project-base/server-skynet-source-3rd-jemalloc
2
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
a62b7ed928
server-skynet-source-3rd-je.../test/unit/cache_bin.c

205 lines
7.0 KiB
C
Raw Normal View History

Redesign the cache bin metadata for fast path. Implement the pointer-based metadata for tcache bins -- - 3 pointers are maintained to represent each bin; - 2 of the pointers are compressed on 64-bit; - is_full / is_empty done through pointer comparison; Comparing to the previous counter based design -- - fast-path speed up ~15% in benchmarks - direct pointer comparison and de-reference - no need to access tcache_bin_info in common case
2019-08-10 13:12:47 +08:00
#include "test/jemalloc_test.h"
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
static void
do_fill_test(cache_bin_t *bin, cache_bin_info_t *info, void **ptrs,
cache_bin_sz_t ncached_max, cache_bin_sz_t nfill_attempt,
cache_bin_sz_t nfill_succeed) {
bool success;
void *ptr;
assert_true(cache_bin_ncached_get(bin, info) == 0, "");
CACHE_BIN_PTR_ARRAY_DECLARE(arr, nfill_attempt);
cache_bin_init_ptr_array_for_fill(bin, info, &arr, nfill_attempt);
for (cache_bin_sz_t i = 0; i < nfill_succeed; i++) {
arr.ptr[i] = &ptrs[i];
}
cache_bin_finish_fill(bin, info, &arr, nfill_succeed);
expect_true(cache_bin_ncached_get(bin, info) == nfill_succeed, "");
cache_bin_low_water_set(bin);
Redesign the cache bin metadata for fast path. Implement the pointer-based metadata for tcache bins -- - 3 pointers are maintained to represent each bin; - 2 of the pointers are compressed on 64-bit; - is_full / is_empty done through pointer comparison; Comparing to the previous counter based design -- - fast-path speed up ~15% in benchmarks - direct pointer comparison and de-reference - no need to access tcache_bin_info in common case
2019-08-10 13:12:47 +08:00
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
for (cache_bin_sz_t i = 0; i < nfill_succeed; i++) {
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
ptr = cache_bin_alloc(bin, &success);
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
expect_true(success, "");
expect_ptr_eq(ptr, (void *)&ptrs[i],
"Should pop in order filled");
expect_true(cache_bin_low_water_get(bin, info)
== nfill_succeed - i - 1, "");
}
expect_true(cache_bin_ncached_get(bin, info) == 0, "");
expect_true(cache_bin_low_water_get(bin, info) == 0, "");
}
Redesign the cache bin metadata for fast path. Implement the pointer-based metadata for tcache bins -- - 3 pointers are maintained to represent each bin; - 2 of the pointers are compressed on 64-bit; - is_full / is_empty done through pointer comparison; Comparing to the previous counter based design -- - fast-path speed up ~15% in benchmarks - direct pointer comparison and de-reference - no need to access tcache_bin_info in common case
2019-08-10 13:12:47 +08:00
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
static void
do_flush_test(cache_bin_t *bin, cache_bin_info_t *info, void **ptrs,
cache_bin_sz_t nfill, cache_bin_sz_t nflush) {
Redesign the cache bin metadata for fast path. Implement the pointer-based metadata for tcache bins -- - 3 pointers are maintained to represent each bin; - 2 of the pointers are compressed on 64-bit; - is_full / is_empty done through pointer comparison; Comparing to the previous counter based design -- - fast-path speed up ~15% in benchmarks - direct pointer comparison and de-reference - no need to access tcache_bin_info in common case
2019-08-10 13:12:47 +08:00
bool success;
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
assert_true(cache_bin_ncached_get(bin, info) == 0, "");
for (cache_bin_sz_t i = 0; i < nfill; i++) {
success = cache_bin_dalloc_easy(bin, &ptrs[i]);
expect_true(success, "");
}
CACHE_BIN_PTR_ARRAY_DECLARE(arr, nflush);
cache_bin_init_ptr_array_for_flush(bin, info, &arr, nflush);
for (cache_bin_sz_t i = 0; i < nflush; i++) {
