move perf counter to plugin

PSeitz · Oct 14, 2024 · 1132596 · 1132596
1 parent 06a7b4f
commit 1132596
Show file tree

Hide file tree

Showing 9 changed files with 74 additions and 62 deletions.
diff --git a/README.md b/README.md
@@ -27,7 +27,7 @@ It is designed to be simple to use and to provide a good overview of the perform
 ### Example
 
 ```rust
-use binggan::{black_box, plugins::CacheTrasher, InputGroup, PeakMemAlloc, INSTRUMENTED_SYSTEM};
+use binggan::{black_box, plugins::*, InputGroup, PeakMemAlloc, INSTRUMENTED_SYSTEM};
 
 #[global_allocator]
 pub static GLOBAL: &PeakMemAlloc<std::alloc::System> = &INSTRUMENTED_SYSTEM;
@@ -43,12 +43,12 @@ fn test_hashmap(data: &Vec<usize>) {
 fn bench_group(mut runner: InputGroup<Vec<usize>>) {
     runner.set_alloc(GLOBAL); // Set the peak mem allocator. This will enable peak memory reporting.
 
-    // Enables the perf integration. Only on Linux, noop on other OS.
-    runner.config().enable_perf();
-    // Trashes the CPU cache between runs
     runner
         .get_plugin_manager()
+        // Trashes the CPU cache between runs
         .add_plugin(CacheTrasher::default());
+        // Enables the perf integration. Only on Linux, noop on other OS.
+        .add_plugin(PerfCounterPlugin::default());
     // Enables throughput reporting
     runner.throughput(|input| input.len() * std::mem::size_of::<usize>());
     runner.register("vec", |data| {

diff --git a/benches/bench_group.rs b/benches/bench_group.rs
@@ -1,6 +1,10 @@
 use std::collections::HashMap;
 
-use binggan::{black_box, plugins::CacheTrasher, BenchRunner, PeakMemAlloc, INSTRUMENTED_SYSTEM};
+use binggan::{
+    black_box,
+    plugins::{CacheTrasher, PerfCounterPlugin},
+    BenchRunner, PeakMemAlloc, INSTRUMENTED_SYSTEM,
+};
 
 #[global_allocator]
 pub static GLOBAL: &PeakMemAlloc<std::alloc::System> = &INSTRUMENTED_SYSTEM;
@@ -34,10 +38,10 @@ fn run_bench() {
     let mut runner: BenchRunner = BenchRunner::new();
     runner.set_alloc(GLOBAL); // Set the peak mem allocator. This will enable peak memory reporting.
 
-    runner.config().enable_perf();
     runner
         .get_plugin_manager()
-        .add_plugin(CacheTrasher::default());
+        .add_plugin(CacheTrasher::default())
+        .add_plugin(PerfCounterPlugin::default());
 
     for (input_name, data) in inputs.iter() {
         let mut group = runner.new_group();

diff --git a/benches/bench_input.rs b/benches/bench_input.rs
@@ -1,6 +1,10 @@
 use std::collections::HashMap;
 
-use binggan::{black_box, plugins::CacheTrasher, InputGroup, PeakMemAlloc, INSTRUMENTED_SYSTEM};
+use binggan::{
+    black_box,
+    plugins::{CacheTrasher, PerfCounterPlugin},
+    InputGroup, PeakMemAlloc, INSTRUMENTED_SYSTEM,
+};
 
 #[global_allocator]
 pub static GLOBAL: &PeakMemAlloc<std::alloc::System> = &INSTRUMENTED_SYSTEM;
@@ -26,12 +30,12 @@ fn test_hashmap(data: &Vec<usize>) -> HashMap<usize, i32> {
 fn bench_group(mut runner: InputGroup<Vec<usize>, u64>) {
     runner.set_alloc(GLOBAL); // Set the peak mem allocator. This will enable peak memory reporting.
 
-    // Enables the perf integration. Only on Linux, noop on other OS.
-    runner.config().enable_perf();
-    // Trashes the CPU cache between runs
     runner
         .get_plugin_manager()
-        .add_plugin(CacheTrasher::default());
+        // Trashes the CPU cache between runs
+        .add_plugin(CacheTrasher::default())
+        // Enables the perf integration. Only on Linux, noop on other OS.
