From 778b8c8ca708b082a1192acfb114a6751b2ad7c9 Mon Sep 17 00:00:00 2001 From: jdybnis Date: Mon, 9 Mar 2009 21:02:28 +0000 Subject: [PATCH] add perf test driver improve perf test add single threaded ("unsafe") skiplist for comparison minor refactoring --- include/common.h | 25 ++- makefile | 12 +- map/hashtable.c | 32 +-- map/list.c | 6 +- map/skiplist.c | 186 ++++++------------ map/unsafe_skiplist.c | 355 +++++++++++++++++++++++++++++++++ perf.sh | 10 + runtime/lwt.c | 2 +- runtime/mem.c | 24 ++- runtime/mem2.c | 410 +++++++++++++++++++++++++++++++++++++++ runtime/mem_class_calc.c | 134 +++++++++++++ runtime/rcu.c | 4 +- test/haz_test.c | 3 +- test/map_test1.c | 2 +- test/map_test2.c | 19 +- test/perf_test.c | 80 +++++--- test/rcu_test.c | 2 +- 17 files changed, 1093 insertions(+), 213 deletions(-) create mode 100644 map/unsafe_skiplist.c create mode 100755 perf.sh create mode 100644 runtime/mem2.c create mode 100644 runtime/mem_class_calc.c diff --git a/include/common.h b/include/common.h index d8811a9..fbd64bc 100644 --- a/include/common.h +++ b/include/common.h @@ -11,18 +11,31 @@ #include #include -#define MAX_NUM_THREADS 4 // make this whatever you want, but make it a power of 2 - #define CACHE_LINE_SIZE 64 // 64 byte cache line on x86 and x86-64 #define CACHE_LINE_SCALE 6 // log base 2 of the cache line size #define EXPECT_TRUE(x) __builtin_expect(!!(x), 1) #define EXPECT_FALSE(x) __builtin_expect(!!(x), 0) -#define SYNC_SWAP __sync_lock_test_and_set -#define SYNC_CAS __sync_val_compare_and_swap -#define SYNC_ADD __sync_add_and_fetch -#define SYNC_FETCH_AND_OR __sync_fetch_and_or +#ifndef NBD_SINGLE_THREADED + +#define MAX_NUM_THREADS 16 // make this whatever you want, but make it a power of 2 + +#define SYNC_SWAP(addr,x) __sync_lock_test_and_set(addr,x) +#define SYNC_CAS(addr,old,x) __sync_val_compare_and_swap(addr,old,x) +#define SYNC_ADD(addr,n) __sync_add_and_fetch(addr,n) +#define SYNC_FETCH_AND_OR(addr,x) __sync_fetch_and_or(addr,x) +#else// NBD_SINGLE_THREADED + +#define MAX_NUM_THREADS 1 + +#define SYNC_SWAP(addr,x) ({ typeof(*(addr)) _old = *(addr); *(addr) = (x); _old; }) +#define SYNC_CAS(addr,old,x) ({ typeof(*(addr)) _old = *(addr); *(addr) = (x); _old; }) +//#define SYNC_CAS(addr,old,x) ({ typeof(*(addr)) _old = *(addr); if ((old) == _old) { *(addr) = (x); } _old; }) +#define SYNC_ADD(addr,n) ({ typeof(*(addr)) _old = *(addr); *(addr) += (n); _old; }) +#define SYNC_FETCH_AND_OR(addr,x) ({ typeof(*(addr)) _old = *(addr); *(addr) |= (x); _old; }) + +#endif//NBD_SINGLE_THREADED #define COUNT_TRAILING_ZEROS __builtin_ctz diff --git a/makefile b/makefile index dbb0b8d..5ede683 100644 --- a/makefile +++ b/makefile @@ -5,17 +5,15 @@ # Makefile for building programs with whole-program interfile optimization ################################################################################################### CFLAGS0 := -Wall -Werror -std=gnu99 -lpthread #-m32 -DNBD32 -CFLAGS1 := $(CFLAGS0) -g #-O3 #-DNDEBUG #-fwhole-program -combine +CFLAGS1 := $(CFLAGS0) -g -O3 -DNDEBUG #-fwhole-program -combine CFLAGS2 := $(CFLAGS1) #-DENABLE_TRACE CFLAGS3 := $(CFLAGS2) #-DLIST_USE_HAZARD_POINTER -CFLAGS := $(CFLAGS3) #-DUSE_SYSTEM_MALLOC #-DTEST_STRING_KEYS +CFLAGS := $(CFLAGS3) #-DNBD_SINGLE_THREADED #-DUSE_SYSTEM_MALLOC #-DTEST_STRING_KEYS INCS := $(addprefix -I, include) -TESTS := output/perf_test output/map_test2 output/map_test1 output/txn_test \ - output/rcu_test output/haz_test +TESTS := output/perf_test output/map_test1 output/map_test2 output/rcu_test output/txn_test #output/haz_test OBJS := $(TESTS) -RUNTIME_SRCS := runtime/runtime.c runtime/rcu.c runtime/lwt.c runtime/mem.c datatype/nstring.c \ - runtime/hazard.c +RUNTIME_SRCS := runtime/runtime.c runtime/rcu.c runtime/lwt.c runtime/mem.c datatype/nstring.c #runtime/hazard.c MAP_SRCS := map/map.c map/list.c map/skiplist.c map/hashtable.c haz_test_SRCS := $(RUNTIME_SRCS) test/haz_test.c @@ -47,7 +45,7 @@ $(addsuffix .log, $(TESTS)) : %.log : % # in gcc. It chokes when -MM is used with -combine. ################################################################################################### $(OBJS): output/% : output/%.d makefile - gcc $(CFLAGS:-combine:) $(INCS) -MM -MT $@ $($*_SRCS) > $@.d + gcc $(CFLAGS) $(INCS) -MM -MT $@ $($*_SRCS) > $@.d gcc $(CFLAGS) $(INCS) -o $@ $($*_SRCS) asm: $(addsuffix .s, $(OBJS)) diff --git a/map/hashtable.c b/map/hashtable.c index bff782f..85ffe41 100644 --- a/map/hashtable.c +++ b/map/hashtable.c @@ -9,7 +9,7 @@ * Note: This is code uses synchronous atomic operations because that is all that x86 provides. * Every atomic operation is also an implicit full memory barrier. The upshot is that it simplifies * the code a bit, but it won't be as fast as it could be on platforms that provide weaker - * operations like and unfenced CAS which would still do the job. + * operations like unfenced CAS which would still do the job. * * 11FebO9 - Bug fix in ht_iter_next() from Rui Ueyama */ @@ -57,8 +57,8 @@ struct ht { const datatype_t *key_type; }; -static const map_val_t COPIED_VALUE = TAG_VALUE(DOES_NOT_EXIST, TAG1); -static const map_val_t TOMBSTONE = STRIP_TAG(-1, TAG1); +static const map_val_t COPIED_VALUE = TAG_VALUE(DOES_NOT_EXIST, TAG1); +static const map_val_t TOMBSTONE = STRIP_TAG(-1, TAG1); static const unsigned ENTRIES_PER_BUCKET = CACHE_LINE_SIZE/sizeof(entry_t); static const unsigned ENTRIES_PER_COPY_CHUNK = CACHE_LINE_SIZE/sizeof(entry_t)*2; @@ -143,7 +143,7 @@ static hti_t *hti_alloc (hashtable_t *parent, int scale) { memset(hti, 0, sizeof(hti_t)); hti->scale = scale; - size_t sz = sizeof(entry_t) * (1 << scale); + size_t sz = sizeof(entry_t) * (1ULL << scale); #ifdef USE_SYSTEM_MALLOC hti->unaligned_table_ptr = nbd_malloc(sz + CACHE_LINE_SIZE - 1); hti->table = (void *)(((size_t)hti->unaligned_table_ptr + CACHE_LINE_SIZE - 1) & ~(CACHE_LINE_SIZE - 1)); @@ -153,7 +153,7 @@ static hti_t *hti_alloc (hashtable_t *parent, int scale) { memset((void *)hti->table, 0, sz); // When searching for a key probe a maximum of 1/4 of the buckets up to 1000 buckets. - hti->max_probe = ((1 << (hti->scale - 2)) / ENTRIES_PER_BUCKET) + 4; + hti->max_probe = ((1ULL << (hti->scale - 2)) / ENTRIES_PER_BUCKET) + 4; if (hti->max_probe > MAX_BUCKETS_TO_PROBE) { hti->max_probe = MAX_BUCKETS_TO_PROBE; } @@ -177,8 +177,8 @@ static void hti_start_copy (hti_t *hti) { // heuristics to determine the size of the new table size_t count = ht_count(hti->ht); unsigned int new_scale = hti->scale; - new_scale += (count > (1 << (new_scale - 2))); // double size if more than 1/4 full - new_scale += (count > (1 << (new_scale - 2))); // double size again if more than 1/2 full + new_scale += (count > (1ULL << (new_scale - 2))); // double size if more than 1/4 full + new_scale += (count > (1ULL << (new_scale - 2))); // double size again if more than 1/2 full // Allocate the new table and attempt to install it. hti_t *next = hti_alloc(hti->ht, new_scale); @@ -205,7 +205,7 @@ static int hti_copy_entry (hti_t *ht1, volatile entry_t *ht1_ent, uint32_t key_h assert(ht1); assert(ht1->next); assert(ht2); - assert(ht1_ent >= ht1->table && ht1_ent < ht1->table + (1 << ht1->scale)); + assert(ht1_ent >= ht1->table && ht1_ent < ht1->table + (1ULL << ht1->scale)); #ifndef NBD32 assert(key_hash == 0 || ht1->ht->key_type == NULL || (key_hash >> 16) == (ht1_ent->key >> 48)); #endif @@ -473,9 +473,9 @@ static int hti_help_copy (hti_t *hti) { size_t x = hti->copy_scan; TRACE("h1", "ht_cas: help copy. scan is %llu, size is %llu", x, 1<scale); - if (total_copied != (1 << hti->scale)) { + if (total_copied != (1ULL << hti->scale)) { // Panic if we've been around the array twice and still haven't finished the copy. - int panic = (x >= (1 << (hti->scale + 1))); + int panic = (x >= (1ULL << (hti->scale + 1))); if (!panic) { limit = ENTRIES_PER_COPY_CHUNK; @@ -491,26 +491,26 @@ static int hti_help_copy (hti_t *hti) { TRACE("h1", "ht_cas: help copy panic", 0, 0); // scan the whole table ent = hti->table; - limit = (1 << hti->scale); + limit = (1ULL << hti->scale); } // Copy the entries for (int i = 0; i < limit; ++i) { num_copied += hti_copy_entry(hti, ent++, 0, hti->next); - assert(ent <= hti->table + (1 << hti->scale)); + assert(ent <= hti->table + (1ULL << hti->scale)); } if (num_copied != 0) { total_copied = SYNC_ADD(&hti->num_entries_copied, num_copied); } } - return (total_copied == (1 << hti->scale)); + return (total_copied == (1ULL << hti->scale)); } static void hti_defer_free (hti_t *hti) { assert(hti->ref_count == 0); - for (uint32_t i = 0; i < (1 << hti->scale); ++i) { + for (uint32_t i = 0; i < (1ULL << hti->scale); ++i) { map_key_t key = hti->table[i].key; map_val_t val = hti->table[i].val; if (val == COPIED_VALUE) @@ -628,7 +628,7 @@ void ht_print (hashtable_t *ht) { hti_t *hti = ht->hti; while (hti) { printf("hti:%p scale:%u count:%d copied:%d\n", hti, hti->scale, hti->count, hti->num_entries_copied); - for (int i = 0; i < (1 << hti->scale); ++i) { + for (int i = 0; i < (1ULL << hti->scale); ++i) { volatile entry_t *ent = hti->table + i; printf("[0x%x] 0x%llx:0x%llx\n", i, (uint64_t)ent->key, (uint64_t)ent->val); if (i > 30) { @@ -667,7 +667,7 @@ map_val_t ht_iter_next (ht_iter_t *iter, map_key_t *key_ptr) { volatile entry_t *ent; map_key_t key; map_val_t val; - size_t table_size = (1 << iter->hti->scale); + size_t table_size = (1ULL << iter->hti->scale); do { iter->idx++; if (iter->idx == table_size) { diff --git a/map/list.c b/map/list.c index 3b17873..992de52 100644 --- a/map/list.c +++ b/map/list.c @@ -118,7 +118,7 @@ static int find_pred (node_t **pred_ptr, node_t **item_ptr, list_t *ll, map_key_ // Unlink logically removed items. TRACE("l3", "find_pred: unlinking marked item %p next is %p", item, next); - markable_t other = SYNC_CAS(&pred->next, item, STRIP_MARK(next)); + markable_t other = SYNC_CAS(&pred->next, (markable_t)item, (markable_t)STRIP_MARK(next)); if (other == (markable_t)item) { TRACE("l2", "find_pred: unlinked item %p from pred %p", item, pred); item = STRIP_MARK(next); @@ -230,7 +230,7 @@ map_val_t ll_cas (list_t *ll, map_key_t key, map_val_t expectation, map_val_t ne map_key_t new_key = ll->key_type == NULL ? key : (map_key_t)ll->key_type->clone((void *)key); node_t *new_item = node_alloc(new_key, new_val); markable_t next = new_item->next = (markable_t)old_item; - markable_t other = SYNC_CAS(&pred->next, next, new_item); + markable_t other = SYNC_CAS(&pred->next, (markable_t)next, (markable_t)new_item); if (other == next) { TRACE("l1", "ll_cas: successfully inserted new item %p", new_item, 0); return DOES_NOT_EXIST; // success @@ -310,7 +310,7 @@ map_val_t ll_remove (list_t *ll, map_key_t key) { // item earlier, we logically removed it. TRACE("l2", "ll_remove: unlink the item by linking its pred %p to its successor %p", pred, next); markable_t other; - if ((other = SYNC_CAS(&pred->next, item, next)) != (markable_t)item) { + if ((other = SYNC_CAS(&pred->next, (markable_t)item, next)) != (markable_t)item) { TRACE("l1", "ll_remove: unlink failed; pred's link changed from %p to %p", item, other); return val; } diff --git a/map/skiplist.c b/map/skiplist.c index 70cb6e5..73bad7b 100644 --- a/map/skiplist.c +++ b/map/skiplist.c @@ -26,13 +26,19 @@ #include "mem.h" #include "rcu.h" -// Setting MAX_LEVEL to 0 essentially makes this data structure the Harris-Michael lock-free list (in list.c). -#define MAX_LEVEL 31 +// Setting MAX_LEVELS to 1 essentially makes this data structure the Harris-Michael lock-free list (see list.c). +#define MAX_LEVELS 32 + +enum unlink { + FORCE_UNLINK, + ASSIST_UNLINK, + DONT_UNLINK +}; typedef struct node { map_key_t key; map_val_t val; - int top_level; + unsigned num_levels; markable_t next[1]; } node_t; @@ -59,22 +65,22 @@ static inline node_t * STRIP_MARK(markable_t x) { return ((node_t *)STRIP_TAG(x, #define STRIP_MARK(x) ((node_t *)STRIP_TAG((x), 0x1)) #endif -static int random_level (void) { +static int random_levels (void) { unsigned r = nbd_rand(); - int n = __builtin_ctz(r) / 2; - if (n > MAX_LEVEL) { n = MAX_LEVEL; } + int n = __builtin_ctz(r) / 2 + 1; + if (n > MAX_LEVELS) { n = MAX_LEVELS; } return n; } -static node_t *node_alloc (int level, map_key_t key, map_val_t val) { - assert(level >= 0 && level <= MAX_LEVEL); - size_t sz = sizeof(node_t) + level * sizeof(node_t *); +static node_t *node_alloc (int num_levels, map_key_t key, map_val_t val) { + assert(num_levels >= 0 && num_levels <= MAX_LEVELS); + size_t sz = sizeof(node_t) + (num_levels - 1) * sizeof(node_t *); node_t *item = (node_t *)nbd_malloc(sz); memset(item, 0, sz); item->key = key; item->val = val; - item->top_level = level; - TRACE("s2", "node_alloc: new node %p (%llu levels)", item, level); + item->num_levels = num_levels; + TRACE("s2", "node_alloc: new node %p (%llu levels)", item, num_levels); return item; } @@ -82,8 +88,8 @@ skiplist_t *sl_alloc (const datatype_t *key_type) { skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t)); sl->key_type = key_type; sl->high_water = 0; - sl->head = node_alloc(MAX_LEVEL, 0, 0); - memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *)); + sl->head = node_alloc(MAX_LEVELS, 0, 0); + memset(sl->head->next, 0, MAX_LEVELS * sizeof(skiplist_t *)); return sl; } @@ -111,26 +117,22 @@ size_t sl_count (skiplist_t *sl) { return count; } -static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, map_key_t key, int help_remove) { +static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, map_key_t key, enum unlink unlink) { node_t *pred = sl->head; node_t *item = NULL; TRACE("s2", "find_preds: searching for key %p in skiplist (head is %p)", key, pred); int d = 0; - int start_level = sl->high_water; - if (EXPECT_FALSE(start_level < n)) { - start_level = n; - } // Traverse the levels of from the top level to the bottom - for (int level = start_level; level >= 0; --level) { + for (int level = sl->high_water - 1; level >= 0; --level) { markable_t next = pred->next[level]; - if (next == DOES_NOT_EXIST && level > n) + if (next == DOES_NOT_EXIST && level >= n) continue; TRACE("s3", "find_preds: traversing level %p starting at %p", level, pred); if (EXPECT_FALSE(HAS_MARK(next))) { TRACE("s2", "find_preds: pred %p is marked for removal (next %p); retry", pred, next); - ASSERT(level == pred->top_level || HAS_MARK(pred->next[level+1])); - return find_preds(preds, succs, n, sl, key, help_remove); // retry + ASSERT(level == pred->num_levels - 1 || HAS_MARK(pred->next[level+1])); + return find_preds(preds, succs, n, sl, key, unlink); // retry } item = GET_NODE(next); while (item != NULL) { @@ -139,7 +141,7 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl // A tag means an item is logically removed but not physically unlinked yet. while (EXPECT_FALSE(HAS_MARK(next))) { TRACE("s3", "find_preds: found marked item %p (next is %p)", item, next); - if (!help_remove) { + if (unlink == DONT_UNLINK) { // Skip over logically removed items. item = STRIP_MARK(next); @@ -149,14 +151,14 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl } else { // Unlink logically removed items. - markable_t other = SYNC_CAS(&pred->next[level], item, STRIP_MARK(next)); + markable_t other = SYNC_CAS(&pred->next[level], (markable_t)item, (markable_t)STRIP_MARK(next)); if (other == (markable_t)item) { TRACE("s3", "find_preds: unlinked item from pred %p", pred, 0); item = STRIP_MARK(next); } else { TRACE("s3", "find_preds: lost race to unlink item pred %p's link changed to %p", pred, other); if (HAS_MARK(other)) - return find_preds(preds, succs, n, sl, key, help_remove); // retry + return find_preds(preds, succs, n, sl, key, unlink); // retry item = GET_NODE(other); } next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST; @@ -177,7 +179,9 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl d = sl->key_type->cmp((void *)item->key, (void *)key); } - if (d >= 0) + if (d > 0) + break; + if (d == 0 && unlink != FORCE_UNLINK) break; pred = item; @@ -186,8 +190,7 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl TRACE("s3", "find_preds: found pred %p next %p", pred, item); - // The cast to unsigned is for the case when n is -1. - if ((unsigned)level <= (unsigned)n) { + if (level < n) { if (preds != NULL) { preds[level] = pred; } @@ -197,14 +200,6 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl } } - // fill in empty levels - if (n == -1 && item != NULL && preds != NULL) { - assert(item->top_level <= MAX_LEVEL); - for (int level = start_level + 1; level <= item->top_level; ++level) { - preds[level] = sl->head; - } - } - if (d == 0) { TRACE("s2", "find_preds: found matching item %p in skiplist, pred is %p", item, pred); return item; @@ -213,72 +208,10 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl return NULL; } -static void sl_unlink (skiplist_t *sl, map_key_t key) { - node_t *pred = sl->head; - node_t *item = NULL; - TRACE("s2", "sl_unlink: unlinking marked item with key %p", key, 0); - int d = 0; - - // Traverse the levels of from the top level to the bottom - for (int level = sl->high_water; level >= 0; --level) { - markable_t next = pred->next[level]; - if (next == DOES_NOT_EXIST) - continue; - TRACE("s3", "sl_unlink: traversing level %p starting at %p", level, pred); - if (EXPECT_FALSE(HAS_MARK(next))) { - TRACE("s2", "sl_unlink: lost a race; pred %p is marked for removal (next %p); retry", pred, next); - ASSERT(level == pred->top_level || HAS_MARK(pred->next[level+1])); - return sl_unlink(sl, key); // retry - } - item = GET_NODE(next); - while (item != NULL) { - next = item->next[level]; - - while (HAS_MARK(next)) { - TRACE("s3", "sl_unlink: found marked item %p (next is %p)", item, next); - - markable_t other = SYNC_CAS(&pred->next[level], item, STRIP_MARK(next)); - if (other == (markable_t)item) { - TRACE("s3", "sl_unlink: unlinked item from pred %p", pred, 0); - item = STRIP_MARK(next); - } else { - TRACE("s3", "sl_unlink: lost race to unlink item, pred %p's link changed to %p", pred, other); - if (HAS_MARK(other)) - return sl_unlink(sl, key); // retry - item = GET_NODE(other); - } - next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST; - } - - if (EXPECT_FALSE(item == NULL)) { - TRACE("s3", "sl_unlink: past the last item in the skiplist", 0, 0); - break; - } - - TRACE("s4", "sl_unlink: visiting item %p (next is %p)", item, next); - TRACE("s4", "sl_unlink: key %p val %p", STRIP_MARK(item->key), item->val); - - if (EXPECT_TRUE(sl->key_type == NULL)) { - d = item->key - key; - } else { - d = sl->key_type->cmp((void *)item->key, (void *)key); - } - - if (d > 0) - break; - - pred = item; - item = GET_NODE(next); - } - - TRACE("s3", "sl_unlink: at pred %p next %p", pred, item); - } -} - // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors. map_val_t sl_lookup (skiplist_t *sl, map_key_t key) { TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key, sl); - node_t *item = find_preds(NULL, NULL, 0, sl, key, FALSE); + node_t *item = find_preds(NULL, NULL, 0, sl, key, DONT_UNLINK); // If we found an matching the return its value. if (item != NULL) { @@ -338,11 +271,15 @@ map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_ TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val); ASSERT((int64_t)new_val > 0); - node_t *preds[MAX_LEVEL+1]; - node_t *nexts[MAX_LEVEL+1]; + node_t *preds[MAX_LEVELS]; + node_t *nexts[MAX_LEVELS]; node_t *new_item = NULL; - int n = random_level(); - node_t *old_item = find_preds(preds, nexts, n, sl, key, TRUE); + int n = random_levels(); + if (n > sl->high_water) { + n = SYNC_ADD(&sl->high_water, 1); + TRACE("s2", "sl_cas: incremented high water mark to %p", n, 0); + } + node_t *old_item = find_preds(preds, nexts, n, sl, key, ASSIST_UNLINK); // If there is already