expect_ptr_eq(cache_bin_ptr_array_get(&arr, i), &ptrs[i], "");
Redesign the cache bin metadata for fast path. Implement the pointer-based metadata for tcache bins -- - 3 pointers are maintained to represent each bin; - 2 of the pointers are compressed on 64-bit; - is_full / is_empty done through pointer comparison; Comparing to the previous counter based design -- - fast-path speed up ~15% in benchmarks - direct pointer comparison and de-reference - no need to access tcache_bin_info in common case
2019-08-10 13:12:47 +08:00
}
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
cache_bin_finish_flush(bin, info, &arr, nflush);
expect_true(cache_bin_ncached_get(bin, info) == nfill - nflush, "");
while (cache_bin_ncached_get(bin, info) > 0) {
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
cache_bin_alloc(bin, &success);
Redesign the cache bin metadata for fast path. Implement the pointer-based metadata for tcache bins -- - 3 pointers are maintained to represent each bin; - 2 of the pointers are compressed on 64-bit; - is_full / is_empty done through pointer comparison; Comparing to the previous counter based design -- - fast-path speed up ~15% in benchmarks - direct pointer comparison and de-reference - no need to access tcache_bin_info in common case
2019-08-10 13:12:47 +08:00
}
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
}
TEST_BEGIN(test_cache_bin) {
bool success;
void *ptr;
cache_bin_t bin;
cache_bin_info_t info;
cache_bin_info_init(&info, TCACHE_NSLOTS_SMALL_MAX);
size_t size;
size_t alignment;
cache_bin_info_compute_alloc(&info, 1, &size, &alignment);
void *mem = mallocx(size, MALLOCX_ALIGN(alignment));
assert_ptr_not_null(mem, "Unexpected mallocx failure");
size_t cur_offset = 0;
cache_bin_preincrement(&info, 1, mem, &cur_offset);
cache_bin_init(&bin, &info, mem, &cur_offset);
cache_bin_postincrement(&info, 1, mem, &cur_offset);
assert_zu_eq(cur_offset, size, "Should use all requested memory");
/* Initialize to empty; should then have 0 elements. */
cache_bin_sz_t ncached_max = cache_bin_info_ncached_max(&info);
expect_true(cache_bin_ncached_get(&bin, &info) == 0, "");
expect_true(cache_bin_low_water_get(&bin, &info) == 0, "");
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
ptr = cache_bin_alloc_easy(&bin, &success);
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
expect_false(success, "Shouldn't successfully allocate when empty");
expect_ptr_null(ptr, "Shouldn't get a non-null pointer on failure");
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
ptr = cache_bin_alloc(&bin, &success);
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
expect_false(success, "Shouldn't successfully allocate when empty");
expect_ptr_null(ptr, "Shouldn't get a non-null pointer on failure");
/*
* We allocate one more item than ncached_max, so we can test cache bin
* exhaustion.
*/
void **ptrs = mallocx(sizeof(void *) * (ncached_max + 1), 0);
assert_ptr_not_null(ptrs, "Unexpected mallocx failure");
for (cache_bin_sz_t i = 0; i < ncached_max; i++) {
expect_true(cache_bin_ncached_get(&bin, &info) == i, "");
success = cache_bin_dalloc_easy(&bin, &ptrs[i]);
expect_true(success,
"Should be able to dalloc into a non-full cache bin.");
expect_true(cache_bin_low_water_get(&bin, &info) == 0,
"Pushes and pops shouldn't change low water of zero.");
}
expect_true(cache_bin_ncached_get(&bin, &info) == ncached_max, "");
success = cache_bin_dalloc_easy(&bin, &ptrs[ncached_max]);
expect_false(success, "Shouldn't be able to dalloc into a full bin.");
cache_bin_low_water_set(&bin);
for (cache_bin_sz_t i = 0; i < ncached_max; i++) {
expect_true(cache_bin_low_water_get(&bin, &info)
== ncached_max - i, "");
expect_true(cache_bin_ncached_get(&bin, &info)
== ncached_max - i, "");
/*
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
* This should fail -- the easy variant can't change the low
* water mark.
