Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions Makefile.am
Original file line number Diff line number Diff line change
@@ -1,5 +1,11 @@

EXTRA_DIST = README.html
EXTRA_DIST += \
benchmarks/README.md \
benchmarks/benchmark.c \
benchmarks/run.sh \
benchmarks/runner.py \
benchmarks/selftest.py

SUBDIRS = . tests

Expand Down
3 changes: 3 additions & 0 deletions benchmarks/.gitignore
Original file line number Diff line number Diff line change
@@ -0,0 +1,3 @@
artifacts/
raw/
__pycache__/
64 changes: 64 additions & 0 deletions benchmarks/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,64 @@
# libfastjson performance benchmarks

This suite isolates the libfastjson operations that dominate rsyslog's `$!`
variable path. It measures case-insensitive object lookup by width and key
position, replacement writes, string creation around inline-storage
boundaries, and narrow-object construction/destruction both with normal
duplicate checking and with the existing `FJSON_OBJECT_ADD_KEY_IS_NEW` flag.

Every workload uses a correctness checksum and structural oracle. The runner
performs one calibration, then 11 measured trials, records resumable raw JSON,
and reports medians, median absolute deviation, and outliers. Paired mode runs
baseline and candidate builds in alternating order.

Build libfastjson first:

```sh
CFLAGS='-O2 -g -fno-omit-frame-pointer' ./autogen.sh --configure
make -j"$(nproc)"
```

Benchmark build directories must be clean Git checkouts. The runner rejects
uncommitted library sources so recorded measurements can be reproduced.

Record a characterization session:

```sh
benchmarks/run.sh \
--build-dir "$PWD" \
--label current-main-a \
--output benchmarks/raw/current-main-a.json \
--summary-json benchmarks/results/current-main-a.json \
--summary-markdown benchmarks/results/current-main-a.md
```

Use `--smoke` for a reduced matrix. Filters `--operation`, `--width`,
`--position`, and `--value-bytes` are repeatable.

For candidate acceptance, use paired mode with separate worktrees/builds:

```sh
benchmarks/run.sh \
--build-dir /path/to/baseline \
--label baseline \
--output benchmarks/raw/baseline.json \
--pair-build-dir /path/to/candidate \
--pair-label candidate \
--pair-output benchmarks/raw/candidate.json \
--comparison-json benchmarks/results/comparison.json \
--comparison-markdown benchmarks/results/comparison.md
```

A runtime candidate should be retained only when two independent paired
sessions each improve its targeted workload by at least 10%, no core workload
regresses by more than 5%, and relative MAD is at most 0.05 after one rerun.

Run deterministic harness tests with:

```sh
python3 benchmarks/selftest.py
```

Tracked current-main characterization and direct-lookup candidate results are
under `benchmarks/results/`. Raw trial data remain ignored under
`benchmarks/raw/`.
269 changes: 269 additions & 0 deletions benchmarks/benchmark.c
Original file line number Diff line number Diff line change
@@ -0,0 +1,269 @@
/*
* Microbenchmark for libfastjson operations used heavily by rsyslog.
*
* Setup is outside the timed interval for lookup and replacement workloads.
* Every workload verifies its final value, member count, and accumulated
* result so an optimizing compiler or a broken library cannot silently turn
* the timed loop into a no-op.
*/
#define _POSIX_C_SOURCE 200809L

#include <errno.h>
#include <inttypes.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>

#include "json.h"

static volatile uint64_t result_sink;

static void
fail(const char *const message)
{
fprintf(stderr, "benchmark error: %s\n", message);
exit(1);
}

static uint64_t
parse_u64(const char *const text, const char *const name)
{
char *end = NULL;
errno = 0;
const unsigned long long value = strtoull(text, &end, 10);
if (errno != 0 || end == text || *end != '\0') {
fprintf(stderr, "benchmark error: invalid %s: %s\n", name, text);
exit(2);
}
return (uint64_t) value;
}

static uint64_t
now_ns(void)
{
struct timespec now;
if (clock_gettime(CLOCK_MONOTONIC_RAW, &now) != 0)
fail("clock_gettime failed");
return (uint64_t) now.tv_sec * UINT64_C(1000000000) + (uint64_t) now.tv_nsec;
}

