In this assignment you'll parallelize your solver with std::jthreads. You can use any threading paradigm you'd like, but the number of threads spawned must be constant regardless of simulation time--you may not spawn threads on each iteration. Functionality and performance requirements are the same as they were for the first optimization assignment--given 8 threads on a whole m9 node, wavesolve_thread must run in 20 seconds on 2d-medium-in.wo, and checkpointing must still work.
The environment variable SOLVER_NUM_THREADS will control how many threads are used. If it is not set or isn't a positive integer, one thread should be used.
The first choice you'll have to make is how to split work among threads. The example code divides work into evenly-sized chunks and assigns each thread one such chunk. Some people find it easier to split work into chunks, put these chunks on a shared queue, and have a pool of worker threads that pull chunks from said queue on each iteration. Choose whichever method makes the most sense to you--performance will be comparable assuming sane work splitting.
Update your CMakeLists.txt to create wavesolve_thread (making sure to compile with C++ threads), develop in a branch named phase6 or tag the commit you'd like me to grade from phase6, and push it.
This phase is worth 20 points. The following deductions, up to 20 points total, will apply for a program that doesn't work as laid out by the spec:
| Defect | Deduction |
|---|---|
Failure to compile wavesolve_thread |
4 points |
Failure of wavesolve_thread to work on each of 3 test files |
4 points each |
Failure of wavesolve_thread to checkpoint correctly |
4 points |
Failure of wavesolve_thread to run on 2d-medium-in.wo on m9 with 8 threads in 20 seconds |
4 points |
| ...in 60 seconds | 4 points |
wavesolve_thread isn't a threaded program, or doesn't distribute work evenly between threads |
1-20 points |
wavesolve_thread spawns more than a small, constant number of threads |
10 points |