Lecture 2

Measuring Speed

• Most CPU clocks can't be more accurate than 10ms

• At 3GHz, 10ms = 30,000,000 cycles

  • ~1$\rightarrow$5 cycles per instruction
  • $\therefore$ 10ms $\approx$ 10,000,000 instructions

• Code that takes 15ms is reported as having taken 10 or 20 ms

  • Large inaccuracy
  • 10ms $\pm$ 10ms

• Instead, run it 100 times, 1500 ms reported as either 1490, 1500 or 1510

  • lower relative inaccuracy

• Issue: compilers are too smart, and will attempt to optimise re-runs, therefore, subsequent runs may be much faster than initial

  • This can be mostly negated by flushing caches between each run
  • Alternatively, ignore initial run.

• Do not include I/O in your timings

  • remove delay statements / logging

Posix Threads

• language independent parallel execution model (Currently only implemented in C)
  • Thread management
  • Mutexes
    • Mutually exclusive lock on a variable
  • Condition variables
    • lock until condition becomes True
  • Read/ Write locks
    • Allows multiple readers or a single writer
  • Barriers
    • Threads whicb hit the barrier (code location) have to wait until all threads in the group also reach that barrier.
    • Allows thread groups to syncronise

Basics of pthreads

pthread_attr_init()

Gets default thread attributes & stores ina pthread_attr_t struct

pthread_attr_setdetachstate()

Sets the detach state value.

• Joinable
• Detachable

pthread_create()

• Creates a new thread with attributes defined in a pthread_attr_t struct
• Starts running on the function specified with the specified argument & stores thread id in a specific variable

pthread_exit()

• Terminates the current thread
  • if joinable:
    • its return value is available to any other thread that calls pthread_join() on this thread's id

pthread_join()

• Waits for the specific joinable thread to terminate & gets return value from it.

Note: Making threads joinable requires more memory $\therefore$if not required, make threads detached