static char **
make_keys(const size_t width)
{
char **const keys = calloc(width == 0 ? 1 : width, sizeof(*keys));
if (keys == NULL)
fail("key table allocation failed");
for (size_t i = 0 ; i < width ; ++i) {
keys[i] = malloc(32);
if (keys[i] == NULL)
fail("key allocation failed");
snprintf(keys[i], 32, "key-%03zu", i);
}
return keys;
}

static void
free_keys(char **const keys, const size_t width)
{
for (size_t i = 0 ; i < width ; ++i)
free(keys[i]);
free(keys);
}

static size_t
position_index(const char *const position, const size_t width)
{
if (!strcmp(position, "first"))
return 0;
if (!strcmp(position, "middle"))
return width / 2;
if (!strcmp(position, "last"))
return width - 1;
if (!strcmp(position, "miss"))
return width;
fail("unknown position");
return 0;
}

static struct fjson_object *
make_object(char **const keys, const size_t width, const int assume_new)
{
struct fjson_object *const object = fjson_object_new_object();
if (object == NULL)
fail("object allocation failed");
for (size_t i = 0 ; i < width ; ++i) {
struct fjson_object *const value = fjson_object_new_int64((int64_t) i);
if (value == NULL)
fail("integer allocation failed");
if (assume_new)
fjson_object_object_add_ex(object, keys[i], value, FJSON_OBJECT_ADD_KEY_IS_NEW);
else
fjson_object_object_add(object, keys[i], value);
}
if ((size_t) fjson_object_object_length(object) != width)
fail("object setup member count mismatch");
return object;
}

static uint64_t
run_lookup(const size_t width, const char *const position, const uint64_t iterations)
{
if (width == 0)
fail("lookup width must be positive");
char **const keys = make_keys(width + 1);
struct fjson_object *const object = make_object(keys, width, 0);
const size_t index = position_index(position, width);
uint64_t checksum = 0;
const uint64_t start = now_ns();
for (uint64_t i = 0 ; i < iterations ; ++i) {
struct fjson_object *value = NULL;
const fjson_bool found = fjson_object_object_get_ex(object, keys[index], &value);
if (index == width) {
checksum += found ? UINT64_C(1000003) : 1;
} else {
if (!found || value == NULL)
fail("lookup unexpectedly missed");
checksum += (uint64_t) fjson_object_get_int64(value) + 1;
}
}
const uint64_t elapsed = now_ns() - start;
const uint64_t expected = iterations * (index == width ? 1 : index + 1);
if (checksum != expected)
fail("lookup checksum mismatch");
result_sink = checksum;
fjson_object_put(object);
free_keys(keys, width + 1);
return elapsed;
}

static uint64_t
run_replace(const size_t width, const char *const position, const uint64_t iterations)
{
if (width == 0 || !strcmp(position, "miss"))
fail("replacement requires a present key");
char **const keys = make_keys(width);
struct fjson_object *const object = make_object(keys, width, 0);
const size_t index = position_index(position, width);
uint64_t checksum = 0;
const uint64_t start = now_ns();
for (uint64_t i = 0 ; i < iterations ; ++i) {
struct fjson_object *const value = fjson_object_new_int64((int64_t) i);
if (value == NULL)
fail("replacement value allocation failed");
fjson_object_object_add(object, keys[index], value);
checksum += i & 1;
}
const uint64_t elapsed = now_ns() - start;
struct fjson_object *final_value = NULL;
if (!fjson_object_object_get_ex(object, keys[index], &final_value) || final_value == NULL)
fail("replacement final lookup failed");
if (fjson_object_get_int64(final_value) != (int64_t) (iterations - 1))
fail("replacement final value mismatch");
if ((size_t) fjson_object_object_length(object) != width)
fail("replacement changed member count");
if (checksum != iterations / 2)
fail("replacement checksum mismatch");
result_sink = checksum;
fjson_object_put(object);
free_keys(keys, width);
return elapsed;
}

