B7 - scheduling

threads get roughly equal numbers of loop iterations
some iterations may be more time-consuming

chunks

contiguous ranges of iterations, the number in a given range being the chunk size

$N, P$ and $C$ are the trip-count, number of threads, and the chunk size (if specified)

static scheduling

allocates chunks before loop execution
by default: schedule(static), with $P$ chunks of size $C = N / P$
with schedule(static, C), iterations are divided into $N / C$ chunks of size $C$ allocated on a round-robin basis

dynamic scheduling

allocates chunks during loop execution
the aim is that the fastest processor waits at most $C$ iterations at the end of loop
but the overhead goes up as $C$ decreases
schedule(dynamic, C) divides iterations into $N / C$ chunks of size $C$
when a thread finished a chunk, it gets a new one, helping balance the workload
the disadvantage is scheduling overhead as OpenMP must coordinate the allocation of chunks during execution
by default, C=1

guided

schedule(guided, C) works similarly to dynamic scheduling but the chunk size generally decreases as fewer iterations remain up to a specified minimum chunk size
roughly:

chunk size \approx remaining iterations / number of threads

large initial chunks means less scheduling overhead
small chunks in the end means more flexible distribution of remaining work
by default, C=1

runtime

allows schedule to be set at runtime rather than compilation
this is specified with environment variable or function call
eg: export OMP_SCHEDULE = "static, 4" or omp_set_schedule(omp_sched_static, 4)
useful for experimenting:

export OMP_NUM_THREADS=4

export OMP_SCHEDULE="static,2"
./scheduling

export OMP_SCHEDULE="dynamic,2"
./scheduling

export OMP_SCHEDULE="guided,2"
./scheduling