parallel computing and the os [3db6ebce]

talks, operating systems

https://dl.acm.org/doi/10.1145/2830903.2830911 www.youtube.com/embed/KZ9hm5AUmWU

• Minicomputers -> after timesharing

  • parallel computing sort of dropped off with this
  • logs and interrupts
  • multiprocess parallelism sort of died

• client/server computing make IO concurrcy big

  • must exploit I/O concurrency

  • microkernel organizations

    • User level services for I/O -> moves into microkernels, pthreads standard

• events vs threads

  • no sharing of handlers, no locks, etc, CSP style

    • most webservers today are events style
    • john usterhat

• multiprocessors make a return, NUMA stuff becomes big again

  • but uniprocessors performance keeps improving

• multicore processors come roaring back

  • cores per processor increase yet again performance goes back up
  • moore's law is dead on single cores, but multi cores are back

  • make, for example, increases and scale more

    • reading a cache line from a different core is expensive

    • congestion on cache lines, but avoiding cache line sharing is challenging

      • even a RW lock is expensive
      • basically the only real optimization is to stop spraying your data structures everywhere

    • RCU!

      • writers must follow careful discipline

• phase 4? OS community? Commmodity chips? -> Cloud computing, map reduce, moving data is expensive, move compute

  • shared memory locks have gotten better!
  • parallelism ahs moved up the software stack driven by commodity hardware