This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] Add malloc micro benchmark
Hi DJ,
> Looks good to me, although I'd like some additional comments in the test
> code.
Thanks for the review - I've added some extra comments:
+/* Benchmark the malloc/free performance of a varying number of blocks of a
+ given size. This enables performance tracking of the t-cache and fastbins.
+ It tests 3 different scenarios: single-threaded using main arena,
+ multi-threaded using thread-arena, and main arena with SINGLE_THREAD_P
+ false. */
> + else \
> + for thr in 8 16 32 64 128 256 512 1024 2048 4096; do \
> + echo "Running $${run} $${thr}"; \
> + $(run-bench) $${thr} > $${run}-$${thr}.out; \
> + done;\
> + fi;\
> done
> I wonder if this could be done more elegantly, but I'm OK with a simple
> approach for now. If we end up adding many more such tests we might
> need to revisit this part.
The main concern was to get a clean state so that the test of a previous block
size doesn't affect subsequent results.
> +#define NUM_ITERS 1000000
> +#define NUM_ALLOCS 4
> +#define MAX_ALLOCS 1600
> How long does this test take to run, on average, compared to other
> tests? Do we have to worry about increasing timeouts for slow hosts?
All the tests together runs finish in a fraction of the time taken by a single
test of bench-malloc-thread, so if anything we need to reduce the time of
that one by an order of magnitude (it takes ~5 minutes!).
> +static void
> +do_benchmark (malloc_args *args, int **arr)
> +{
> + timing_t start, stop;
> + size_t iters = args->iters;
> + size_t size = args->size;
> + int n = args->n;
> +
> + TIMING_NOW (start);
> +
> + for (int j = 0; j < iters; j++)
> + {
> + for (int i = 0; i < n; i++)
> + arr[i] = malloc (size);
> +
> + for (int i = 0; i < n; i++)
> + free (arr[i]);
> + }
> +
> + TIMING_NOW (stop);
> +
> + TIMING_DIFF (args->elapsed, start, stop);
> +}
> Simple loop, but doesn't test for malloc returning NULL.
Yeah, the benchmark doesn't need to care since the amount we allocate
is tiny (6.4MBytes).
Cheers,
Wilco
I've committed this:
Add a malloc micro benchmark to enable accurate testing of the
various paths in malloc and free. The benchmark does a varying
number of allocations of a given block size, then frees them again.
It tests 3 different scenarios: single-threaded using main arena,
multi-threaded using thread-arena, main arena with SINGLE_THREAD_P
false.
OK for commit?
ChangeLog:
2019-02-14 Wilco Dijkstra <wdijkstr@arm.com>
* benchtests/Makefile: Add malloc-simple benchmark.
* benchtests/bench-malloc-simple.c: New benchmark.
--
diff --git a/benchtests/Makefile b/benchtests/Makefile
index 12036b1935dc7ea84b421f024d6fe3190ae35a6e..09f7cb8e475a312268eebb4d346edde70d22bb3d 100644
--- a/benchtests/Makefile
+++ b/benchtests/Makefile
@@ -90,7 +90,7 @@ CFLAGS-bench-trunc.c += -fno-builtin
CFLAGS-bench-truncf.c += -fno-builtin
ifeq (${BENCHSET},)
-bench-malloc := malloc-thread
+bench-malloc := malloc-thread malloc-simple
else
bench-malloc := $(filter malloc-%,${BENCHSET})
endif
@@ -98,7 +98,7 @@ endif
$(addprefix $(objpfx)bench-,$(bench-math)): $(libm)
$(addprefix $(objpfx)bench-,$(math-benchset)): $(libm)
$(addprefix $(objpfx)bench-,$(bench-pthread)): $(shared-thread-library)
-$(objpfx)bench-malloc-thread: $(shared-thread-library)
+$(addprefix $(objpfx)bench-,$(bench-malloc)): $(shared-thread-library)
@@ -165,7 +165,7 @@ bench-clean:
ifneq ($(strip ${BENCHSET}),)
VALIDBENCHSETNAMES := bench-pthread bench-math bench-string string-benchset \
wcsmbs-benchset stdlib-benchset stdio-common-benchset math-benchset \
- malloc-thread
+ malloc-thread malloc-simple
INVALIDBENCHSETNAMES := $(filter-out ${VALIDBENCHSETNAMES},${BENCHSET})
ifneq (${INVALIDBENCHSETNAMES},)
$(info The following values in BENCHSET are invalid: ${INVALIDBENCHSETNAMES})
@@ -194,10 +194,18 @@ bench-set: $(binaries-benchset)
bench-malloc: $(binaries-bench-malloc)
for run in $^; do \
+ echo "$${run}"; \
+ if [ `basename $${run}` = "bench-malloc-thread" ]; then \
for thr in 1 8 16 32; do \
echo "Running $${run} $${thr}"; \
- $(run-bench) $${thr} > $${run}-$${thr}.out; \
- done;\
+ $(run-bench) $${thr} > $${run}-$${thr}.out; \
+ done;\
+ else \
+ for thr in 8 16 32 64 128 256 512 1024 2048 4096; do \
+ echo "Running $${run} $${thr}"; \
+ $(run-bench) $${thr} > $${run}-$${thr}.out; \
+ done;\
+ fi;\
done
# Build and execute the benchmark functions. This target generates JSON
diff --git a/benchtests/bench-malloc-simple.c b/benchtests/bench-malloc-simple.c
new file mode 100644
index 0000000000000000000000000000000000000000..83203ff3187654a1710c9ef81016f854957b9d64
--- /dev/null
+++ b/benchtests/bench-malloc-simple.c
@@ -0,0 +1,188 @@
+/* Benchmark malloc and free functions.