an item in the skiplist that matches the key just update its value. if (old_item != NULL) { @@ -362,24 +299,18 @@ map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_ // Create a new node and insert it into the skiplist. TRACE("s3", "sl_cas: attempting to insert a new item between %p and %p", preds[0], nexts[0]); map_key_t new_key = sl->key_type == NULL ? key : (map_key_t)sl->key_type->clone((void *)key); - if (n > sl->high_water) { - n = sl->high_water + 1; - int x = SYNC_ADD(&sl->high_water, 1); - x = x; - TRACE("s2", "sl_cas: incremented high water mark to %p", x, 0); - } new_item = node_alloc(n, new_key, new_val); // Set 's next pointers to their proper values markable_t next = new_item->next[0] = (markable_t)nexts[0]; - for (int level = 1; level <= new_item->top_level; ++level) { + for (int level = 1; level < new_item->num_levels; ++level) { new_item->next[level] = (markable_t)nexts[level]; } // Link into from the bottom level up. After is inserted into the bottom level // it is officially part of the skiplist. node_t *pred = preds[0]; - markable_t other = SYNC_CAS(&pred->next[0], next, new_item); + markable_t other = SYNC_CAS(&pred->next[0], next, (markable_t)new_item); if (other != next) { TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other); @@ -393,22 +324,23 @@ map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_ TRACE("s3", "sl_cas: successfully inserted a new item %p at the bottom level", new_item, 0); - for (int level = 1; level <= new_item->top_level; ++level) { + ASSERT(new_item->num_levels <= MAX_LEVELS); + for (int level = 1; level < new_item->num_levels; ++level) { TRACE("s3", "sl_cas: inserting the new item %p at level %p", new_item, level); do { node_t * pred = preds[level]; ASSERT(new_item->next[level]==(markable_t)nexts[level] || new_item->next[level]==MARK_NODE(nexts[level])); TRACE("s3", "sl_cas: attempting to to insert the new item between %p and %p", pred, nexts[level]); - markable_t other = SYNC_CAS(&pred->next[level], nexts[level], (markable_t)new_item); + markable_t other = SYNC_CAS(&pred->next[level], (markable_t)nexts[level], (markable_t)new_item); if (other == (markable_t)nexts[level]) break; // successfully linked into the skiplist at the current TRACE("s3", "sl_cas: lost a race. failed to change pred's link. expected %p found %p", nexts[level], other); // Find 's new preds and nexts. - find_preds(preds, nexts, new_item->top_level, sl, key, TRUE); + find_preds(preds, nexts, new_item->num_levels, sl, key, ASSIST_UNLINK); - for (int i = level; i <= new_item->top_level; ++i) { + for (int i = level; i < new_item->num_levels; ++i) { markable_t old_next = new_item->next[i]; if ((markable_t)nexts[i] == old_next) continue; @@ -416,12 +348,12 @@ map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_ // Update 's inconsistent next pointer before trying again. Use a CAS so if another thread // is trying to remove the new item concurrently we do not stomp on the mark it places on the item. TRACE("s3", "sl_cas: attempting to update the new item's link from %p to %p", old_next, nexts[i]); - other = SYNC_CAS(&new_item->next[i], old_next, nexts[i]); + other = SYNC_CAS(&new_item->next[i], old_next, (markable_t)nexts[i]); ASSERT(other == old_next || other == MARK_NODE(old_next)); // If another thread is removing this item we can stop linking it into to skiplist if (HAS_MARK(other)) { - sl_unlink(sl, key); // see comment below + find_preds(NULL, NULL, 0, sl, key, FORCE_UNLINK); // see comment below return DOES_NOT_EXIST; } } @@ -432,8 +364,8 @@ map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_ // make sure it is completely unlinked before we return. We might have lost a race and inserted the new item // at some level after the other thread thought it was fully removed. That is a problem because once a thread // thinks it completely unlinks a node it queues it to be freed - if (HAS_MARK(new_item->next[new_item->top_level])) { - sl_unlink(sl, key); + if (HAS_MARK(new_item->next[new_item->num_levels - 1])) { + find_preds(NULL, NULL, 0, sl, key, FORCE_UNLINK); } return DOES_NOT_EXIST; // success, inserted a new item @@ -441,8 +373,8 @@ map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_ map_val_t sl_remove (skiplist_t *sl, map_key_t key) { TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key, sl); - node_t *preds[MAX_LEVEL+1]; - node_t *item = find_preds(preds, NULL, -1, sl, key, TRUE); + node_t *preds[MAX_LEVELS]; + node_t *item = find_preds(preds, NULL, sl->high_water, sl, key, ASSIST_UNLINK); if (item == NULL) { TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0); return DOES_NOT_EXIST; @@ -451,7 +383,7 @@ map_val_t sl_remove (skiplist_t *sl, map_key_t key) { // Mark at each level of from the top down. If multiple threads try to concurrently remove // the same item only one of them should succeed. Marking the bottom level establishes which of them succeeds. markable_t old_next = 0; - for (int level = item->top_level; level >= 0; --level) { + for (int level = item->num_levels - 1; level >= 0; --level) { markable_t next; old_next = item->next[level]; do { @@ -473,7 +405,7 @@ map_val_t sl_remove (skiplist_t *sl, map_key_t key) { TRACE("s2", "sl_remove: replaced item %p's value with DOES_NOT_EXIT", item, 0); // unlink the item - sl_unlink(sl, key); + find_preds(NULL, NULL, 0, sl, key, FORCE_UNLINK); // free the node if (sl->key_type != NULL) { @@ -487,7 +419,7 @@ map_val_t sl_remove (skiplist_t *sl, map_key_t key) { void sl_print (skiplist_t *sl) { printf("high water: %d levels\n", sl->high_water); - for (int level = MAX_LEVEL; level >= 0; --level) { + for (int level = MAX_LEVELS - 1; level >= 0; --level) { node_t *item = sl->head; if (item->next[level] == DOES_NOT_EXIST) continue; @@ -511,11 +443,11 @@ void sl_print (skiplist_t *sl) { int is_marked = HAS_MARK(item->next[0]); printf("%s%p:0x%llx ", is_marked ? "*" : "", item, (uint64_t)item->key); if (item != sl->head) { - printf("[%d]", item->top_level); + printf("[%d]", item->num_levels); } else { printf("[HEAD]"); } - for (int level = 1; level <= item->top_level; ++level) { + for (int level = 1; level < item->num_levels; ++level) { node_t *next = STRIP_MARK(item->next[level]); is_marked = HAS_MARK(item->next[0]); printf(" %p%s", next, is_marked ? "*" : ""); @@ -535,7 +467,7 @@ void sl_print (skiplist_t *sl) { sl_iter_t *sl_iter_begin (skiplist_t *sl, map_key_t key) { sl_iter_t *iter = (sl_iter_t *)nbd_malloc(sizeof(sl_iter_t)); if (key != DOES_NOT_EXIST) { - find_preds(NULL, &iter->next, 0, sl, key, FALSE); + find_preds(NULL, &iter->next, 