3 ms·

> There are a lot of reasons to do big science projects other than accomplishing the science itself One is national prestige. I'm not sure that's a good reason

> There are a lot of reasons to do big science projects other than accomplishing the science itself One is national prestige. I'm not sure that's a good reason (if that's what's happened here).

godelski 7y ago
I think it definitely does help. National pride is definitely a factor that matters. You can say that your country is on the forefront of science. Personally I'm very happy to spend a lot of money on science. It helps push humanity forward and is good for the economy (ROI is high in the long run, though maybe not in the short term. Which is why government funding is better than commercial)
- tempguy9999 7y ago
  At the risk of misrepresenting chinese efforts, my point was a bigger scientific instrument might just be willy waving. Blowing resources doesn't necessarily get results. If someone claims a new supercomputer place, the first thing to ask is whether they're getting real performance out of it. (I'll repeat that this a question, and not an assertion that china's making a mistake)
  - godelski 7y ago
    Oh I think I misunderstood what you were saying. > Blowing resources doesn't necessarily get results. > If someone claims a new supercomputer place, the first thing to ask is whether they're getting real performance out of it. So like TaihuLight[0] and Tianhe[1]? ([2]) I still think these machines add to national pride, but you're right that the money would be much better spent if it ALSO did actual science. I don't think these kinds of moves are as beneficial, because the benefits that I described in my first post are well known in the community (though not sure if they are well known by the public or even Rick Perry (though I know he's been told)). [0] https://top500.org/system/178764 https://top500.org/system/178764 [1] https://top500.org/system/177999 https://top500.org/system/177999 [2] I work on super computers and from my understanding (I haven't personally used Taihu or Tianhe) is that these computers are "Linpack machines" (linpack is the metric used to quantify the rankings). That these computers are extremely difficult to work with (all super computers are difficult to work with, and so these are HPC guys telling me that they don't want to work with those machines and are much happier to work on machines like Summit, Sierra, or even Titan) and so not much science is actually done on them.
    - tempguy9999 7y ago
      Whoa, you work in HPC? Wow. I understood that the TaihuLight etc. were based on SW26010 cpus and as these use scratchpad rather than automatic cache hierarchies I assumed they were a bugger to program therefore were less efficient (it's not that simple but you know what I mean). But I checked on https://en.wikipedia.org/wiki/TOP500#TOP_500 https://en.wikipedia.org/wiki/TOP500#TOP_500 which gave, for this machine, Rpeak (theoretical peak) of 125 PFlops[0] and Rmax ("the highest score measured using the LINPACK benchmarks suite") as 93 PFlops. Which really wasn't at all bad compared to the US machines, so I wondered where that left my assumptions. But you now mention they're hard to use. I'd assumed that if linpack was written and tested, you just used it and got the peformance. Apparently not? (I don't work with supercomputers, just in case you hadn't noticed)
      - godelski 7y ago
        So a big part of what's hard in HPC (really depends on what you do in the area. I mainly just use the machines to run stuff. You don't write software on the machines) is getting things to install (and run as expected). The people that maintain the machines will generally have your basic stuff installed (things like GCC, clang, your intel compilers, python, hdf5, etc) and you can load different versions using modules. Anything not there you need to do a local build of (your software and anything that your software depends on that isn't a super common library). Building these packages can be a real pain because architecture is so different (some people work on things called Data Parallel Primitives that try to make these easier. One common software that uses this is VTKm. They us a lot of template metaprogramming, which results in crazy errors that are hard to read but many times works more easily). Basically take the frustration you have for building things from source and multiply that by at least 2. So basically what I've heard about Taihu is that installing any packages and writing code for the system is even more of a pain (like adds months to your work (probably an exaggeration)) to write for. Libraries that are commonly used aren't supported and so many things need to be written by hand (including all the optimization). So they go (from my knowledge) fairly unused. Hence calling them Linpack machines (someone got linpack running and submitted test results. Doesn't say how utilized the machine is). But this could also be people in the labs playing up lab pride too. So take everything with a grain of salt. I want to try to be fair here. (US labs do have lots of internationals at them, not just western countries, plenty of Chinese nationals. They just want smart people). As for performance, it depends on a lot of things (it is actually REALLY hard to measure). Linpack is just a good baseline and has been highly optimized. But it doesn't tell you how fast your software is going to be because frankly your software could be completely unoptimized. It is highly dependent. That said, having the fastest/biggest HPC computer doesn't mean too much. No one is running across the entire machine at one time. It is generally a lot of different programs and groups working on the same machine. There are benefits though. It is easier to have a single big machine than many smaller machines and it does enable you to do the science you want to do if you do end up needing a huge capacity. More bang for your buck. And for only a few hundred million dollars too, not a bad price.

