"""# Fugaku Experiments Folder

📂 Folder Structure

CHIARA/
├── allgather/
│   ├── all_gather_radix_batch.cpp
│   └── main.cpp
├── allreduce/
│   ├── all_reduce_radix_batch.cpp
│   └── main.cpp
├── reduce-scatter/
│   ├── reduce_scatter_radix_batch.cpp
│   └── main.cpp
└── README.md

Each main.cpp follows a consistent structure and benchmarking flow:

1. Initializes MPI and parses runtime parameters.
2. Iterates over different message sizes (count = base << i).
3. Benchmarks radix- and batch-parameterized algorithm variants.
4. Records results to a CSV file (results<num_nodes>.csv).

---

⚙️ Compilation

All programs can be compiled using mpic++:

mpic++ -O3 -std=c++17 -o allgather main.cpp all_gather_radix_batch.cpp
mpic++ -O3 -std=c++17 -o allreduce main.cpp all_reduce_radix_batch.cpp
mpic++ -O3 -std=c++17 -o reduce_scatter main.cpp reduce_scatter_radix_batch.cpp

---

🚀 Usage

Each executable accepts the same set of parameters:

mpirun -np <nprocs> ./program <n_iter> [--overwrite] [b=<value>] [base=<value>] [num_nodes=<value>] [radix_increment=<value>]

Arguments

Parameter	Description	Default
<n_iter>	Number of iterations (required)	—
--overwrite	Overwrites existing CSV file	false
b=<value>	Batch size parameter (upper limit for radix loop)	16
base=<value>	Base message size multiplier	1
num_nodes=<value>	Number of nodes in the system (used for CSV naming)	1
radix_increment=<value>	Step size for increasing radix r	1

Example:

mpirun -np 16 ./reduce_scatter 8 --overwrite b=32 base=4 num_nodes=4 radix_increment=2

---

🧠 Output

Each run produces a CSV file named:

results<num_nodes>.csv

The CSV contains one row per experiment configuration:

algorithm_name	k	b	nprocs	send_count	time	is_correct
reduce_scatter_radix_batch	4	16	32	4096	1.23e-05	1
reduce_scatter_standard	—	—	32	4096	2.45e-05	1

• time is the measured execution latency in seconds.
• is_correct indicates whether the result was validated against the MPI reference collective.

---

🧩 Implementation Notes

• Each collective (AllGather, AllReduce, ReduceScatter) defines:
  • A parameterized radix/batch variant (*_radix_batch.cpp)
  • A standard MPI baseline for comparison.
• The test harness in main.cpp automatically sweeps message sizes and radix values.
• Designed for scaling studies on HPC clusters (e.g., Fugaku, Polaris, Aurora).

---

📜 Example Run

# Example: Run AllReduce experiment on 8 processes
mpirun -np 8 ./allreduce 6 --overwrite b=16 base=2 num_nodes=8 radix_increment=2

Output:

Parsed parameters:
  n_iter          = 6 
  overwrite       = true
  b               = 16
  base            = 2
  num_nodes       = 8
  radix_increment = 2

The resulting file results8.csv will contain performance data for all tested configurations.

"""