1, sl, key, DONT_UNLINK); } else { iter->next = GET_NODE(sl->head->next[0]); } diff --git a/map/unsafe_skiplist.c b/map/unsafe_skiplist.c new file mode 100644 index 0000000..f54e10d --- /dev/null +++ b/map/unsafe_skiplist.c @@ -0,0 +1,355 @@ +/* + * Written by Josh Dybnis and released to the public domain, as explained at + * http://creativecommons.org/licenses/publicdomain + * + * non thread safe skiplist + */ + +#include +#include + +#include "common.h" +#include "skiplist.h" +#include "runtime.h" +#include "mem.h" + +#define MAX_LEVEL 31 + +typedef struct node { + map_key_t key; + map_val_t val; + int top_level; + struct node *next[1]; +} node_t; + +struct sl_iter { + node_t *next; +}; + +struct sl { + node_t *head; + const datatype_t *key_type; + int high_water; // max level of any item in the list +}; + +static int random_level (void) { + unsigned r = nbd_rand(); + int n = __builtin_ctz(r) / 2; + if (n > MAX_LEVEL) { n = MAX_LEVEL; } + return n; +} + +static node_t *node_alloc (int level, map_key_t key, map_val_t val) { + assert(level >= 0 && level <= MAX_LEVEL); + size_t sz = sizeof(node_t) + level * sizeof(node_t *); + node_t *item = (node_t *)nbd_malloc(sz); + memset(item, 0, sz); + item->key = key; + item->val = val; + item->top_level = level; + TRACE("s2", "node_alloc: new node %p (%llu levels)", item, level); + return item; +} + +skiplist_t *sl_alloc (const datatype_t *key_type) { + skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t)); + sl->key_type = key_type; + sl->high_water = 0; + sl->head = node_alloc(MAX_LEVEL, 0, 0); + memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *)); + return sl; +} + +void sl_free (skiplist_t *sl) { + node_t *item = sl->head->next[0]; + while (item) { + node_t *next = item->next[0]; + if (sl->key_type != NULL) { + nbd_free((void *)item->key); + } + nbd_free(item); + item = next; + } +} + +size_t sl_count (skiplist_t *sl) { + size_t count = 0; + node_t *item = sl->head->next[0]; + while (item) { + count++; + item = item->next[0]; + } + return count; +} + +static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, map_key_t key, int help_remove) { + node_t *pred = sl->head; + node_t *item = NULL; + TRACE("s2", "find_preds: searching for key %p in skiplist (head is %p)", key, pred); + int d = 0; + int start_level = sl->high_water; + if (EXPECT_FALSE(start_level < n)) { + start_level = n; + } + + // Traverse the levels of from the top level to the bottom + for (int level = start_level; level >= 0; --level) { + node_t *next = pred->next[level]; + if (next == DOES_NOT_EXIST && level > n) + continue; + TRACE("s3", "find_preds: traversing level %p starting at %p", level, pred); + item = next; + while (item != NULL) { + next = item->next[level]; + + if (EXPECT_TRUE(sl->key_type == NULL)) { + d = item->key - key; + } else { + d = sl->key_type->cmp((void *)item->key, (void *)key); + } + + if (d >= 0) + break; + + pred = item; + item = next; + } + + TRACE("s3", "find_preds: found pred %p next %p", pred, item); + + // The cast to unsigned is for the case when n is -1. + if ((unsigned)level <= (unsigned)n) { + if (preds != NULL) { + preds[level] = pred; + } + if (succs != NULL) { + succs[level] = item; + } + } + } + + // fill in empty levels + if (n == -1 && item != NULL && preds != NULL) { + assert(item->top_level <= MAX_LEVEL); + for (int level = start_level + 1; level <= item->top_level; ++level) { + preds[level] = sl->head; + } + } + + if (d == 0) { + TRACE("s2", "find_preds: found matching item %p in skiplist, pred is %p", item, pred); + return item; + } + TRACE("s2", "find_preds: found proper place for key %p in skiplist, pred is %p. returning null", key, pred); + return NULL; +} + +static void sl_unlink (skiplist_t *sl, map_key_t key) { + node_t *pred = sl->head; + node_t *item = NULL; + TRACE("s2", "sl_unlink: unlinking marked item with key %p", key, 0); + int d = 0; + + // Traverse the levels of + for (int level = sl->high_water; level >= 0; --level) { + node_t *next = pred->next[level]; + if (next == DOES_NOT_EXIST) + continue; + TRACE("s3", "sl_unlink: traversing level %p starting at %p", level, pred); + item = next; + while (item != NULL) { + next = item->next[level]; + + if (EXPECT_TRUE(sl->key_type == NULL)) { + d = item->key - key; + } else { + d = sl->key_type->cmp((void *)item->key, (void *)key); + } + + if (d == 0) { + pred->next[level] = next; + TRACE("s3", "sl_unlink: unlinked item from pred %p", pred, 0); + item = next; + next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST; + break; + } + if (d > 0) + break; + + pred = item; + item = next; + } + + TRACE("s3", "sl_unlink: at pred %p next %p", pred, item); + } +} + +// Fast find that does not return the node's predecessors. +map_val_t sl_lookup (skiplist_t *sl, map_key_t key) { + TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key, sl); + node_t *item = find_preds(NULL, NULL, 0, sl, key, FALSE); + + // If we found an matching the return its value. + if (item != NULL) { + map_val_t val = item->val; + return val; + } + + TRACE("l1", "sl_lookup: no item in the skiplist matched the key", 0, 0); + return DOES_NOT_EXIST; +} + +map_key_t sl_min_key (skiplist_t *sl) { + node_t *item = sl->head->next[0]; + while (item != NULL) + return item->key; + return DOES_NOT_EXIST; +} + +map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_t new_val) { + TRACE("s1", "sl_cas: key %p skiplist %p", key, sl); + TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val); + ASSERT((int64_t)new_val > 0); + + node_t *preds[MAX_LEVEL+1]; + node_t *nexts[MAX_LEVEL+1]; + node_t *new_item = NULL; + int n = random_level(); + node_t *old_item = find_preds(preds, nexts, n, sl, key, TRUE); + + // If there is already an item in the skiplist that matches the key just update its value. + if (old_item != NULL) { + map_val_t old_val = old_item->val; + if (expectation == CAS_EXPECT_DOES_NOT_EXIST || + (expectation != CAS_EXPECT_WHATEVER && expectation != CAS_EXPECT_EXISTS && expectation != old_val)) { + TRACE("s1", "update_item: found an item %p in the skiplist that matched the key. the expectation was " + "not met, the skiplist was not changed", item, old_val); + return old_val; + } + old_item->val = new_val; + return old_val; + } + + if (EXPECT_FALSE(expectation != CAS_EXPECT_DOES_NOT_EXIST && expectation != CAS_EXPECT_WHATEVER)) { + TRACE("l1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0); + return DOES_NOT_EXIST; // failure, the caller expected an item for the to already exist + } + + TRACE("s3", "sl_cas: inserting a new item