μBench

μBench is a microbenchmarking tool to evaluate microarchitectural features on run real hardware and gem5 simulator. The tool is designed to be extensible, allowing users to add easily add their own microbenchmarks.

Refer to the benchmarks README for more information on the available benchmarks and how to add new ones.

Build the Benchmarks

Native Build

```bash

Creates a build directory and runs cmake to generate the build files

The -DARCH flag is used to specify the architecture for which the benchmarks are being built.

The default is x86_64, but you can specify aarch64

cmake -B build -DARCH=x86_64

Build the benchmark.

cmake --build ./build -j $(nproc) ```

Using Docker + Cross Compilation

μBench is designed to be cross-platform and can be built for different architectures using the -DARCH flag. We use docker (dockcross) as a cross-compilation toolchain.

You can use the provided Makefile to build the benchmarks for different architectures which will automatically download the required docker images and run the cross-compilation commands.

bash make <arch>

Where <arch> can be one of the following: - x86: Build for x86_64 architecture (default) - arm64: Build for ARM64 architecture - riscv64: Build for RISC-V architecture - all: Build for all architectures

You will find the compiled binaries in the corresponing build-<arch> folder.

If you want to cross-compile the microbenchmarking tool for ARM or X86 we can use dockcross

```bash

The following command will create an invocation script for the cross compilation toolchain for ARM64

docker run --rm dockcross/linux-arm64 > ./dockcross-arm64 chmod +x ./dockcross-arm64

Once you have the invocation script, you can run the cmake command to generate the build files for ARM64

./dockcross-arm64 cmake -B -DARCH=aarch64

Build the benchmark.

./dockcross-arm64 cmake --build -j $(nproc) ```

The same for x86_64:

```bash

The following command will create an invocation script for the cross compilation toolchain for ARM64

docker run --rm dockcross/linux-x64 > ./dockcross-x64 chmod +x ./dockcross-x64

Once you have the invocation script, you can run the cmake command to generate the build files for ARM64

./dockcross-x64 cmake -B

Build the benchmark.

./dockcross-x64 cmake --build -j $(nproc) ```

Running a benchmark

To run a benchmark, you need to specify the benchmark to run and optionally the number of repetitions. There are two ways to select the benchmark to run. (1) By name passing the benchmark name with the --bmname/-b flag to ubench or (2) by providing a configuration file with the --config/-c flag. The configuration file is a YAML file that can contain further configuration options for the benchmark. See ./configs for example configuration files.

```bash

Run the benchmark with the benchmark name

./build/ubench --bmname --repeats

Run the benchmark with a configuration file

./build/ubench --config --repeats ```

Run the benchmark in gem5

The folder contains a minimal gem5 configuration to run the microbenchmarks in system call emulation mode (SE). ```bash

Run the benchmark in gem5

/build//gem5.opt ./gem5-configs/se-simple.py --isa build-/ubench --bmname --repeats ```

Example: ```bash

Run the benchmark in gem5

/build/X86/gem5.opt ./gem5-configs/se-simple.py --isa X86 build-x64/ubench --bmname simple_loop --repeats 10 ```

M5Ops

The microbenchmarks is compiled using gem5's magic instructions (m5ops). These instructions are useful the exit the simulation loop at specific point during the simulation. By passing the --m5ops/-m flag to the microbenchmark, the benchmark execute a workitem_start before a repetition starts and a workitem_end after the repetition ends. The user can then define in the gem5 configuration file what to do when the workitem_start and workitem_end instructions are executed. For example dumping or resetting the stats. Refer to the se-simple.py for an example of how to process the workitem exit instructions.

Benchmarks

All the benchmarks are located in the benchmarks folder. Refer to the benchmarks README for more information on the benchmarks and how to add new ones.