Allocation profiler

NEWS.md

@@ -135,6 +135,9 @@ Standard library changes
   Further, percent utilization is now reported as a total or per-thread, based on whether the thread is idle or not at
   each sample. `Profile.fetch()` by default strips out the new metadata to ensure backwards compatibility with external
   profiling data consumers, but can be included with the `include_meta` kwarg. ([#41742])
+* The new `Profile.Allocs` module allows memory allocations to be profiled. The stack trace, type, and size of each
+  allocation is recorded, and a `sample_rate` argument allows a tunable amount of allocations to be skipped,
+  reducing performance overhead. ([#42768])
 
 #### Random
 

doc/src/manual/profile.md

@@ -336,6 +336,31 @@ and how much garbage it collects each time. This can be enabled with
 [`GC.enable_logging(true)`](@ref), which causes Julia to log to stderr every time
 a garbage collection happens.
 
+### Allocation Profiler
+
+The allocation profiler records the stack trace, type, and size of each
+allocation while it is running. It can be invoked with
+[`Profile.Allocs.@profile`](@ref).
+
+This information about the allocations is returned as an array of `Alloc`
+objects, wrapped in an `AllocResults` object. The best way to visualize
+these is currently with the [PProf.jl](https://github.com/JuliaPerf/PProf.jl)
+library, which can visualize the call stacks which are making the most
+allocations.
+
+The allocation profiler does have significant overhead, so a `sample_rate`
+argument can be passed to speed it up by making it skip some allocations.
+Passing `sample_rate=1.0` will make it record everything (which is slow);
+`sample_rate=0.1` will record only 10% of the allocations (faster), etc.
+
+!!! note
+
+    The current implementation of the Allocations Profiler _does not
+    capture all allocations._ You can read more about the missing allocations
+    and the plan to improve this, here: https://github.com/JuliaLang/julia/issues/43688.
+    Calling `Profile.Allocs.fetch()` will print a log line reporting the percentage
+    of missed allocations, so you can understand the accuracy of your profile.
+
 ## External Profiling
 
 Currently Julia supports `Intel VTune`, `OProfile` and `perf` as external profiling tools.

src/Makefile

@@ -45,7 +45,7 @@ RUNTIME_SRCS := \
 	jltypes gf typemap smallintset ast builtins module interpreter symbol \
 	dlload sys init task array dump staticdata toplevel jl_uv datatype \
 	simplevector runtime_intrinsics precompile \
-	threading partr stackwalk gc gc-debug gc-pages gc-stacks method \
+	threading partr stackwalk gc gc-debug gc-pages gc-stacks gc-alloc-profiler method \
 	jlapi signal-handling safepoint timing subtype \
 	crc32c APInt-C processor ircode opaque_closure codegen-stubs coverage
 SRCS := jloptions runtime_ccall rtutils
@@ -288,7 +288,7 @@ $(BUILDDIR)/disasm.o $(BUILDDIR)/disasm.dbg.obj: $(SRCDIR)/debuginfo.h $(SRCDIR)
 $(BUILDDIR)/dump.o $(BUILDDIR)/dump.dbg.obj: $(addprefix $(SRCDIR)/,common_symbols1.inc common_symbols2.inc builtin_proto.h serialize.h)
 $(BUILDDIR)/gc-debug.o $(BUILDDIR)/gc-debug.dbg.obj: $(SRCDIR)/gc.h
 $(BUILDDIR)/gc-pages.o $(BUILDDIR)/gc-pages.dbg.obj: $(SRCDIR)/gc.h
-$(BUILDDIR)/gc.o $(BUILDDIR)/gc.dbg.obj: $(SRCDIR)/gc.h
+$(BUILDDIR)/gc.o $(BUILDDIR)/gc.dbg.obj: $(SRCDIR)/gc.h $(SRCDIR)/gc-alloc-profiler.h
 $(BUILDDIR)/init.o $(BUILDDIR)/init.dbg.obj: $(SRCDIR)/builtin_proto.h
 $(BUILDDIR)/interpreter.o $(BUILDDIR)/interpreter.dbg.obj: $(SRCDIR)/builtin_proto.h
 $(BUILDDIR)/jitlayers.o $(BUILDDIR)/jitlayers.dbg.obj: $(SRCDIR)/jitlayers.h $(SRCDIR)/codegen_shared.h

