JMK no matter what

준비한 것이 많은 것 같던데 발표하다 시간이 늦어져서 많이 스킵해서 아쉬웠다. 발표자료라도 봐야 할 듯? 재밌었음. aio 파이썬에서 쓰는 얘기 같은 거 좀 더 보고 싶은데.

• (Under Linux)

• (Usual) Buffered I/O

  • There are three buffers; one in user space, and two buffers (read/write) in Kernel
  • Lots of copying around. -.-

• Blocked I/Os

• Non-blocking I/O == styll synchronous!!

  • Still sequential
  • It's a property of the FD
  • returns EWOULDBLOCK or # of available bytes
  • when asked, kernel will get the data, but that's all
  • You have to poll the data out of kernel's buffer: read() again
  • and this is ugly

• I/O Multiplexing

  • poll()

• Event Driven I/O

  • The readiness of an FD for I/O is refered as an event
  • Libraries and frameworks
  • A (possibly) thin layer around I/O multiplexing?
  • Still everything is synchronous: once kernel's buffers fill up, there's no I/O happenning until you read() or write() it
  • So if the handler:
    • page faults
    • blocks on mutex
    • goes compute bound
    • => we're hosed
  • Why better? => far less memory, less context switching, do a lot of I/O without threads -- without GIL
  • libev, pyev
  • libevent
• Non-blocking Frameworks
  • Tornado, Twisted, Asyncore
• So we don't really have asynchronous I/O in Linux?
  • Windows has IOCP
  • aio_read aio_write in Linux: examples in Python plz orz
• differences in poll/epoll
  • poll: give list every time
  • epoll: we don't want to give a list of 10000 fds; register once, get events triggered
• eventlet?