Memory Bandwidth Test

Benchmark application for measuring memory read/write bandwidth using PXL runtime and MU kernels.

Overview

The memory_test example measures memory subsystem performance by executing parallel read/write operations across multiple tasks. Located in example/memory_test directory.

Use cases:

  • Evaluate memory bandwidth performance
  • Compare different memory access patterns
  • Benchmark system configuration changes

Build and Run

Build 

cd example/memory_test
./build.sh

Run 

./memory_test [options]

Options:

Option Description Default
-s Total memory size (GiB) 192
-t Number of parallel tasks 3072
-o Operation: read or write read
-i Number of iterations 1
-u Loop unrolling: 0 or 1 1

Examples:

# Default test: 192 GiB read with 3072 tasks
./memory_test

# Write test with 64 GiB, 1024 tasks
./memory_test -s 64 -t 1024 -o write

# Compare with/without loop unrolling
./memory_test -s 128 -t 2048 -u 1
./memory_test -s 128 -t 2048 -u 0

Output example (QEMU) 

./memory_test -s 1
=== Memory Bandwidth Test ===
Operation        : read
Total Memory Size: 1023.94 MiB (1.00 GiB)
Task Count       : 3072
Size per Task    : 349504 bytes (341.31 KiB)
Iterations       : 1
Use Unrolling    : 1
-----------------------------
Elapsed Time     : 35275.891 ms
Throughput       : 0.028 GiB/s (0.030 GB/s)

Code Structure

Host Application

main.cpp performs:

  1. Parse command-line arguments
  2. Initialize PXL runtime and load kernel
  3. Allocate memory and flush host cache
  4. Execute kernel across all tasks
  5. Measure time and calculate throughput

MU Kernel

mu_kernel/mu_memory_test.cpp implements:

  • Memory read/write operations
  • Loop unrolling optimization (8 x 64-bit accesses per iteration)
  • Per-task memory region calculation using mu::getTaskIdx()

Next Steps