Skip to content

Reduce timing overhead in examples 20 and 21 by measuring actual work #265

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 3 commits into
base: main
Choose a base branch
from
Draft

Reduce timing overhead in examples 20 and 21 by measuring actual work #265

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
89419e3
Initial plan
Copilot Oct 19, 2025
2427d5a
Fix timing discrepancy when using --gemm_only or --comm_only flags
Copilot Oct 19, 2025
80869af
Reduce timing overhead by measuring actual work instead of do_bench w…
Copilot Oct 19, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 22 additions & 2 deletions examples/20_gemm_all_scatter_independent/benchmark.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -220,6 +220,12 @@ def _worker(local_rank: int, world_size: int, init_url: str, args: dict):
"ms": 0,
"experiments": 0,
},
"total": {
"start_event": torch.cuda.Event(enable_timing=True),
"end_event": torch.cuda.Event(enable_timing=True),
"ms": 0,
"experiments": 0,
},
}

# Allocate Timestamps
Expand All @@ -237,6 +243,9 @@ def run_experiment():
timestamps.reset()
shmem.barrier()

# Record start of actual work (after initial barrier)
kernel_timing["total"]["start_event"].record()

# Determine what to run based on flags
run_gemm = not args["only_comm"]
run_comm = not args["only_gemm"]
Expand Down Expand Up @@ -303,6 +312,10 @@ def run_experiment():
kernel_timing["communication"]["experiments"] += 1
torch.cuda.nvtx.range_pop()

# Record end of actual work (before final barrier)
kernel_timing["total"]["end_event"].record()
kernel_timing["total"]["experiments"] += 1

shmem.barrier()

# Update timing for operations that were run
Expand All @@ -312,6 +325,10 @@ def run_experiment():
if run_comm:
ms = kernel_timing["communication"]["start_event"].elapsed_time(kernel_timing["communication"]["end_event"])
kernel_timing["communication"]["ms"] += ms

# Update total timing
ms = kernel_timing["total"]["start_event"].elapsed_time(kernel_timing["total"]["end_event"])
kernel_timing["total"]["ms"] += ms

torch.cuda.nvtx.range_pop()

Expand All @@ -323,7 +340,7 @@ def run_experiment():

shmem.barrier()

for k in ["gemm", "communication"]:
for k in ["gemm", "communication", "total"]:
kernel_timing[k]["ms"] = 0
kernel_timing[k]["experiments"] = 0

Expand Down Expand Up @@ -387,7 +404,10 @@ def run_experiment():
matmul.set_debug(False)
shmem.info("Benchmarking...")
perf = lambda ms: 2 * args["m"] * args["n"] * args["k"] * 1e-12 / (ms * 1e-3)
triton_ms = iris.do_bench(run_experiment, shmem.barrier)
# Run benchmark to populate all timing measurements
iris.do_bench(run_experiment, shmem.barrier)
# Use the total timing recorded inside run_experiment (excludes do_bench overhead)
triton_ms = kernel_timing["total"]["ms"] / kernel_timing["total"]["experiments"]
triton_tflops = perf(triton_ms)
algo_string = "all_scatter"

Expand Down
24 changes: 22 additions & 2 deletions examples/21_gemm_one_shot_all_reduce_independent/benchmark.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -211,6 +211,12 @@ def _worker(local_rank: int, world_size: int, init_url: str, args: dict):
"ms": 0,
"experiments": 0,
},
"total": {
"start_event": torch.cuda.Event(enable_timing=True),
"end_event": torch.cuda.Event(enable_timing=True),
"ms": 0,
"experiments": 0,
},
}

# Allocate Timestamps
Expand All @@ -227,6 +233,9 @@ def run_experiment():
timestamps.reset()
shmem.barrier()

# Record start of actual work (after initial barrier)
kernel_timing["total"]["start_event"].record()

# Determine what to run based on flags
run_gemm = not args["only_comm"]
run_comm = not args["only_gemm"]
Expand Down Expand Up @@ -297,6 +306,10 @@ def run_experiment():
kernel_timing["communication"]["experiments"] += 1
torch.cuda.nvtx.range_pop()

# Record end of actual work (before final barrier)
kernel_timing["total"]["end_event"].record()
kernel_timing["total"]["experiments"] += 1

shmem.barrier()

# Update timing for operations that were run
Expand All @@ -306,6 +319,10 @@ def run_experiment():
if run_comm:
ms = kernel_timing["communication"]["start_event"].elapsed_time(kernel_timing["communication"]["end_event"])
kernel_timing["communication"]["ms"] += ms

# Update total timing
ms = kernel_timing["total"]["start_event"].elapsed_time(kernel_timing["total"]["end_event"])
kernel_timing["total"]["ms"] += ms

torch.cuda.nvtx.range_pop()

Expand All @@ -317,7 +334,7 @@ def run_experiment():

shmem.barrier()

for k in ["gemm", "communication"]:
for k in ["gemm", "communication", "total"]:
kernel_timing[k]["ms"] = 0
kernel_timing[k]["experiments"] = 0

Expand Down Expand Up @@ -381,7 +398,10 @@ def run_experiment():
matmul.set_debug(False)
shmem.info("Benchmarking...")
perf = lambda ms: 2 * args["m"] * args["n"] * args["k"] * 1e-12 / (ms * 1e-3)
triton_ms = iris.do_bench(run_experiment, shmem.barrier)
# Run benchmark to populate all timing measurements
iris.do_bench(run_experiment, shmem.barrier)
# Use the total timing recorded inside run_experiment (excludes do_bench overhead)
triton_ms = kernel_timing["total"]["ms"] / kernel_timing["total"]["experiments"]
triton_tflops = perf(triton_ms)

# Determine what was run based on flags
Expand Down