Paper Critique 1
Mohammad Ali Farazdaghi
ELEC873
A Large-Scale Study of MPI Usage in Open-Source HPC Applications
Ignacio Laguna, Ryan Marshall, Kathryn Mohror, Martin Ruefenacht, Anthony Skjellum, Nawrin Sultana
Before We Start
Will not discuss:
- Most charts
- Time-Based Analysis (6)
- Complexity Analysis (7)
- Programming Model Analysis (8)
What they've done:
- Taken 100+ opensource projects using MPI
- Figured out MPI call usage statistics
- Labeled each program
There are 13 labels:
- Uses P2P Comms
- Uses Collective Comms
- Uses FileIO
- ...
How They've done it:
- Run
mpiusage.pyon each project - Performs static source code analysis
- Figures which MPI symbols are used
- All done using regexs
- Label projects accordingly
$ ./mpiusage.py /path/to/project/src/
{"MPI_COMM_SIZE": 3, "MPI_REDUCE": 3, "MPI_WAITALL": 2,
"MPI_WAIT": 116, "MPI_COMM_RANK": 15, "MPI_ALLREDUCE": 3,
"MPI_INIT": 3, "MPI_FINALIZE": 3, "MPI_ABORT": 12, "MPI_WTIME":
3, "MPI_IRECV": 26, "MPI_ISEND": 52, "MPI_BARRIER": 3,
"OPENMP": 0, "OPENACC": 0, "CUDA": 168, "OPENCL": 0,
"C_LINES": 0, "CPP_LINES": 297, "C_CPP_H_LINES": 365,
"FORTRAN_LINES": 0, "LINES_OF_CODE": 14574}
How did they figure out MPI Usage statistics?
Static Source Code Analysis!
Static Analysis?
Types of Program Analysis
-
Static AnalysisStatic Source Code Analysis
- Dynamic Analysis
Strengths
A Large Sample Size (100+)
→ Interesting Results
- C++ is the most popular
- 90% of all projects use P2P comms
- Most use MPI 1
Weaknesses
- Static sourcecode analysis may be too simplistic
- Behaviors that depend on runtime env
- Deadcode
- Compile time switches
- C++'s grammar is Turing Complete
- What's in
mpiusage.py?
Runtime Dependent behavior
in OpenFOAM
OpenFoam labels:
- P2P
- Collective
- Comm Group
void Foam::reduce(..., ...., ...., ...., ....) {
#ifdef MPIX_COMM_TYPE_SHARED
MPIX_Ireduce(...);
#else
reduce(..., ...., ...., ....); // -> allReduce(...)
#endif
}
void Foam::allReduce(...) {
if (small_enough()) {
// use p2p: MPI_Recv, MPI_Send
} else {
MPI_Allreduce(...); }
}
if (almost_always_returns_false()) {
MPI_Rare_API(...);
}
for (size_t i = 0; i < very_large_runtime_val; i++) {
MPI_Rare_API(...);
}
#ifdef USE_EXOTIC_HW_FEATURE_X
MPI_Rare_API(...);
#endif
Couldn't Find mpiusage.py anywhere
Cannot reproduce results
Not sure how they parsed C++
with regexes!