+        .add_plugin(PerfCounterPlugin::default());
     // Enables throughput reporting
     runner.throughput(|input| input.len() * std::mem::size_of::<usize>());
     runner.register("vec", |data| {

diff --git a/benches/test_throughput.rs b/benches/test_throughput.rs
@@ -1,6 +1,9 @@
 use std::time::{Duration, Instant};
 
-use binggan::{plugins::CacheTrasher, BenchRunner, PeakMemAlloc, INSTRUMENTED_SYSTEM};
+use binggan::{
+    plugins::{CacheTrasher, PerfCounterPlugin},
+    BenchRunner, PeakMemAlloc, INSTRUMENTED_SYSTEM,
+};
 
 #[global_allocator]
 pub static GLOBAL: &PeakMemAlloc<std::alloc::System> = &INSTRUMENTED_SYSTEM;
@@ -9,10 +12,10 @@ fn run_bench() {
     let mut runner: BenchRunner = BenchRunner::new();
     runner.set_alloc(GLOBAL); // Set the peak mem allocator. This will enable peak memory reporting.
 
-    runner.config().enable_perf();
     runner
         .get_plugin_manager()
-        .add_plugin(CacheTrasher::default());
+        .add_plugin(CacheTrasher::default())
+        .add_plugin(PerfCounterPlugin::default());
     runner.config().set_num_iter_for_group(128);
 
     let mut group = runner.new_group();

diff --git a/src/bench_runner.rs b/src/bench_runner.rs
@@ -139,14 +139,6 @@ impl BenchRunner {
         if group.is_empty() {
             return;
         }
-        #[cfg(target_os = "linux")]
-        {
-            use crate::plugins::perf_counter::PerfCounterPlugin;
-            if self.config().enable_perf {
-                self.plugins
-                    .add_plugin_if_absent(PerfCounterPlugin::default());
-            }
-        }
 
         self.plugins.emit(PluginEvents::GroupStart {
             runner_name: self.name.as_deref(),

diff --git a/src/config.rs b/src/config.rs
@@ -8,8 +8,6 @@ pub struct Config {
     /// The filter for the benchmarks
     /// This is read from the command line by default.
     pub filter: Option<String>,
-    /// Enable/disable perf integration
-    pub enable_perf: bool,
     /// Verbose output of binggan. Prints the number of iterations.
     pub verbose: bool,
     /// Manually set the number of iterations the benchmarks registered afterwards are called.
@@ -26,7 +24,6 @@ impl Default for Config {
         Config {
             interleave: true,
             filter: None,
-            enable_perf: false,
             verbose: false,
             num_iter_bench: None,
             num_iter_group: None,
@@ -85,35 +82,6 @@ impl Config {
         self.interleave = interleave;
         self
     }
-
-    /// Enable perf profiling + report
-    ///
-    /// The numbers are reported with the following legend:
-    /// ```bash
-    /// L1dA: L1 data access
-    /// L1dM: L1 data misses
-    /// Br: branches
-    /// BrM: missed branches
-    /// ```
-    /// e.g.
-    /// ```bash
-    /// fibonacci    Memory: 0 B       Avg: 135ns      Median: 136ns     132ns          140ns    
-    ///              L1dA: 809.310     L1dM: 0.002     Br: 685.059       BrM: 0.010     
-    /// baseline     Memory: 0 B       Avg: 1ns        Median: 1ns       1ns            1ns      
-    ///              L1dA: 2.001       L1dM: 0.000     Br: 6.001         BrM: 0.000     
-    /// ```
-    ///
-    /// # Note:
-    /// This is only available on Linux. On other OSs this does nothing.
-    ///
-    /// Perf may run into limitations where all counters are reported as zero. <https://github.com/jimblandy/perf-event/issues/2>.
-    /// Disabling the NMI watchdog should help:
-    ///
-    /// `sudo sh -c "echo '0' > /proc/sys/kernel/nmi_watchdog"`
-    pub fn enable_perf(&mut self) -> &mut Self {
-        self.enable_perf = true;
-        self
-    }
 }
 
 pub(crate) fn parse_args() -> Config {

diff --git a/src/lib.rs b/src/lib.rs
@@ -37,11 +37,11 @@
 //!
 //! # Perf Integration
 //! Binggan can integrate with perf to report hardware performance counters.
-//! See [Config::enable_perf](crate::Config::enable_perf) for more information.
+//! See [PerfCounterPlugin](crate::plugins::PerfCounterPlugin) for more information.
 //!
