A rust library to collect and convert Heap profiling data from the jemalloc allocator and convert it to the pprof format.

To understand how to use this together with Polar Signals Cloud to continuously collect profiling data, refer to the Use with Polar Signals Cloud section.

This code was originally developed as part of Materialize, and then in a collaboration extracted into this standalone library.

Currently, this library only supports Linux.

Furthermore, you must be able to switch your allocator to jemalloc. If you need to continue using the default system allocator for any reason, this library will not be useful.

Internally this library uses tikv-jemalloc-ctl to interact with jemalloc, so to use it, you must use the jemalloc allocator via the tikv-jemallocator library.

When adding tikv-jemallocator as a dependency, make sure to enable the profiling feature.

[dependencies]
[target.'cfg(not(target_env = "msvc"))'.dependencies]
tikv-jemallocator = { version = "0.5.4", features = ["profiling", "unprefixed_malloc_on_supported_platforms"] }

Note: We also recommend enabling the unprefixed_malloc_on_supported_platforms feature, not strictly necessary, but will influence the rest of the usage.

Then configure the global allocator and configure it with profiling enabled.

#[cfg(not(target_env = "msvc"))]
#[global_allocator]
static ALLOC: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;

#[allow(non_upper_case_globals)]
#[export_name = "malloc_conf"]
pub static malloc_conf: &[u8] = b"prof:true,prof_active:true,lg_prof_sample:19\0";

If you do not use the unprefixed_malloc_on_supported_platforms feature, you have to name it _rjem_malloc_conf it instead of malloc_conf.

2^19 bytes (512KiB) is the default configuration for the sampling period, but we recommend being explicit. To understand more about jemalloc sampling check out the detailed docs on it.

We recommend serving the profiling data on an HTTP server such as axum, that could look like this, and we'll intentionally include a 4mb allocation to trigger sampling.

#[tokio::main]
async fn main() {
    let mut v = vec![];
    for i in 0..1000000 {
        v.push(i);
    }

    let app = axum::Router::new()
        .route("/debug/pprof/heap", axum::routing::get(handle_get_heap));

    // run our app with hyper, listening globally on port 3000
    let listener = tokio::net::TcpListener::bind("0.0.0.0:3000").await.unwrap();
    axum::serve(listener, app).await.unwrap();
}

use axum::http::StatusCode;
use axum::response::IntoResponse;

pub async fn handle_get_heap() -> Result<impl IntoResponse, (StatusCode, String)> {
    let mut prof_ctl = jemalloc_pprof::PROF_CTL.as_ref().unwrap().lock().await;
    require_profiling_activated(&prof_ctl)?;
    let pprof = prof_ctl
        .dump_pprof()
        .map_err(|err| (StatusCode::INTERNAL_SERVER_ERROR, err.to_string()))?;
    Ok(pprof)
}

/// Checks whether jemalloc profiling is activated an returns an error response if not.
fn require_profiling_activated(prof_ctl: &jemalloc_pprof::JemallocProfCtl) -> Result<(), (StatusCode, String)> {
    if prof_ctl.activated() {
        Ok(())
    } else {
        Err((axum::http::StatusCode::FORBIDDEN, "heap profiling not activated".into()))
    }
}

Then running the application, we can capture a profile and view it the pprof toolchain.

curl localhost:3000/debug/pprof/heap > heap.pb.gz
pprof -http=:8080 heap.pb.gz

Note: The profiling data is not symbolized, so either addr2line or llvm-addr2line needs to be available in the path and pprof needs to be able to discover the respective debuginfos.

The way this library works is that it creates a new temporary file (in the platform-specific default temp dir), and instructs jemalloc to dump a profile into that file. Therefore the platform respective temporary directory must be writeable by the process. After reading and converting it to pprof, the file is cleaned up via the destructor. A single profile tends to be only a few kilobytes large, so it doesn't require a significant space, but it's non-zero and needs to be writeable.

Polar Signals Cloud allows continuously collecting heap profiling data, so you always have the right profiling data available, and don't need to search for the right data, you already have it!

Polar Signals Cloud supports anything in the pprof format, so a process exposing the above explained pprof endpoint, can then be scraped as elaborated in the scraping docs.

The functionality to dump the current jemalloc heap profile in pprof format is exposed to C and C++ (or any other language that can use jemalloc and can link against libraries via the C ABI). This functionality is exposed via the capi (C API) package.

The following prerequisites are necessary to build the C API package:

• Working Rust and C toolchains. The former can be installed by following the instructions at https://rustup.rs . The latter can be installed via the distribution's package manager. For example, on Ubuntu, run sudo apt install build-essential.
• jemalloc and its development headers. For example, on Ubuntu, run sudo apt install jemalloc-dev.

