Add support for memory-profiling on subsystem-bench

Cargo.toml

@@ -810,6 +810,7 @@ is-terminal = { version = "0.4.9" }
 is_executable = { version = "1.0.1" }
 isahc = { version = "1.2" }
 itertools = { version = "0.11" }
+jemalloc_pprof = { version = "0.4" }
 jobserver = { version = "0.1.26" }
 jsonpath_lib = { version = "0.3" }
 jsonrpsee = { version = "0.24.3" }

polkadot/node/subsystem-bench/Cargo.toml

@@ -19,7 +19,11 @@ path = "src/cli/subsystem-bench.rs"
 # Prevent rustdoc error. Already documented from top-level Cargo.toml.
 doc = false
 
+
 [dependencies]
+tikv-jemallocator = { features = ["profiling", "unprefixed_malloc_on_supported_platforms"], workspace = true, optional = true }
+jemalloc_pprof = { workspace = true, optional = true }
+polkadot-service = { workspace = true, default-features = true }
 polkadot-node-subsystem = { workspace = true, default-features = true }
 polkadot-node-subsystem-util = { workspace = true, default-features = true }
 polkadot-node-subsystem-types = { workspace = true, default-features = true }
@@ -93,3 +97,7 @@ strum = { features = ["derive"], workspace = true, default-features = true }
 
 [features]
 default = []
+memprofile = [
+	"dep:jemalloc_pprof",
+	"dep:tikv-jemallocator",
+]

polkadot/node/subsystem-bench/README.md

@@ -260,6 +260,41 @@ This file is best interpreted with `cg_annotate --auto=yes cachegrind.out.<pid>`
 
 For finer profiling of cache misses, better use `perf` on a bare-metal machine.
 
+### Profile memory usage using jemalloc
+
+Bellow you can find instructions how to setup and run profiling with jemalloc, this is complementary
+with using other memory profiling tools like: <https://github.com/koute/bytehound?tab=readme-ov-file#basic-usage>.
+
+#### Prerequisites
+
+Install tooling with:
+
+```
+sudo apt install libjemalloc-dev graphviz
+```
+
+#### Generate memory usage snapshots
+
+Memory usage can be profiled by running any subsystem benchmark with `--features memprofile`, e.g:
+
+```
+RUSTFLAGS=-g cargo run -p polkadot-subsystem-bench --release --features memprofile -- polkadot/node/subsystem-bench/examples/approvals_throughput.yaml
+```
+
+#### Interpret the results
+
+After the benchmark ran the memory usage snapshots can be found in `/tmp/subsystem-bench*`, to extract the information
+from a snapshot you can use `jeprof` like this:
+
+```
+jeprof --text PATH_TO_EXECUTABLE_WITH_DEBUG_SYMBOLS /tmp/subsystem-bench.1222895.199.i199.heap > statistics.txt
+```
+
+Useful links:
+
+- Tutorial: <https://www.magiroux.com/rust-jemalloc-profiling/>
+- Jemalloc configuration options: <https://jemalloc.net/jemalloc.3.html>
+
 ## Create new test objectives
 
 This tool is intended to make it easy to write new test objectives that focus individual subsystems,

polkadot/node/subsystem-bench/src/cli/subsystem-bench.rs

@@ -182,6 +182,17 @@ impl BenchCli {
 	}
 }
 
+#[cfg(feature = "memprofile")]
+#[global_allocator]
+static ALLOC: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
+
+#[cfg(feature = "memprofile")]
+#[allow(non_upper_case_globals)]
+#[export_name = "malloc_conf"]
+// See https://jemalloc.net/jemalloc.3.html for more information on the configuration options.
+pub static malloc_conf: &[u8] =
+	b"prof:true,prof_active:true,lg_prof_interval:30,lg_prof_sample:21,prof_prefix:/tmp/subsystem-bench\0";
+
 fn main() -> eyre::Result<()> {
 	color_eyre::install()?;
 	sp_tracing::try_init_simple();