 //! # Example for InputGroup
 //! ```rust
-//! use binggan::{black_box, InputGroup, PeakMemAlloc, INSTRUMENTED_SYSTEM};
+//! use binggan::{black_box, InputGroup, PeakMemAlloc, INSTRUMENTED_SYSTEM, plugins::*};
 //!
 //! #[global_allocator]
 //! pub static GLOBAL: &PeakMemAlloc<std::alloc::System> = &INSTRUMENTED_SYSTEM;
@@ -64,7 +64,9 @@
 //! // Run the benchmark for the group with input `Vec<usize>`
 //! fn bench_group(mut runner: InputGroup<Vec<usize>, u64>) {
 //!     runner.set_alloc(GLOBAL); // Set the peak mem allocator. This will enable peak memory reporting.
-//!     runner.config().enable_perf(); // Enable perf integration. This only works on linux.
+//!     runner
+//!        .get_plugin_manager()
+//!        .add_plugin(PerfCounterPlugin::default());
 //!     runner.register("vec", move |data| {
 //!         let vec = test_vec(data);
 //!         Some(vec.len() as u64)
@@ -135,7 +137,6 @@
 //!     let mut runner: BenchRunner = BenchRunner::new();
 //!     runner.set_alloc(GLOBAL); // Set the peak mem allocator. This will enable peak memory reporting.
 //!
-//!     runner.config().enable_perf();
 //!     runner
 //!        .get_plugin_manager()
 //!        .add_plugin(CacheTrasher::default());

diff --git a/src/plugins/perf_counter/linux.rs b/src/plugins/perf_counter/linux.rs
@@ -97,9 +97,44 @@ impl PerfCounters {
     }
 }
 
-/// Perf Counter Plugin.
 ///
-/// Stores one counter group per bench id.
+/// Plugin to report perf counters.
+///
+/// The numbers are reported with the following legend:
+/// ```bash
+/// Br: Branches
+/// MBr: Missed Branches
+/// L1dA: L1 Data Access
+/// L1dM: L1 Data Access Misses
+/// TLBdA: Translation Lookaside Buffer Data Access
+/// TLBdM: Translation Lookaside Buffer Data Access Misses
+/// ```
+/// e.g.
+/// ```bash
+/// fibonacci    Memory: 0 B       Avg: 135ns      Median: 136ns     132ns          140ns    
+///              L1dA: 809.310     L1dM: 0.002     Br: 685.059       MBr: 0.010     
+/// baseline     Memory: 0 B       Avg: 1ns        Median: 1ns       1ns            1ns      
+///              L1dA: 2.001       L1dM: 0.000     Br: 6.001         MBr: 0.000     
+/// ```
+///
+/// # Note:
+/// This is only available on Linux. On other OSs this does nothing.
+///
+/// Perf may run into limitations where all counters are reported as zero. <https://github.com/jimblandy/perf-event/issues/2>.
+/// Disabling the NMI watchdog should help:
+///
+/// `sudo sh -c "echo '0' > /proc/sys/kernel/nmi_watchdog"`
+///
+/// ## Usage Example
+/// ```rust
+/// use binggan::{*, plugins::*}
+///
+/// let mut runner = BenchRunner::new();
+/// runner
+///    .get_plugin_manager()
+///    .add_plugin(PerfCounterPlugin::default());
+/// ```
+
 #[derive(Default)]
 pub struct PerfCounterPlugin {
     perf_per_bench: PerBenchData<Option<PerfCounters>>,

diff --git a/src/plugins/perf_counter/mod.rs b/src/plugins/perf_counter/mod.rs
@@ -4,6 +4,12 @@ pub(crate) mod linux;
 #[cfg(target_os = "linux")]
 pub use linux::*;
 
+#[cfg(not(target_os = "linux"))]
+pub(crate) mod dummy;
+
+#[cfg(not(target_os = "linux"))]
+pub use dummy::*;
+
 use crate::stats::*;
 use miniserde::Deserialize;
 use miniserde::Serialize;
@@ -92,7 +98,6 @@ impl CounterValues {
             print_counter_value("L1dM", &self, other, |stats| stats.l1d_miss_count),
             print_counter_value("TLBdA", &self, other, |stats| stats.tlbd_access_count),
             print_counter_value("TLBdM", &self, other, |stats| stats.tlbd_miss_count),
-            print_counter_value("L1dA", &self, other, |stats| stats.l1d_access_count),
         ]
     }
 }