A platform to run conditional branch prediction experiments on real binaries. It can:

1. Capture branch trace from real binaries
2. Run SimPoint methodology on branch traces to reduce size and simulation speed
3. Simulate branch predictors (from CBP2016, etc.) on the branch traces

1. Clone this repository, and:

cargo build --release

2. Build tracers under tracers directory if necessary. See tracers/README.md for details.
3. Run experiments:
   1. Prepare benchmark config under benchmarks/[config]/config.json, see Configuration section
   2. Record traces: cargo run --release --bin benchmark -- record --tracer [tracer] --config-name [config], traces are stored under benchmarks/[config]/traces/final
   3. Display trace statistics: cargo run --release --bin benchmark -- info --config-name [config]
   4. Run SimPoint clustering: cargo run --release --bin benchmark -- simpoint --config-name [config] --size [instructions]
   5. Run branch prediction: cargo run --release --bin benchmark -- simulate --config-name [config] --predictor [predictor]
   6. Find results under: benchmarks/[config]/[simulate]/[datetime]-[predictor]/per-benchmark or use cargo run --release --bin benchmark -- report

Example #1 leela benchmark:

cargo run --release --bin benchmark -- record --config-name leela --tracer intel-pt 
cargo run --release --bin benchmark -- info --config-name leela
cargo run --release --bin benchmark -- simpoint --config-name leela --size 100000000
cargo run --release --bin benchmark -- simulate --config-name leela --predictor AndreSeznec-TAGE-SC-L-8KB

Example #2 some dynamically linked binaries:

cargo run --release --bin benchmark -- record --config-name test --tracer intel-pt 
cargo run --release --bin benchmark -- info --config-name test
cargo run --release --bin benchmark -- simpoint --config-name test --size 1000
cargo run --release --bin benchmark -- simulate --config-name test --predictor AndreSeznec-TAGE-SC-L-8KB

Example #3 full spec int 2017 rate:

# it takes 20+ hours
cargo run --release --bin benchmark -- record --config-name spec-int-2017-rate --tracer pin
cargo run --release --bin benchmark -- info --config-name spec-int-2017-rate
cargo run --release --bin benchmark -- simpoint --config-name spec-int-2017-rate --size 100000000
cargo run --release --bin benchmark -- simulate --config-name spec-int-2017-rate --predictor AndreSeznec-TAGE-SC-L-8KB

Configuration hierarchy:

1. config, e.g. spec-int-2017-rate-1
2. benchmark, e.g. leela
3. command, e.g. leela ref.sgf
4. simpoint, e.g. the first simpoint phase of leela ref.sgf

You can add benchmarks to benchmarks/[config]/config.json, like:

{
    "benchmarks": [{
        "name": "ls",
        "commands": [{
            "command": "/usr/bin/ls ."
        }]
    }, {
        "name": "cat",
        "commands": [{
            "command": "/usr/bin/cat /proc/self/maps"
        }]
    }]
}

fork breaks intel pt converter due to missing filtering of events from child processes