static uint64_t
run_string(const size_t value_bytes, const uint64_t iterations)
{
if (value_bytes == SIZE_MAX)
fail("string value bytes too large");
char *const value = malloc(value_bytes + 1);
Comment thread
rgerhards marked this conversation as resolved.
if (value == NULL)
fail("string source allocation failed");
for (size_t i = 0 ; i < value_bytes ; ++i)
value[i] = (char) ('a' + (i % 23));
value[value_bytes] = '\0';
uint64_t checksum = 0;
const uint64_t start = now_ns();
for (uint64_t i = 0 ; i < iterations ; ++i) {
struct fjson_object *const object = fjson_object_new_string(value);
if (object == NULL)
fail("string object allocation failed");
if ((size_t) fjson_object_get_string_len(object) != value_bytes)
fail("string length mismatch");
const char *const copy = fjson_object_get_string(object);
if (copy == NULL || (value_bytes > 0 && copy[value_bytes - 1] != value[value_bytes - 1]))
fail("string content mismatch");
checksum += value_bytes + (value_bytes == 0 ? 0 : (unsigned char) copy[0]);
fjson_object_put(object);
}
const uint64_t elapsed = now_ns() - start;
const uint64_t expected = iterations * (value_bytes + (value_bytes == 0 ? 0 : (unsigned char) value[0]));
if (checksum != expected)
fail("string checksum mismatch");
result_sink = checksum;
free(value);
return elapsed;
}

static uint64_t
run_object(const size_t width, const uint64_t iterations, const int assume_new)
{
char **const keys = make_keys(width);
uint64_t checksum = 0;
const uint64_t start = now_ns();
for (uint64_t i = 0 ; i < iterations ; ++i) {
struct fjson_object *const object = make_object(keys, width, assume_new);
checksum += (uint64_t) fjson_object_object_length(object) + 1;
fjson_object_put(object);
}
const uint64_t elapsed = now_ns() - start;
if (checksum != iterations * (width + 1))
fail("object lifecycle checksum mismatch");
result_sink = checksum;
free_keys(keys, width);
return elapsed;
}

int
main(const int argc, char **const argv)
{
if (argc != 7) {
fprintf(stderr,
"usage: %s OPERATION WIDTH POSITION VALUE_BYTES ITERATIONS CASE_SENSITIVE\n",
argv[0]);
return 2;
}
const char *const operation = argv[1];
const size_t width = (size_t) parse_u64(argv[2], "width");
const char *const position = argv[3];
const size_t value_bytes = (size_t) parse_u64(argv[4], "value bytes");
const uint64_t iterations = parse_u64(argv[5], "iterations");
const uint64_t case_mode = parse_u64(argv[6], "case mode");
if (iterations == 0 || case_mode > 1)
fail("iterations and case mode are out of range");
const int case_sensitive = (int) case_mode;
fjson_global_do_case_sensitive_comparison(case_sensitive);

uint64_t elapsed;
if (!strcmp(operation, "lookup"))
elapsed = run_lookup(width, position, iterations);
else if (!strcmp(operation, "replace"))
elapsed = run_replace(width, position, iterations);
else if (!strcmp(operation, "string"))
elapsed = run_string(value_bytes, iterations);
else if (!strcmp(operation, "object"))
elapsed = run_object(width, iterations, 0);
else if (!strcmp(operation, "object-new"))
elapsed = run_object(width, iterations, 1);
else
fail("unknown operation");

printf("{\"operation\":\"%s\",\"width\":%zu,\"position\":\"%s\","
"\"value_bytes\":%zu,\"iterations\":%" PRIu64 ",\"case_sensitive\":%d,"
"\"elapsed_ns\":%" PRIu64 ",\"ns_per_operation\":%.9f,"
"\"operations_per_second\":%.6f,\"oracle\":true}\n",
operation, width, position, value_bytes, iterations, case_sensitive, elapsed,
(double) elapsed / (double) iterations,
(double) iterations * 1000000000.0 / (double) elapsed);
return result_sink == UINT64_MAX ? 3 : 0;
}
Loading
Loading