mirror of
https://github.com/harvard-edge/cs249r_book.git
synced 2026-05-07 18:18:42 -05:00
1eb30f5f86
Align the MLSys·im code, docs, paper, website, workflows, and lab wheel for the 0.1.1 release. This also fixes runtime/API issues found during release review and prepares the paper PDF plus archive package.
81 lines
3.0 KiB
Python
81 lines
3.0 KiB
Python
"""
|
|
Example 04: The Data Wall
|
|
-------------------------
|
|
This script demonstrates the "Data Wall" concept from Volume 2, Lab 4.
|
|
It shows how faster GPUs can actually result in lower utilization if the
|
|
storage bandwidth cannot keep up with the compute demand.
|
|
"""
|
|
import mlsysim
|
|
from mlsysim.core.constants import Q_
|
|
|
|
def main():
|
|
print("Evaluating ResNet-50 Training Data Pipeline...\n")
|
|
|
|
# 1. Define the Workload
|
|
model = mlsysim.Models.Vision.ResNet50
|
|
batch_size = 256
|
|
|
|
# Calculate data demand per step
|
|
# ResNet50 input: 224x224x3 FP16 (2 bytes)
|
|
bytes_per_sample = 224 * 224 * 3 * 2
|
|
batch_bytes = bytes_per_sample * batch_size * Q_("1 byte")
|
|
|
|
# 2. Define the Hardware Targets
|
|
v100 = mlsysim.Hardware.Cloud.V100
|
|
a100 = mlsysim.Hardware.Cloud.A100
|
|
h100 = mlsysim.Hardware.Cloud.H100
|
|
|
|
# 3. Define the Storage (A standard NVMe SSD)
|
|
storage_bw = Q_("3.0 GB/s")
|
|
|
|
print(f"Storage Bandwidth: {storage_bw}")
|
|
print(f"Batch Data Size: {batch_bytes.to('MB'):.2f}\n")
|
|
|
|
print(f"{'GPU':<10} | {'Compute Time':<15} | {'I/O Time':<15} | {'GPU Utilization':<15}")
|
|
print("-" * 65)
|
|
|
|
for hw in [v100, a100, h100]:
|
|
# Calculate Compute Time
|
|
prof = mlsysim.Engine.solve(
|
|
model=model,
|
|
hardware=hw,
|
|
batch_size=batch_size,
|
|
precision="fp16",
|
|
is_training=True
|
|
)
|
|
t_compute = prof.latency
|
|
|
|
# Calculate I/O Time
|
|
t_io = (batch_bytes / storage_bw).to("ms")
|
|
|
|
# Calculate Utilization (assuming no perfect overlap for simplicity, or just taking the ratio)
|
|
# If I/O takes longer than compute, GPU is idle waiting for data.
|
|
# Utilization = T_compute / (T_compute + T_io) (if purely sequential)
|
|
# Or if pipelined: T_compute / max(T_compute, T_io)
|
|
utilization = t_compute / max(t_compute, t_io)
|
|
|
|
print(f"{hw.name:<10} | {t_compute.m_as('ms'):.1f} ms | {t_io.m_as('ms'):.1f} ms | {utilization.m_as('dimensionless') * 100:.1f}%")
|
|
|
|
print("\nConclusion: As GPUs get faster (V100 -> H100), the compute time drops.")
|
|
print("But because storage bandwidth is fixed, the GPU spends more time waiting for data,")
|
|
print("causing utilization to plummet. This is the Data Wall.")
|
|
|
|
if __name__ == "__main__":
|
|
main()
|
|
|
|
# Expected output (mlsysim v0.1.1):
|
|
# Evaluating ResNet-50 Training Data Pipeline...
|
|
#
|
|
# Storage Bandwidth: 3.0 GB/s
|
|
# Batch Data Size: 77.07 MB
|
|
#
|
|
# GPU | Compute Time | I/O Time | GPU Utilization
|
|
# -----------------------------------------------------------------
|
|
# NVIDIA V100 | 51.9 ms | 25.7 ms | 100.0%
|
|
# NVIDIA A100 | 41.9 ms | 25.7 ms | 100.0%
|
|
# NVIDIA H100 | 7.9 ms | 25.7 ms | 30.7%
|
|
#
|
|
# Conclusion: As GPUs get faster (V100 -> H100), the compute time drops.
|
|
# But because storage bandwidth is fixed, the GPU spends more time waiting for data,
|
|
# causing utilization to plummet. This is the Data Wall.
|