src/array.c

@@ -508,6 +508,7 @@ JL_DLLEXPORT jl_value_t *jl_alloc_string(size_t len)
         s = jl_gc_big_alloc(ptls, allocsz);
     }
     jl_set_typeof(s, jl_string_type);
+    maybe_record_alloc_to_profile(s, len);
     *(size_t*)s = len;
     jl_string_data(s)[len] = 0;
     return s;

src/gc-alloc-profiler.cpp

@@ -0,0 +1,139 @@
+// This file is a part of Julia. License is MIT: https://julialang.org/license
+
+#include "gc-alloc-profiler.h"
+
+#include "julia_internal.h"
+#include "gc.h"
+
+#include <string>
+#include <vector>
+
+using std::string;
+using std::vector;
+
+struct jl_raw_backtrace_t {
+    jl_bt_element_t *data;
+    size_t size;
+};
+
+struct jl_raw_alloc_t {
+    jl_datatype_t *type_address;
+    jl_raw_backtrace_t backtrace;
+    size_t size;
+};
+
+// == These structs define the global singleton profile buffer that will be used by
+// callbacks to store profile results. ==
+struct jl_per_thread_alloc_profile_t {
+    vector<jl_raw_alloc_t> allocs;
+};
+
+struct jl_alloc_profile_t {
+    double sample_rate;
+
+    vector<jl_per_thread_alloc_profile_t> per_thread_profiles;
+};
+
+struct jl_combined_results {
+    vector<jl_raw_alloc_t> combined_allocs;
+};
+
+// == Global variables manipulated by callbacks ==
+
+jl_alloc_profile_t g_alloc_profile;
+int g_alloc_profile_enabled = false;
+jl_combined_results g_combined_results; // Will live forever.
+
+// === stack stuff ===
+
+jl_raw_backtrace_t get_raw_backtrace() {
+    // A single large buffer to record backtraces onto
+    static jl_bt_element_t static_bt_data[JL_MAX_BT_SIZE];
+
+    size_t bt_size = rec_backtrace(static_bt_data, JL_MAX_BT_SIZE, 2);
+
+    // Then we copy only the needed bytes out of the buffer into our profile.
+    size_t bt_bytes = bt_size * sizeof(jl_bt_element_t);
+    jl_bt_element_t *bt_data = (jl_bt_element_t*) malloc(bt_bytes);
+    memcpy(bt_data, static_bt_data, bt_bytes);
+
+    return jl_raw_backtrace_t{
+        bt_data,
+        bt_size
+    };
+}
+
+// == exported interface ==
+
+extern "C" {  // Needed since these functions doesn't take any arguments.
+
+JL_DLLEXPORT void jl_start_alloc_profile(double sample_rate) {
+    // We only need to do this once, the first time this is called.
+    while (g_alloc_profile.per_thread_profiles.size() < jl_n_threads) {
+        g_alloc_profile.per_thread_profiles.push_back(jl_per_thread_alloc_profile_t{});
+    }
+
+    g_alloc_profile.sample_rate = sample_rate;
+    g_alloc_profile_enabled = true;
+}
+
+JL_DLLEXPORT jl_profile_allocs_raw_results_t jl_fetch_alloc_profile() {
+    // combine allocs
+    // TODO: interleave to preserve ordering
+    for (auto& profile : g_alloc_profile.per_thread_profiles) {
+        for (const auto& alloc : profile.allocs) {
+            g_combined_results.combined_allocs.push_back(alloc);
+        }
+
+        profile.allocs.clear();
+    }
+
+    return jl_profile_allocs_raw_results_t{
+        g_combined_results.combined_allocs.data(),
+        g_combined_results.combined_allocs.size(),
+    };
+}
+
+JL_DLLEXPORT void jl_stop_alloc_profile() {
+    g_alloc_profile_enabled = false;
+}
+
+JL_DLLEXPORT void jl_free_alloc_profile() {
+    // Free any allocs that remain in the per-thread profiles, that haven't
+    // been combined yet (which happens in fetch_alloc_profiles()).
+    for (auto& profile : g_alloc_profile.per_thread_profiles) {
+        for (auto alloc : profile.allocs) {
+            free(alloc.backtrace.data);
+        }
+        profile.allocs.clear();
+    }
+
+    // Free the allocs that have been already combined into the combined results object.
+    for (auto alloc : g_combined_results.combined_allocs) {
+        free(alloc.backtrace.data);
+    }
+
+    g_combined_results.combined_allocs.clear();
+}
+
+// == callback called into by the outside ==
+
+void _maybe_record_alloc_to_profile(jl_value_t *val, size_t size) JL_NOTSAFEPOINT {
+    auto& global_profile = g_alloc_profile;
+    auto& profile = global_profile.per_thread_profiles[jl_threadid()];
+
+    auto sample_val = double(rand()) / double(RAND_MAX);
+    auto should_record = sample_val <= global_profile.sample_rate;
+    if (!should_record) {
+        return;
+    }
+
+    auto type = (jl_datatype_t*)jl_typeof(val);
+    profile.allocs.emplace_back(jl_raw_alloc_t{
+        type,
+        get_raw_backtrace(),
+        size
+    });
+}
+
+}  // extern "C"

