cs161 2004 Lecture 4 (non traditional, faster, I/O)

We've been talking about being careful about I/O.  Under what circumstances
will we still be CPU bound?

Look at code, find the inefficiencies.

int select(int n, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
           struct timeval *timeout);
 fixed size, copy all, scan all
int poll(struct pollfd *ufds, unsigned int nfds, int timeout);
 copy all, scan all (redo work for each fd for each call)

/dev/poll
sys_epoll (epoll_create, epoll_ctl, epoll_wait)

edge vs level
 If level, you must register and deregister more often
 If edge, you must track more on your own
   If you "miss an event" you're in trouble.  Likely?

Remaining problems?
 Unifying event sources (signals vs aio vs timers vs poll vs futexes (locks))
 Zero-copy
 Small frames or packets: TCP_CORK, ACK+data
 Behave under overload (Seda, livelock)



io-lite

background
 files: buffer cache / vm
 net: mbufs

new primitives:
 immutable buffers, page aligned
 mutable "buffer aggregates" - ordered list of <pointer,length> pairs
 what happens with a IOL_read, followed by IOL_write in another process?

optimizations
 get rid of copies (memory IO and cpu caching effects)
 get rid of redundant buffers (more room for memory caches)
 cross-layer (checksums)
 buffer control (Greedy Dual Size)
   Assign a value of cost/size to each cached resource.
   Add value of lowest resources to the value of an accessed resource

made faster:
 disk i/o? no (mmap)
 IPC? yes (no copy)  caveat: decide ahead of time who you will send to
 net? yes (copyless, and readless when repeated (chksum))
  WAN is particularly nice, memory savings

what wouldn't be helped?
  ssl server
  very small I/Os (VM ops cost more than small copies)
  compressed content? (if there are dynamic pieces, can't cache the rest)
  new bug? - unintentended sharing (mmu is page level)
    Actually, ACL must be predeclared
    So, rather than the bug, IOL forces a copy in that situation
    But one common case, CGIs is analogous to user/kernel.  Just trust.
    A case that isn't?  Databases
  stdio? Can pipe work transparently?  Seems to need ACL.  So what does?