Files
TDT4200/exercise3/report.md
2025-10-07 15:18:14 +02:00

2.5 KiB

title, date, author
title date author
tdt4200ex3 -- 1d wave equation mpi 2025-09-22 fredrik robertsen

theory questions

1)

by parallelizing the code i get quite a speed-up when running four processes. this is as i expected, because my laptop has four cores available for parallel processing. i have done threading previously, which also gave a speed-up for four threads. thus it makes sense that we get a speed-up here. it means we must have more intensive calculations than the overhead of using mpi.

here's an excerpt from make check:

./sequential
Total elapsed time: 5.115875 seconds

mpiexec -n 1 --oversubscribe ./parallel
Total elapsed time: 5.347622 seconds

mpiexec -n 4 --oversubscribe ./parallel
Total elapsed time: 1.729224 seconds

mpiexec -n 13 --oversubscribe ./parallel
Total elapsed time: 2.717839 seconds

assuming i have implemented timing correctly, mpi is slightly slower with only a single process, as expected. similarly, for 13 processes we get a speed-up, but not by as much due to the overhead of having more processes that are not necessarily capable of running in parallel -- they are limited by my hardware.

2)

this is a ring-buffer (halo) topology 3-stencil. a process only communicates with its nearest neighbor. the calculations are a form of finite differences.

3)

well-spoken AI answer:

point-to-point communication:

  • communication between exactly two processes (sender ↔ receiver)
  • examples: mpi_send, mpi_recv, mpi_sendrecv

collective communication:

  • communication involving all processes in a communicator
  • examples: mpi_bcast, mpi_reduce, mpi_allreduce, mpi_gather, mpi_scatter

key differences:

aspect point-to-point collective
participants 2 processes all processes
synchronization optional required
optimization manual automatic
flexibility high limited patterns

4)

int* a, b is a classic c pitfall. my auto-formatter is configured to correctly format such a string to int *a, b, since pointers are a part of the variable name definition, not the type. as such, b is merely an int.

this is slightly relevant for the entertaining c-ism buffer[0] == 0[buffer] (this relates to the fact that indexing brackets are syntactic sugar for pointer addition, which is commutative).