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
*/
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
ptr = cache_bin_alloc_easy(&bin, &success);
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
expect_ptr_null(ptr, "");
expect_false(success, "");
expect_true(cache_bin_low_water_get(&bin, &info)
== ncached_max - i, "");
expect_true(cache_bin_ncached_get(&bin, &info)
== ncached_max - i, "");
/* This should succeed, though. */
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
ptr = cache_bin_alloc(&bin, &success);
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
expect_true(success, "");
expect_ptr_eq(ptr, &ptrs[ncached_max - i - 1],
"Alloc should pop in stack order");
expect_true(cache_bin_low_water_get(&bin, &info)
== ncached_max - i - 1, "");
expect_true(cache_bin_ncached_get(&bin, &info)
== ncached_max - i - 1, "");
}
/* Now we're empty -- all alloc attempts should fail. */
expect_true(cache_bin_ncached_get(&bin, &info) == 0, "");
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
ptr = cache_bin_alloc_easy(&bin, &success);
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
expect_ptr_null(ptr, "");
expect_false(success, "");
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
ptr = cache_bin_alloc(&bin, &success);
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
expect_ptr_null(ptr, "");
expect_false(success, "");
for (cache_bin_sz_t i = 0; i < ncached_max / 2; i++) {
cache_bin_dalloc_easy(&bin, &ptrs[i]);
}
cache_bin_low_water_set(&bin);
for (cache_bin_sz_t i = ncached_max / 2; i < ncached_max; i++) {
cache_bin_dalloc_easy(&bin, &ptrs[i]);
}
expect_true(cache_bin_ncached_get(&bin, &info) == ncached_max, "");
for (cache_bin_sz_t i = ncached_max - 1; i >= ncached_max / 2; i--) {
/*
* Size is bigger than low water -- the reduced version should
* succeed.
*/
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
ptr = cache_bin_alloc_easy(&bin, &success);
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
expect_true(success, "");
expect_ptr_eq(ptr, &ptrs[i], "");
}
/* But now, we've hit low-water. */
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
ptr = cache_bin_alloc_easy(&bin, &success);
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
expect_false(success, "");
expect_ptr_null(ptr, "");
/* We're going to test filling -- we must be empty to start. */
while (cache_bin_ncached_get(&bin, &info)) {
Fast path: allow low-water mark changes. This lets us put more allocations on an "almost as fast" path after a flush. This results in around a 4% reduction in malloc cycles in prod workloads (corresponding to about a 0.1% reduction in overall cycles).
2020-03-05 00:58:42 +08:00
cache_bin_alloc(&bin, &success);
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
expect_true(success, "");
}
/* Test fill. */
/* Try to fill all, succeed fully. */
do_fill_test(&bin, &info, ptrs, ncached_max, ncached_max, ncached_max);
/* Try to fill all, succeed partially. */
do_fill_test(&bin, &info, ptrs, ncached_max, ncached_max,
ncached_max / 2);
/* Try to fill all, fail completely. */
do_fill_test(&bin, &info, ptrs, ncached_max, ncached_max, 0);
/* Try to fill some, succeed fully. */
do_fill_test(&bin, &info, ptrs, ncached_max, ncached_max / 2,
ncached_max / 2);
/* Try to fill some, succeed partially. */
do_fill_test(&bin, &info, ptrs, ncached_max, ncached_max / 2,
ncached_max / 2);
/* Try to fill some, fail completely. */
do_fill_test(&bin, &info, ptrs, ncached_max, ncached_max / 2, 0);
Redesign the cache bin metadata for fast path. Implement the pointer-based metadata for tcache bins -- - 3 pointers are maintained to represent each bin; - 2 of the pointers are compressed on 64-bit; - is_full / is_empty done through pointer comparison; Comparing to the previous counter based design -- - fast-path speed up ~15% in benchmarks - direct pointer comparison and de-reference - no need to access tcache_bin_info in common case
2019-08-10 13:12:47 +08:00
Cache bin: Write the unit test in terms of the API I.e. stop allowing the unit test to have secret access to implementation internals.
2020-03-03 06:14:08 +08:00
do_flush_test(&bin, &info, ptrs, ncached_max, ncached_max);
do_flush_test(&bin, &info, ptrs, ncached_max, ncached_max / 2);
do_flush_test(&bin, &info, ptrs, ncached_max, 0);
do_flush_test(&bin, &info, ptrs, ncached_max / 2, ncached_max / 2);
do_flush_test(&bin, &info, ptrs, ncached_max / 2, ncached_max / 4);
do_flush_test(&bin, &info, ptrs, ncached_max / 2, 0);
Redesign the cache bin metadata for fast path. Implement the pointer-based metadata for tcache bins -- - 3 pointers are maintained to represent each bin; - 2 of the pointers are compressed on 64-bit; - is_full / is_empty done through pointer comparison; Comparing to the previous counter based design -- - fast-path speed up ~15% in benchmarks - direct pointer comparison and de-reference - no need to access tcache_bin_info in common case
2019-08-10 13:12:47 +08:00
}
TEST_END
int
main(void) {
return test(test_cache_bin);
}
Reference in New Issue Copy Permalink