.com/ http://chapel.cray.com/ and X10 http://x10-lang.org/ http://x10-lang.org/ ( the proposed language Fortress died an early death). Toy languages like C, C++ and Java that microcomputer developers play with simply are not suitable for supercomputers. Among other problems, those joke languages simply do not scale to 64,000 heterogenous processors.
rbanffy 15y ago
Plan 9 was a very interesting OS, but it always gave me the impression to be too desktop/workstation influenced. While you could have your heavy lifting carried out on big processing nodes on the net, it depended on having the same one-size-fits-all OS on all boxes. There was a time when we envisioned powerful computers running powerful software on every desktop. That time has more or less passed - most people now are very happy with the browser, the modern equivalent of the 3270. And I have to tell you that being able to sit down at any 3270 in any company building and to have access to your stuff readily available was, in several ways, a better (if uglier) experience than what desktop workstation provided. Plan 9 also committed the cardinal sin of being different for the sake of being different. You can build a better Unix without making "vi" a MIPS emulator. We invent things that are ahead of their time. It's a shame so much of the technology is lost when that happens. As for the Fortran 77 thing... Well... They have lots of code that has already been written and tested. Do you really want to port it to something else? Besides, legend says F77 compiles to very efficient code.
- tedunangst 15y ago
  I don't get that impression for plan 9 at all. It included lots of features that suggest it was designed for heterogeneous clusters. Like simple convenient cross compiler support. You could run mips, your office mate could run sparc, and you could effectively and somewhat transparently use both to get your work done. I think plan 9 was actually trying to transform the network to be like your 3270 example.
  - rbanffy 15y ago
    > It included lots of features that suggest it was designed for heterogeneous clusters. ... all running Plan 9. I see your point - when it was conceived we had RISC workstations and Pentium PCs and they all could run some flavor of Plan 9, but today we have x86 PCs and ARM gadgets running browsers and x86 boxes running app servers and databases buried in data-centers. It could be built on Plan 9, but that's not how it ended up being.
- jff 15y ago
  Possibly the biggest motivation for porting Plan 9 was that it provides a very light-weight yet featureful kernel. Linux is way too big, but the compute node kernels IBM and Cray have been coming up with may be too simple. Plan 9 provides a rich environment but doesn't demand much of the environment. And yes, Fortran 77 can be quite efficient... but on the other hand, we're also seeing a lot of new software being written in Python, which (at least according to our process traces) tries to open thousands of non-existent files on startup. So efficiency isn't everything they care about... it's partly that they were used to Fortran, and now they're used to Python, and something scarily new like Plan 9 is, well, scary. Oh, and good luck trying to convince people that MPI isn't necessarily the only solution for parallel programming. If the community at large has its way, we'll still be using Linux and MPI in 2040.
  - rbanffy 15y ago
    > and something scarily new like Plan 9 is, well, scary Plan 9 would be more OK if it didn't try to be so not Unix. That's too bad. It saddens me that, of the three more advanced widely used OS, two are variations of the 70's Unix theme while the third is the bastard child of VMS with lots of lipstick on it. :-( Plan 9 deserved better.

by jff 15y ago

mebesilly 15y ago

The only real competitors to Fortran on supercomputers are Chapel http://chapel.cray.com/ http://chapel.cray.com/ and X10 http://x10-lang.org/ http://x10-lang.org/ ( the proposed language Fortress died an early death). Toy languages like C, C++ and Java that microcomputer developers play with simply are not suitable for supercomputers. Among other problems, those joke languages simply do not scale to 64,000 heterogenous processors.

rbanffy 15y ago

Plan 9 was a very interesting OS, but it always gave me the impression to be too desktop/workstation influenced. While you could have your heavy lifting carried out on big processing nodes on the net, it depended on having the same one-size-fits-all OS on all boxes. There was a time when we envisioned powerful computers running powerful software on every desktop. That time has more or less passed - most people now are very happy with the browser, the modern equivalent of the 3270. And I have to tell you that being able to sit down at any 3270 in any company building and to have access to your stuff readily available was, in several ways, a better (if uglier) experience than what desktop workstation provided. Plan 9 also committed the cardinal sin of being different for the sake of being different. You can build a better Unix without making "vi" a MIPS emulator. We invent things that are ahead of their time. It's a shame so much of the technology is lost when that happens. As for the Fortran 77 thing... Well... They have lots of code that has already been written and tested. Do you really want to port it to something else? Besides, legend says F77 compiles to very efficient code.

tedunangst 15y ago
I don't get that impression for plan 9 at all. It included lots of features that suggest it was designed for heterogeneous clusters. Like simple convenient cross compiler support. You could run mips, your office mate could run sparc, and you could effectively and somewhat transparently use both to get your work done. I think plan 9 was actually trying to transform the network to be like your 3270 example.
- rbanffy 15y ago
  > It included lots of features that suggest it was designed for heterogeneous clusters. ... all running Plan 9. I see your point - when it was conceived we had RISC workstations and Pentium PCs and they all could run some flavor of Plan 9, but today we have x86 PCs and ARM gadgets running browsers and x86 boxes running app servers and databases buried in data-centers. It could be built on Plan 9, but that's not how it ended up being.
jff 15y ago
Possibly the biggest motivation for porting Plan 9 was that it provides a very light-weight yet featureful kernel. Linux is way too big, but the compute node kernels IBM and Cray have been coming up with may be too simple. Plan 9 provides a rich environment but doesn't demand much of the environment. And yes, Fortran 77 can be quite efficient... but on the other hand, we're also seeing a lot of new software being written in Python, which (at least according to our process traces) tries to open thousands of non-existent files on startup. So efficiency isn't everything they care about... it's partly that they were used to Fortran, and now they're used to Python, and something scarily new like Plan 9 is, well, scary. Oh, and good luck trying to convince people that MPI isn't necessarily the only solution for parallel programming. If the community at large has its way, we'll still be using Linux and MPI in 2040.
- rbanffy 15y ago
  > and something scarily new like Plan 9 is, well, scary Plan 9 would be more OK if it didn't try to be so not Unix. That's too bad. It saddens me that, of the three more advanced widely used OS, two are variations of the 70's Unix theme while the third is the bastard child of VMS with lots of lipstick on it. :-( Plan 9 deserved better.