Deduplicate walltime perf symbols, unwind data and debug info across pids

AGENTS.md

@@ -17,21 +17,16 @@ cargo build
 # Build in release mode
 cargo build --release
 
-# Run tests (prefer nextest if available)
-cargo nextest run  # preferred if installed
-cargo test         # fallback if nextest is not available
+# Run tests
+cargo test
 
 # Run specific test
-cargo nextest run <test_name>  # with nextest
-cargo test <test_name>         # with cargo test
+cargo test <test_name>
 
 # Run tests with output
-cargo nextest run --nocapture  # with nextest
-cargo test -- --nocapture      # with cargo test
+cargo test -- -nocapture
 ```
 
-**Note**: Always check if `cargo nextest` is available first (with `cargo nextest --version` or `which cargo-nextest`). If available, use it instead of `cargo test` as it provides faster and more reliable test execution.
-
 ### Running the Application
 
 ```bash
@@ -99,7 +94,7 @@ The core functionality for running benchmarks:
 
 The project uses:
 
-- `cargo nextest` (preferred) or standard Rust `cargo test`
+- `cargo test`
 - `insta` for snapshot testing
 - `rstest` for parameterized tests
 - `temp-env` for environment variable testing
@@ -108,5 +103,4 @@ Test files include snapshots in `snapshots/` directories for various run environ
 
 **Important**:
 
-- Always prefer `cargo nextest run` over `cargo test` when running tests, as it provides better performance and reliability.
-- Some walltime executor tests require `sudo` access and will fail in non-interactive environments (e.g., `test_walltime_executor::*`). These failures are expected if sudo is not available.
+- Some tests require `sudo` access. They are skipped by default unless the `GITHUB_ACTIONS` env var is set.

CONTRIBUTING.md

@@ -96,3 +96,7 @@ This ensures only stable runner releases are marked as "latest" in GitHub.
 ## Known issue
 
 - If one of the crates is currenlty in beta version, for example the runner is in beta version 4.4.2-beta.1, any alpha release will fail for the any crate, saying that only minor, major or patch releases is supported.
+
+## Testing
+
+- Some tests require `sudo` access. They are skipped by default unless the `GITHUB_ACTIONS` env var is set.

Cargo.toml

@@ -10,6 +10,10 @@ default-run = "codspeed"
 name = "codspeed"
 path = "src/main.rs"
 
+[[bin]]
+name = "compare_walltime_output"
+path = "src/bin/compare_walltime_output.rs"
+
 
 [dependencies]
 anyhow = { workspace = true }

build.rs

@@ -0,0 +1,6 @@
+fn main() {
+    // Force a rebuild of the test target to be able to run the full test suite locally just by
+    // setting GITHUB_ACTIONS=1 in the environment.
+    // This is because `test_with` is evaluated at build time
+    println!("cargo::rerun-if-env-changed=GITHUB_ACTIONS");
+}

crates/memtrack/build.rs

@@ -3,6 +3,11 @@ use std::{env, path::PathBuf};
 
 #[cfg(feature = "ebpf")]
 fn build_ebpf() {
+    // Force a rebuild of the test target to be able to run the full test suite locally just by
+    // setting GITHUB_ACTIONS=1 in the environment.
+    // This is because `test_with` is evaluated at build time
+    println!("cargo::rerun-if-env-changed=GITHUB_ACTIONS");
+
     use libbpf_cargo::SkeletonBuilder;
 
     println!("cargo:rerun-if-changed=src/ebpf/c");

crates/runner-shared/src/debug_info.rs

@@ -21,6 +21,13 @@ impl std::fmt::Debug for DebugInfo {
     }
 }
 
+/// Per-pid mounting info referencing a deduplicated debug info entry.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct DebugInfoPidMapping {
+    pub debug_info_key: String,
+    pub load_bias: u64,
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct ModuleDebugInfo {
     /// The path to the object file on disk (e.g. `/usr/lib/libc.so.6`)

crates/runner-shared/src/lib.rs

@@ -3,5 +3,6 @@ pub mod debug_info;
 pub mod fifo;
 pub mod metadata;
 pub mod perf_event;
+pub mod perf_map;
 pub mod unwind_data;
 pub mod walltime_results;

crates/runner-shared/src/metadata.rs

@@ -1,10 +1,14 @@
 use anyhow::Context;
 use serde::{Deserialize, Serialize};
+use std::collections::HashMap;
 use std::io::BufWriter;
 use std::path::Path;
+use std::path::PathBuf;
 
-use crate::debug_info::ModuleDebugInfo;
+use crate::debug_info::{DebugInfoPidMapping, ModuleDebugInfo};
 use crate::fifo::MarkerType;
+use crate::perf_map::SymbolPidMapping;
+use crate::unwind_data::UnwindDataPidMapping;
 
 #[derive(Serialize, Deserialize)]
 pub struct PerfMetadata {
@@ -25,9 +29,31 @@ pub struct PerfMetadata {
     #[deprecated(note = "Use ExecutionTimestamps in the 'artifacts' module instead")]
     pub markers: Vec<MarkerType>,
 
-    /// Debug info for all modules across all processes, mapping PID to module debug info
-    #[serde(default, skip_serializing_if = "std::collections::HashMap::is_empty")]
-    pub debug_info_by_pid: std::collections::HashMap<i32, Vec<ModuleDebugInfo>>,
+    /// Kept for backward compatibility, was used before deduplication of debug info entries.
+    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
+    #[deprecated(note = "Use 'debug_info' + 'debug_info_pid_mappings_by_pid' instead")]
+    pub debug_info_by_pid: HashMap<i32, Vec<ModuleDebugInfo>>,
+
+    /// Deduplicated debug info entries, keyed by semantic key
+    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
+    pub debug_info: HashMap<String, ModuleDebugInfo>,
+
+    /// Per-pid debug info references, mapping PID to list of debug info index + load bias
+    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
+    pub debug_info_pid_mappings_by_pid: HashMap<i32, Vec<DebugInfoPidMapping>>,
+
+    /// Per-pid unwind data references, mapping PID to list of unwind data index + mounting info
+    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
+    pub unwind_data_pid_mappings_by_pid: HashMap<i32, Vec<UnwindDataPidMapping>>,
+
+    /// Per-pid symbol references, mapping PID to list of perf map index + load bias
+    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
+    pub symbol_pid_mappings_by_pid: HashMap<i32, Vec<SymbolPidMapping>>,
+
+    /// Mapping from semantic key to original binary path
+    /// Kept for traceability, and if we ever need to reconstruct the original paths from the keys
+    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
+    pub key_to_path: HashMap<String, PathBuf>,
 }
 
 impl PerfMetadata {

crates/runner-shared/src/perf_map.rs

@@ -0,0 +1,11 @@
+use serde::{Deserialize, Serialize};
+
+/// File suffix used when registering module symbols in a PID agnostic way.
+pub const SYMBOLS_MAP_SUFFIX: &str = "symbols.map";
+
+/// Per-pid mounting info referencing a deduplicated perf map entry.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SymbolPidMapping {
+    pub perf_map_key: String,
+    pub load_bias: u64,
+}