> There are a lot of reasons to do big science projects other than accomplishing the science itself One is national prestige. I'm not sure that's a good reason (if that's what's happened here).

godelski 7y ago

I think it definitely does help. National pride is definitely a factor that matters. You can say that your country is on the forefront of science. Personally I'm very happy to spend a lot of money on science. It helps push humanity forward and is good for the economy (ROI is high in the long run, though maybe not in the short term. Which is why government funding is better than commercial)

tempguy9999 7y ago
At the risk of misrepresenting chinese efforts, my point was a bigger scientific instrument might just be willy waving. Blowing resources doesn't necessarily get results. If someone claims a new supercomputer place, the first thing to ask is whether they're getting real performance out of it. (I'll repeat that this a question, and not an assertion that china's making a mistake)
- godelski 7y ago
  Oh I think I misunderstood what you were saying. > Blowing resources doesn't necessarily get results. > If someone claims a new supercomputer place, the first thing to ask is whether they're getting real performance out of it. So like TaihuLight[0] and Tianhe[1]? ([2]) I still think these machines add to national pride, but you're right that the money would be much better spent if it ALSO did actual science. I don't think these kinds of moves are as beneficial, because the benefits that I described in my first post are well known in the community (though not sure if they are well known by the public or even Rick Perry (though I know he's been told)). [0] https://top500.org/system/178764 https://top500.org/system/178764 [1] https://top500.org/system/177999 https://top500.org/system/177999 [2] I work on super computers and from my understanding (I haven't personally used Taihu or Tianhe) is that these computers are "Linpack machines" (linpack is the metric used to quantify the rankings). That these computers are extremely difficult to work with (all super computers are difficult to work with, and so these are HPC guys telling me that they don't want to work with those machines and are much happier to work on machines like Summit, Sierra, or even Titan) and so not much science is actually done on them.
  - tempguy9999 7y ago
    Whoa, you work in HPC? Wow. I understood that the TaihuLight etc. were based on SW26010 cpus and as these use scratchpad rather than automatic cache hierarchies I assumed they were a bugger to program therefore were less efficient (it's not that simple but you know what I mean). But I checked on https://en.wikipedia.org/wiki/TOP500#TOP_500 https://en.wikipedia.org/wiki/TOP500#TOP_500 which gave, for this machine, Rpeak (theoretical peak) of 125 PFlops[0] and Rmax ("the highest score measured using the LINPACK benchmarks suite") as 93 PFlops. Which really wasn't at all bad compared to the US machines, so I wondered where that left my assumptions. But you now mention they're hard to use. I'd assumed that if linpack was written and tested, you just used it and got the peformance. Apparently not? (I don't work with supercomputers, just in case you hadn't noticed)
    - godelski 7y ago
      So a big part of what's hard in HPC (really depends on what you do in the area. I mainly just use the machines to run stuff. You don't write software on the machines) is getting things to install (and run as expected). The people that maintain the machines will generally have your basic stuff installed (things like GCC, clang, your intel compilers, python, hdf5, etc) and you can load different versions using modules. Anything not there you need to do a local build of (your software and anything that your software depends on that isn't a super common library). Building these packages can be a real pain because architecture is so different (some people work on things called Data Parallel Primitives that try to make these easier. One common software that uses this is VTKm. They us a lot of template metaprogramming, which results in crazy errors that are hard to read but many times works more easily). Basically take the frustration you have for building things from source and multiply that by at least 2. So basically what I've heard about Taihu is that installing any packages and writing code for the system is even more of a pain (like adds months to your work (probably an exaggeration)) to write for. Libraries that are commonly used aren't supported and so many things need to be written by hand (including all the optimization). So they go (from my knowledge) fairly unused. Hence calling them Linpack machines (someone got linpack running and submitted test results. Doesn't say how utilized the machine is). But this could also be people in the labs playing up lab pride too. So take everything with a grain of salt. I want to try to be fair here. (US labs do have lots of internationals at them, not just western countries, plenty of Chinese nationals. They just want smart people). As for performance, it depends on a lot of things (it is actually REALLY hard to measure). Linpack is just a good baseline and has been highly optimized. But it doesn't tell you how fast your software is going to be because frankly your software could be completely unoptimized. It is highly dependent. That said, having the fastest/biggest HPC computer doesn't mean too much. No one is running across the entire machine at one time. It is generally a lot of different programs and groups working on the same machine. There are benefits though. It is easier to have a single big machine than many smaller machines and it does enable you to do the science you want to do if you do end up needing a huge capacity. More bang for your buck. And for only a few hundred million dollars too, not a bad price.