src/gc-alloc-profiler.h

@@ -0,0 +1,49 @@
+// This file is a part of Julia. License is MIT: https://julialang.org/license
+
+#ifndef JL_GC_ALLOC_PROFILER_H
+#define JL_GC_ALLOC_PROFILER_H
+
+#include "julia.h"
+#include "ios.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// ---------------------------------------------------------------------
+// The public interface to call from Julia for allocations profiling
+// ---------------------------------------------------------------------
+
+// Forward-declaration to avoid depenency in header file.
+struct jl_raw_alloc_t;  // Defined in gc-alloc-profiler.cpp
+
+typedef struct {
+    struct jl_raw_alloc_t *allocs;
+    size_t num_allocs;
+} jl_profile_allocs_raw_results_t;
+
+JL_DLLEXPORT void jl_start_alloc_profile(double sample_rate);
+JL_DLLEXPORT jl_profile_allocs_raw_results_t jl_fetch_alloc_profile(void);
+JL_DLLEXPORT void jl_stop_alloc_profile(void);
+JL_DLLEXPORT void jl_free_alloc_profile(void);
+
+// ---------------------------------------------------------------------
+// Functions to call from GC when alloc profiling is enabled
+// ---------------------------------------------------------------------
+
+void _maybe_record_alloc_to_profile(jl_value_t *val, size_t size) JL_NOTSAFEPOINT;
+
+extern int g_alloc_profile_enabled;
+
+static inline void maybe_record_alloc_to_profile(jl_value_t *val, size_t size) JL_NOTSAFEPOINT {
+    if (__unlikely(g_alloc_profile_enabled)) {
+        _maybe_record_alloc_to_profile(val, size);
+    }
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+
+#endif  // JL_GC_ALLOC_PROFILER_H

src/gc.h

@@ -24,6 +24,7 @@
 #endif
 #endif
 #include "julia_assert.h"
+#include "gc-alloc-profiler.h"
 
 #ifdef __cplusplus
 extern "C" {

src/julia_internal.h

@@ -9,6 +9,7 @@
 #include "support/hashing.h"
 #include "support/ptrhash.h"
 #include "support/strtod.h"
+#include "gc-alloc-profiler.h"
 #include <uv.h>
 #if !defined(_WIN32)
 #include <unistd.h>
@@ -364,6 +365,7 @@ STATIC_INLINE jl_value_t *jl_gc_alloc_(jl_ptls_t ptls, size_t sz, void *ty)
         v = jl_gc_big_alloc(ptls, allocsz);
     }
     jl_set_typeof(v, ty);
+    maybe_record_alloc_to_profile(v, sz);
     return v;
 }
 

stdlib/Profile/docs/src/index.md

@@ -1,5 +1,7 @@
 # [Profiling](@id lib-profiling)
 
+## CPU Profiling
+
 ```@docs
 Profile.@profile
 ```
@@ -15,3 +17,18 @@ Profile.retrieve
 Profile.callers
 Profile.clear_malloc_data
 ```
+
+## Memory profiling
+
+```@docs
+Profile.Allocs.@profile
+```
+
+The methods in `Profile.Allocs` are not exported and need to be called e.g. as `Profile.Allocs.fetch()`.
+
+```@docs
+Profile.Allocs.clear
+Profile.Allocs.fetch
+Profile.Allocs.start
+Profile.Allocs.stop
+```