Once the prerequisites are installed, the library can be built by running cargo build -p capi --release. There are three files of interest:

• The library itself, produced at target/release/libjemalloc_pprof.so
• A header file, at capi/include/jemalloc_pprof.h
• A manual page, at capi/man/jemalloc_pprof.3.

The procedure for installing and using these files depends on your distribution and build system.

Ensure that your binaries link against both jemalloc and jemalloc_pprof by passing the linker flags -ljemalloc -ljemalloc_pprof. The procedure for ensuring that these flags are passed depends on your build system and is currently outside the scope of this document.

Once that is done, profiling can be enabled either by setting the MALLOC_CONF variable or by defining a symbol called malloc_conf in the binary. For example:

export MALLOC_CONF="prof:true,prof_active:true,lg_prof_sample:19"

See the jemalloc man page for more details. When profiling is enabled, a profile may be dumped in pprof format via the dump_jemalloc_pprof function.

This program allocates between 1 and 10 MiB every 100 milliseconds, and dumps a profile to the file my_profile every 2 seconds.

#include <assert.h>
#include <errno.h>
#include <unistd.h>
#include <stdlib.h>
#include <pthread.h>
#include <stdio.h>

#include <jemalloc_pprof.h>

void
a()
{
        size_t sz = 1 * 1024 * 1024;
        char *x = malloc(sz);
        for (size_t i = 0; i < sz; ++i) {
                x[i] = '\0';
        }
}

void
b()
{
        size_t sz = 2 * 1024 * 1024;
        char *x = malloc(sz);
        for (size_t i = 0; i < sz; ++i) {
                x[i] = '\0';
        }
}

void
c()
{
        size_t sz = 3 * 1024 * 1024;
        char *x = malloc(sz);
        for (size_t i = 0; i < sz; ++i) {
                x[i] = '\0';
        }
}

void
d()
{
        size_t sz = 4 * 1024 * 1024;
        char *x = malloc(sz);
        for (size_t i = 0; i < sz; ++i) {
                x[i] = '\0';
        }
}

void
e()
{
        size_t sz = 5 * 1024 * 1024;
        char *x = malloc(sz);
        for (size_t i = 0; i < sz; ++i) {
                x[i] = '\0';
        }
}

void
f()
{
        size_t sz = 6 * 1024 * 1024;
        char *x = malloc(sz);
        for (size_t i = 0; i < sz; ++i) {
                x[i] = '\0';
        }
}

void
g()
{
        size_t sz = 7 * 1024 * 1024;
        char *x = malloc(sz);
        for (size_t i = 0; i < sz; ++i) {
                x[i] = '\0';
        }
}

void
h()
{
        size_t sz = 8 * 1024 * 1024;
        char *x = malloc(sz);
        for (size_t i = 0; i < sz; ++i) {
                x[i] = '\0';
        }
}

void
j()
{
        size_t sz = 9 * 1024 * 1024;
        char *x = malloc(sz);
        for (size_t i = 0; i < sz; ++i) {
                x[i] = '\0';
        }
}

void
k()
{
        size_t sz = 10 * 1024 * 1024;
        char *x = malloc(sz);
        for (size_t i = 0; i < sz; ++i) {
                x[i] = '\0';
        }
}

void *
repeatedly_dump(void *ignored)
{
        char *buf;
        size_t len = 0;
        int result;
        for (;;) {
                sleep(2);
                result = dump_jemalloc_pprof(&buf, &len);
                if (result != JP_SUCCESS) {
                        fprintf(stderr, "errno: %d\n", errno);
                        continue;
                }
                if (buf) {                        
                        FILE *file = fopen("my_profile", "w");
                        assert(file);

                        fwrite(buf, sizeof(char), len, file);    
                        fclose(file);
                        printf("dumped pprof of size %lu\n", len);
                        free(buf);
                }
        }
        return NULL;
}

int
main()
{
        pthread_t tid;
        int result;

        result = pthread_create(&tid, NULL, repeatedly_dump, NULL);
        assert(!result);
        for (;;) {
                usleep(100000);
                switch (rand() % 10) {
                case 0:
                        a();
                        break;
                case 1:
                        b();
                        break;
                case 2:
                        c();
                        break;
                case 3:
                        d();
                        break;
                case 4:
                        e();
                        break;
                case 5:
                        f();
                        break;
                case 6:
                        g();
                        break;
                case 7:
                        h();
                        break;
                case 8:
                        j();
                        break;
                case 9:
                        k();
                        break;
                }
        }
}