+ Copyright (C) 2019 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <http://www.gnu.org/licenses/>. */
+
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <malloc.h>
+#include <sys/resource.h>
+#include "bench-timing.h"
+#include "json-lib.h"
+
+/* Benchmark the malloc/free performance of a varying number of blocks of a
+ given size. This enables performance tracking of the t-cache and fastbins.
+ It tests 3 different scenarios: single-threaded using main arena,
+ multi-threaded using thread-arena, and main arena with SINGLE_THREAD_P
+ false. */
+
+#define NUM_ITERS 200000
+#define NUM_ALLOCS 4
+#define MAX_ALLOCS 1600
+
+typedef struct
+{
+ size_t iters;
+ size_t size;
+ int n;
+ timing_t elapsed;
+} malloc_args;
+
+static void
+do_benchmark (malloc_args *args, int **arr)
+{
+ timing_t start, stop;
+ size_t iters = args->iters;
+ size_t size = args->size;
+ int n = args->n;
+
+ TIMING_NOW (start);
+
+ for (int j = 0; j < iters; j++)
+ {
+ for (int i = 0; i < n; i++)
+ arr[i] = malloc (size);
+
+ for (int i = 0; i < n; i++)
+ free (arr[i]);
+ }
+
+ TIMING_NOW (stop);
+
+ TIMING_DIFF (args->elapsed, start, stop);
+}
+
+static malloc_args tests[3][NUM_ALLOCS];
+static int allocs[NUM_ALLOCS] = { 25, 100, 400, MAX_ALLOCS };
+
+static void *
+thread_test (void *p)
+{
+ int **arr = (int**)p;
+
+ /* Run benchmark multi-threaded. */
+ for (int i = 0; i < NUM_ALLOCS; i++)
+ do_benchmark (&tests[2][i], arr);
+
+ return p;
+}
+
+void
+bench (unsigned long size)
+{
+ size_t iters = NUM_ITERS;
+ int **arr = (int**) malloc (MAX_ALLOCS * sizeof (void*));
+ unsigned long res;
+
+ TIMING_INIT (res);
+
+ for (int t = 0; t <= 3; t++)
+ for (int i = 0; i < NUM_ALLOCS; i++)
+ {
+ tests[t][i].n = allocs[i];
+ tests[t][i].size = size;
+ tests[t][i].iters = iters / allocs[i];
+
+ /* Do a quick warmup run. */
+ if (t == 0)
+ do_benchmark (&tests[0][i], arr);
+ }
+
+ /* Run benchmark single threaded in main_arena. */
+ for (int i = 0; i < NUM_ALLOCS; i++)
+ do_benchmark (&tests[0][i], arr);
+
+ /* Run benchmark in a thread_arena. */
+ pthread_t t;
+ pthread_create (&t, NULL, thread_test, (void*)arr);
+ pthread_join (t, NULL);
+
+ /* Repeat benchmark in main_arena with SINGLE_THREAD_P == false. */
+ for (int i = 0; i < NUM_ALLOCS; i++)
+ do_benchmark (&tests[1][i], arr);
+
+ free (arr);
+
+ json_ctx_t json_ctx;
+
+ json_init (&json_ctx, 0, stdout);
+
+ json_document_begin (&json_ctx);
+
+ json_attr_string (&json_ctx, "timing_type", TIMING_TYPE);
+
+ json_attr_object_begin (&json_ctx, "functions");
+
+ json_attr_object_begin (&json_ctx, "malloc");
+
+ char s[100];
+ double iters2 = iters;
+
+ json_attr_object_begin (&json_ctx, "");
+ json_attr_double (&json_ctx, "malloc_block_size", size);
+
+ struct rusage usage;
+ getrusage (RUSAGE_SELF, &usage);
+ json_attr_double (&json_ctx, "max_rss", usage.ru_maxrss);
+
+ for (int i = 0; i < NUM_ALLOCS; i++)
+ {
+ sprintf (s, "main_arena_st_allocs_%04d_time", allocs[i]);
+ json_attr_double (&json_ctx, s, tests[0][i].elapsed / iters2);
+ }
+
+ for (int i = 0; i < NUM_ALLOCS; i++)
+ {
+ sprintf (s, "main_arena_mt_allocs_%04d_time", allocs[i]);
+ json_attr_double (&json_ctx, s, tests[1][i].elapsed / iters2);
+ }
+
+ for (int i = 0; i < NUM_ALLOCS; i++)
+ {
+ sprintf (s, "thread_arena__allocs_%04d_time", allocs[i]);
+ json_attr_double (&json_ctx, s, tests[2][i].elapsed / iters2);
+ }
+
+ json_attr_object_end (&json_ctx);
+
+ json_attr_object_end (&json_ctx);
+
+ json_attr_object_end (&json_ctx);
+
+ json_document_end (&json_ctx);
+}
+
+static void usage (const char *name)
+{
+ fprintf (stderr, "%s: <alloc_size>\n", name);
+ exit (1);
+}
+
+int
+main (int argc, char **argv)
+{
+ long val = 16;
+ if (argc == 2)
+ val = strtol (argv[1], NULL, 0);
+
+ if (argc > 2 || val <= 0)
+ usage (argv[0]);
+
+ bench (val);
+
+ return 0;
+}