between %p and %p", preds[0], nexts[0]); + + // Create a new node and insert it into the skiplist. + map_key_t new_key = sl->key_type == NULL ? key : (map_key_t)sl->key_type->clone((void *)key); + if (n > sl->high_water) { + n = ++sl->high_water; + TRACE("s2", "sl_cas: incremented high water mark to %p", sl->high_water, 0); + } + new_item = node_alloc(n, new_key, new_val); + + // Set 's next pointers to their proper values + for (int level = 0; level <= new_item->top_level; ++level) { + new_item->next[level] = nexts[level]; + } + + // Link into + for (int level = 0; level <= new_item->top_level; ++level) { + preds[level]->next[level] = new_item; + } + + return DOES_NOT_EXIST; // success, inserted a new item +} + +map_val_t sl_remove (skiplist_t *sl, map_key_t key) { + TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key, sl); + node_t *preds[MAX_LEVEL+1]; + node_t *item = find_preds(preds, NULL, -1, sl, key, TRUE); + if (item == NULL) { + TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0); + return DOES_NOT_EXIST; + } + map_val_t val = item->val; + + // unlink the item + sl_unlink(sl, key); + + // free the node + if (sl->key_type != NULL) { + nbd_free((void *)item->key); + } + nbd_free(item); + + return val; +} + +void sl_print (skiplist_t *sl) { + + printf("high water: %d levels\n", sl->high_water); + for (int level = MAX_LEVEL; level >= 0; --level) { + node_t *item = sl->head; + if (item->next[level] == DOES_NOT_EXIST) + continue; + printf("(%d) ", level); + int i = 0; + while (item) { + node_t *next = item->next[level]; + printf("%p ", item); + item = next; + if (i++ > 30) { + printf("..."); + break; + } + } + printf("\n"); + fflush(stdout); + } + node_t *item = sl->head; + int i = 0; + while (item) { + printf("%p:0x%llx ", item, (uint64_t)item->key); + if (item != sl->head) { + printf("[%d]", item->top_level); + } else { + printf("[HEAD]"); + } + for (int level = 1; level <= item->top_level; ++level) { + node_t *next = item->next[level]; + printf(" %p", next); + if (item == sl->head && item->next[level] == DOES_NOT_EXIST) + break; + } + printf("\n"); + fflush(stdout); + item = item->next[0]; + if (i++ > 30) { + printf("...\n"); + break; + } + } +} + +sl_iter_t *sl_iter_begin (skiplist_t *sl, map_key_t key) { + sl_iter_t *iter = (sl_iter_t *)nbd_malloc(sizeof(sl_iter_t)); + if (key != DOES_NOT_EXIST) { + find_preds(NULL, &iter->next, 0, sl, key, FALSE); + } else { + iter->next = sl->head->next[0]; + } + return iter; +} + +map_val_t sl_iter_next (sl_iter_t *iter, map_key_t *key_ptr) { + assert(iter); + node_t *item = iter->next; + if (item == NULL) { + iter->next = NULL; + return DOES_NOT_EXIST; + } + iter->next = item->next[0]; + if (key_ptr != NULL) { + *key_ptr = item->key; + } + return item->val; +} + +void sl_iter_free (sl_iter_t *iter) { + nbd_free(iter); +} diff --git a/perf.sh b/perf.sh new file mode 100755 index 0000000..ae1be88 --- /dev/null +++ b/perf.sh @@ -0,0 +1,10 @@ +#!/bin/sh +for ks in 28 24 20 16 12 8 4 0 +do + for th in 1 2 4 8 16 + do + output/perf_test $th $ks + done +done + + diff --git a/runtime/lwt.c b/runtime/lwt.c index 3904175..a986567 100644 --- a/runtime/lwt.c +++ b/runtime/lwt.c @@ -11,7 +11,7 @@ #include "mem.h" #define LWT_BUFFER_SCALE 20 -#define LWT_BUFFER_SIZE (1 << LWT_BUFFER_SCALE) +#define LWT_BUFFER_SIZE (1ULL << LWT_BUFFER_SCALE) #define LWT_BUFFER_MASK (LWT_BUFFER_SIZE - 1) volatile int halt_ = 0; diff --git a/runtime/mem.c b/runtime/mem.c index 5b0aa67..420171e 100644 --- a/runtime/mem.c +++ b/runtime/mem.c @@ -14,18 +14,19 @@ #include "rlocal.h" #include "lwt.h" -#define MAX_SCALE 31 // allocate blocks up to 4GB (arbitrary, could be bigger) #ifndef NBD32 +#define MAX_SCALE 36 // allocate blocks up to 64GB (arbitrary, could be bigger) #define MIN_SCALE 3 // smallest allocated block is 8 bytes #define MAX_POINTER_BITS 48 #define PAGE_SCALE 21 // 2MB pages #else +#define MAX_SCALE 31 #define MIN_SCALE 2 // smallest allocated block is 4 bytes #define MAX_POINTER_BITS 32 #define PAGE_SCALE 12 // 4KB pages #endif -#define PAGE_SIZE (1 << PAGE_SCALE) -#define HEADERS_SIZE (((size_t)1 << (MAX_POINTER_BITS - PAGE_SCALE)) * sizeof(header_t)) +#define PAGE_SIZE (1ULL << PAGE_SCALE) +#define HEADERS_SIZE (((size_t)1ULL << (MAX_POINTER_BITS - PAGE_SCALE)) * sizeof(header_t)) typedef struct block { struct block *next; @@ -83,7 +84,7 @@ static void *get_new_region (int block_scale) { return region; } #endif//RECYCLE_PAGES - size_t region_size = (1 << block_scale); + size_t region_size = (1ULL << block_scale); if (region_size < PAGE_SIZE) { region_size = PAGE_SIZE; } @@ -155,13 +156,13 @@ void nbd_free (void *x) { ASSERT(b_scale && b_scale <= MAX_SCALE); #ifdef RECYCLE_PAGES if (b_scale > PAGE_SCALE) { - int rc = munmap(x, 1 << b_scale); + int rc = munmap(x, 1ULL << b_scale); ASSERT(rc == 0); rc = rc; } #endif #ifndef NDEBUG - memset(b, 0xcd, (1 << b_scale)); // bear trap + memset(b, 0xcd, (1ULL << b_scale)); // bear trap #endif tl_t *tl = &tl_[tid_]; // thread-local data if (h->owner == tid_) { @@ -270,6 +271,7 @@ void *nbd_malloc (size_t n) { if (b != NULL) { TRACE("m1", "nbd_malloc: returning block %p", b, 0); return b; + assert(b); } // The free list is empty so process blocks freed from other threads and then check again. @@ -278,6 +280,7 @@ void *nbd_malloc (size_t n) { if (b != NULL) { TRACE("m1", "nbd_malloc: returning block %p", b, 0); return b; + assert(b); } #ifdef RECYCLE_PAGES @@ -292,6 +295,7 @@ void *nbd_malloc (size_t n) { ASSERT(b != NULL); TRACE("m1", "nbd_malloc: returning block %p", b, 0); return b; + assert(b); } // There are no partially allocated pages so get a new page. @@ -304,7 +308,7 @@ void *nbd_malloc (size_t n) { // Break up the remainder of the page into blocks and put them on the free list. Start at the // end of the page so that the free list ends up in increasing order, for ease of debugging. if (b_scale < PAGE_SCALE) { - size_t block_size = (1 << b_scale); + size_t block_size = (1ULL << b_scale); block_t *head = NULL; for (int offset = PAGE_SIZE - block_size; offset > 0; offset -= block_size) { block_t *x = (block_t *)(page + offset); @@ -319,16 +323,20 @@ void *nbd_malloc (size_t n) { } TRACE("m1", "nbd_malloc: returning block %p from new region %p", b, (size_t)b & ~MASK(PAGE_SCALE)); + assert(b); return b; } #else//USE_SYSTEM_MALLOC #include +#include "common.h" +#include "rlocal.h" +#include "lwt.h" void mem_init (void) { return; } -void ndb_free (void *x) { +void nbd_free (void *x) { TRACE("m1", "nbd_free: %p", x, 0); #ifndef NDEBUG memset(x, 0xcd, sizeof(void *)); // bear trap diff --git a/runtime/mem2.c b/runtime/mem2.c new file mode 100644 index 0000000..ddb59c7 --- /dev/null +++ b/runtime/mem2.c @@ -0,0 +1,410 @@ +/* + * Written by Josh Dybnis and released to the public domain, as explained at + * http://creativecommons.org/licenses/publicdomain + * + * fast multi-threaded malloc. + */ +#ifndef USE_SYSTEM_MALLOC +#define _BSD_SOURCE // so we get MAP_ANON on linux +#include +#include +#include +#include +#include "common.h" +#include "rlocal.h" +#include "lwt.h" + +#define CHUNK_SCALE 12 // 4k chunks +#define PAGE_SCALE 21 // 2MB pages +#define PAGE_SIZE (1ULL << PAGE_SCALE) + +// On both linux and Mac OS X the size of the mmap-able virtual address space is between 2^46 and 2^47. Linux has +// no problem when you grab the whole thing. Mac OS X apparently does some O(n) thing on the first page fault +// that takes over 2 seconds if you mmap 2^46 bytes. So on Mac OS X we only take 2^38 bytes of virtual space. Which +// is OK though, since you can only buy a Mac with up to 32GB of RAM (as of 2/09). +#ifndef NBD32 +#ifdef __MACOSX__ +#define TOTAL_SCALE 38 +#else //__MACOSX__ +#define TOTAL_SCALE 46 +#endif//__MACOSX__ +#else// NBD32 +#define TOTAL_SCALE 32 +#endif//NBD32 +#define TOTAL_SIZE (1ULL << TOTAL_SCALE) + +#define INVALID_SLAB_CLASS 255 +#define METASLAB_CLASS_MAX 2 +#define NESTED_4K_SLAB_CLASS_MAX 16 +#define NESTED_32K_SLAB_CLASS_MAX 39 +#define NESTED_256K_SLAB_CLASS_MAX 63 +#define NESTED_SLAB_CLASS_MAX NESTED_256K_SLAB_CLASS_MAX +#define LARGE_SLAB_CLASS_MAX 93 +#define HUGE_SLAB_CLASS_MAX (sizeof(BlockSize) / sizeof(*BlockSize)) +#define SLAB_CLASS_MAX HUGE_SLAB_CLASS_MAX + +#define NESTED_SLAB_CASES NESTED_4K_SLAB_CASES: case NESTED_32K_SLAB_CASES: case NESTED_256K_SLAB_CASES +#define NESTED_4K_SLAB_CASES METASLAB_CLASS_MAX+1 ... NESTED_4K_SLAB_CLASS_MAX +#define NESTED_32K_SLAB_CASES NESTED_4K_SLAB_CLASS_MAX+1 ... NESTED_32K_SLAB_CLASS_MAX: case 0 +#define NESTED_256K_SLAB_CASES NESTED_32K_SLAB_CLASS_MAX+1 ... NESTED_SLAB_CLASS_MAX: case 1 +#define LARGE_SLAB_CASES NESTED_SLAB_CLASS_MAX+1 ... LARGE_SLAB_CLASS_MAX: case 2 +#define HUGE_SLAB_CASES LARGE_SLAB_CLASS_MAX+1 ... HUGE_SLAB_CLASS_MAX + +#define SLAB_CLASS_SCALE(class) ({ \ + int _scale = 0; \ + switch (class) { \ + case NESTED_4K_SLAB_CASES: _scale = 12; break; \ + case NESTED_32K_SLAB_CASES: _scale = 15; break; \ + case NESTED_256K_SLAB_CASES: _scale = 18; break; \ + case LARGE_SLAB_CASES: _scale = 21; break; \ + } \ + _scale; \ +}) + +// indexed by class +static const uint32_t BlockSize[] = { + // meta slab classes (for the nested slabs) + 1 << 12, 1 << 15, 1 << 18 + + // nested slab classes (4kB, 32kB, and 256kB) + 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, + 88, 96, 112, 120, 128, 144, 160, 176, 192, 224, + 256, 288, 320, 352, 384, 416, 448, 480, 512, 576, + 640, 704, 768, 832, 896, 960, 1024, 1152, 1280, 1408, + 1536, 1664, 1856, 2048, 2240, 2432, 2688, 2944, 3200, 3520, + 3840, 4160, 4544, 4928, 5312, 5696, 6144, 6592, 7040, 7488, + 7936, + + // large slab classes (full page, 2MB) + 8896, 9984, 11200, 12544, 14016, 15616, 17408, 19328, 21440, 23744, + 26176, 28800, 31616, 34624, 37760, 41024, 44416, 47936, 51584, 55296, + 59008, 62784, 66496, 70208, 73856, 77376, 80832, 84160, 87360, 90368, + 93248, 95936, 98496, 100864, + + // huge slabs (slabs on huge blocks, 2MB-4MB) + 110912, 121984, 134144, 147520, 162240, 178432, 196224, 215808, 237376, 261056, + 287104, 315776, 347328, 382016, 420160, 462144, 508352, 559168, 615040, 676544, + 744192, 818560, 900416, 990400, 1089408, 1198336, 1318144, 1449920, 1594880, 1754368, + 1929792 +}; + +typedef uint8_t class_t; + +typedef struct block { + struct block *next; +} block_t; + +typedef struct slab { + unsigned valid:1; + unsigned free_list:15; + unsigned num_in_use:9; + unsigned class:6; +} __attribute__((packed)) slab_t; + +typedef struct metaslab { + slab_t slab; + char * data; + slab_t slab[1 << (PAGE_SCALE - CHUNK_SCALE)]; + struct { + struct metaslab *older; + struct metaslab *newer; + } q[NESTED_SLAB_CLASS_MAX+1]; + uint64_t partial_slab_bitmap2[NESTED_32K_SLAB_CLASS_MAX+1]; + uint8_t partial_slab_bitmap1[NESTED_SLAB_CLASS_MAX+1]; +} metaslab_t; + +char *MemBase = NULL; +char *MemEnd = NULL; +char *PageBreak = NULL; +size_t *PageMap = NULL; +block_t *FreePages = NULL; +struct { slab_t *slab; char *slab_base; } ActiveSlab[SLAB_CLASS_MAX + 1] = {}; + +struct { + size_t slabs_in_use; + size_t bytes_requested; + size_t bytes_allocated; + size_t total_bytes_allocated; +} ClassStats[METASLAB_CLASS_MAX+1]; + +struct { + slab_t *oldest; + slab_t *newest; +} PartialSlabQueue[SLAB_CLASS_MAX+1]; + +struct { + slab_t *oldest; +} FreshPartialSlabQueue[SLAB_CLASS_MAX+1]; + +static block_t *get_block (class_t slab_class); + +void mem_init (void) { + ASSERT(INVALID_SLAB_CLASS > SLAB_CLASS_MAX); + + void *buf = mmap(NULL, TOTAL_SIZE, PROT_NONE, MAP_NORESERVE|MAP_ANON|MAP_PRIVATE, -1, 0); + if (buf == (void *)-1) { + perror("mmap"); + exit(-1); + } + MemEnd = buf + TOTAL_SIZE; + MemBase = (char *)( ((size_t)buf + PAGE_SIZE-1) & ~(PAGE_SIZE-1) ); // align to a page boundry + + size_t page_map_size = sizeof(void *) >> (TOTAL_SCALE - PAGE_SCALE); + mprotect(MemBase, chunk_map_size, PROT_READ|PROT_WRITE); + PageBreak = MemBase + chunk_map_size; + PageMap = (size_t *)MemBase; +} + +static class_t get_slab_class (size_t size) { + for (int i = METASLAB_CLASS_MAX + 1; i <= SLAB_CLASS_MAX; ++i) { + if (size <= BlockSize[i]) + return i; + } + return INVALID_SLAB_CLASS; +} + +static class_t get_meta_class (class_t class) { + int scale = SLAB_CLASS_SCALE(class); + if (scale == PAGE_SCALE || scale == 0) + return INVALID_SLAB_CLASS; + return (scale - 12) / 3; +} + +static void *get_page (void) { + block_t *p = FreePages; + if (p == NULL) { + p = (block_t *)PageBreak; + PageBreak += PAGE_SIZE; + return p; + } + FreePages = p->next; + return p; +} + +static void free_page (void *p) { + ASSERT(p < (void *)PageBreak); + block_t *b = (block_t *)p; + b->next = FreePages; + FreePages = b; +} + +static void init_slab (void *b, class_t slab_class) { +} + +static slab_t *new_large_slab (class_t slab_class) { + return NULL; +} + +static int find_partial_slab(metaslab_t *metaslab, class_t target_class, int target_index) { + switch (target_class) { + case NESTED_4K_SLAB_CASSES: + { + // search nearby the target first + int base_index = (target_index & ~0x7); + for (int i = 0; i < 8; ++i) { + if (base_index + i == target_index) + continue; + if (metaslab->slab[base_index + i].class == target_class) + return base_index + i; + } + do { + metaslab->partial_slab_bitmap2[target_class] &= ~(1ULL << (base_index >> 3)); + uint64_t bitmap = metaslab->partial_slab_bitmap2[target_class]; + if (bitmap == 0) + return NULL; + int n = base_index >> 3; + if (bitmap & (0xFF << (n & ~0x7))) { + bitmap &= 0xFF << (n & ~0x7); // search nearby the target first + } + base_index = COUNT_TRAILING_ZEROS(bitmap) << 3; + for (int i = 0; i < 8; ++i) { + if (metaslab->slab[base_index + i].class == target_class) + return base_index + i; + } + } while (1); + } + case NESTED_32K_SLAB_CASSES: + { + uint64_t bitmap = metaslab->partial_slab_bitmap2[target_class]; + if (bitmap == 0) + return NULL; + int n = target_index >> 3; + if (bitmap & (0xFF << (n & ~0x7))) { + bitmap &= 0xFF << (n & ~0x7); // search nearby the target first + } + return COUNT_TRAILING_ZEROS(bitmap) << 3; + } + case NESTED_256K_SLAB_CASSES: + { + uint8_t bitmap = metaslab->partial_slab_bitmap1[target_class]; + if (bitmap == 0) + return NULL; + return COUNT_TRAILING_ZEROS(bitmap) << 6; + } + default: + ASSERT(FALSE); + return -1; + } +} + +static void activate_new_slab (class_t slab_class) { + slab_t *new_slab; + switch (slab_class) { + case NESTED_SLAB_CASES: + int slab_index = ActiveSlab[slab_class].slab_index; + metaslab_t *metaslab = ActiveSlab[slab_class].metaslab; + + // First look for a partial slab on the same metaslab as the old active slab. + new_slab = find_partial_slab(metaslab, slab_class); + if (new_slab == NULL) { + // No partial slab on the same metaslab. Remove a metaslab from the front of the queue. + metaslab_t *metaslab = (metaslab_t *)PartialSlabQueue[slab_class].oldest; + if (metaslab != NULL) { + ASSERT(metaslab->q[slab_class].older == NULL); + PartialSlabQueue[slab_class].newest = (slab_t *)metaslab->q[slab_class].newer; + metaslab->q[slab_class].newer->q[slab_class].older = NULL; + new_slab = find_partial_slab(metaslab, slab_class); + } else { + // Can't find a partial slab; create a new slab. + new_slab = (slab_t *)get_block(get_meta_class(slab_class)); + init_slab(new_slab, slab_class); + } + } + break; + + case LARGE_SLAB_CASES: + case HUGE_SLAB_CASES: + // large or huge slab class + new_slab = PartialSlabQueue[slab_class].oldest; + if (new_slab == NULL) { + ASSERT(new_slab->older == NULL); + PartialSlabQueue[slab_class].newest = new_slab->newer; + new_slab->newer->older = NULL; + } + if (new_slab == NULL) { + if (IS_HUGE_SLAB_CLASS(slab_class)) { + new_slab = new_large_slab(slab_class); + } else { + ASSERT(IS_LARGE_SLAB_CLASS(slab_class)); + new_slab = (slab_t *)get_page(); + } + init_slab(new_slab, slab_class); + } + break; + + default: + ASSERT(FALSE); + } + + ActiveSlab[slab_class] = new_slab; +} + +static void *get_block(class_t slab_class) { + + // Look for a free block on the active slab. + switch (slab_class) { + case NESTED_SLAB_CASES: + int slab_index = ActiveSlab[slab_class].slab_index; + metaslab_t *metaslab = ActiveSlab[slab_class].metaslab; + if (metaslab != NULL) { + slab_t slab = metaslab->slab[slab_index]; + if (slab.free_list) { + char *slab_base = metaslab->data + ( ( slab_index - 1 ) << SLAB_CLASS_SCALE(slab_class) ); + void *b = (void *)( slab_base + ( ( slab.free_list - 1 ) << 3 ) ); + metaslab->slab[slab_index].free_list = *(uint16_t *)b; + return b; + } + } + break; + + case LARGE_SLAB_CASES: + //TODO + break; + + case HUGE_SLAB_CASES: + //TODO + break; + + default: + ASSERT(FALSE); + } + + // Find another slab, activate it, and try again. + activate_new_slab(slab_class); + return get_block(slab_class); // recursive tail-call +} + +void *nbd_malloc (size_t n) { + TRACE("m1", "nbd_malloc: size %llu", n, 0); + if (n == 0) + return NULL; + + block_t *b = get_block( get_slab_class(n) ); + + TRACE("m1", "nbd_malloc: returning block %p", b, 0); + return b; +} + +void nbd_free (void *x) { + TRACE("m1", "nbd_free: block %p", x, 0); + ASSERT(x); + ASSERT(x >= (void *)MemBase && x < (void *)MemEnd); + + block_t *b = (block_t *)x; + size_t page_index = (size_t)b >> PAGE_SCALE; + metaslab_t *metaslab = PageMap[page_index]; + ASSERT(metaslab); + size_t slab_index = ((size_t)b & MASK(PAGE_SCALE)) >> 12; + slab_t slab = metaslab->slab[slab_index]; + + // if is not valid is on a larger slab. + if (slab.valid) { + b->next = slab.free_list; + // the of the block is offset by 1 so 0 can represent NULL. + slab.free_list = ( ((size_t)b & MASK(12)) >> 3 ) + 1; + } else { + // is not on a 4kB slab. + slab_index &= 0x7; // Try the 32kB slab. + slab = metaslab->slab[slab_index]; + if (slab.valid) { + b->next = slab.free_list; + slab.free_list = ( ((size_t)b & MASK(15)) >> 3 ) + 1; + } else { + // is not on a 32kB slab. + slab_index &= 0x3F; // must be on the 256kB slab. + slab = metaslab->slab[slab_index]; + ASSERT(slab.valid); + b->next = slab.free_list; + slab.free_list = ( ((size_t)b & MASK(18)) >> 3 ) + 1; + } + } + --slab.num_in_use; + metaslab->slab[slab_index] = slab; + if (slab.num_in_use == 0) { + free_slab(metaslab, slab_index); + } +} + +#else//USE_SYSTEM_MALLOC +#include + +void mem_init (void) { + return; +} + +void ndb_free (void *x) { + TRACE("m1", "nbd_free: %p", x, 0); +#ifndef NDEBUG + memset(x, 0xcd, sizeof(void *)); // bear trap +#endif//NDEBUG + free(x); + return; +} + +void *nbd_malloc (size_t n) { + TRACE("m1", "nbd_malloc: request size %llu", n, 0); + void *x = malloc(n); + TRACE("m1", "nbd_malloc: returning %p", x, 0); + return x; +} +#endif//USE_SYSTEM_MALLOC diff --git a/runtime/mem_class_calc.c b/runtime/mem_class_calc.c new file mode 100644 index 0000000..93e34c8 --- /dev/null +++ b/runtime/mem_class_calc.c @@ -0,0 +1,134 @@ +#include +#include +#include + +typedef unsigned char uint8_t; +typedef unsigned short uint16_t; +typedef unsigned int uint32_t; + +#define CACHE_LINE_SCALE 6 + +// Return the expected fraction of bytes wasted per slab. +// +// The internal fragmentation due to using size classes is biased by including the space required, +// for a pointer to each block. +double calc_frag(int slab_size, int block_size, int delta) +{ + double quant = (double)delta / 2 / block_size; + assert(quant >= 0.0); + int blocks_per_slab = (int)(slab_size / block_size); + + // internal fragmentation that comes from tiling non-power-of-2 sized blocks in slabs + int extra_space = slab_size - blocks_per_slab * block_size; + assert(extra_space < block_size); + + // number of different cache line colors needed to evenly distribute cache line accesses + int num_colors = block_size >> CACHE_LINE_SCALE; + if (num_colors <= 1) + return (double)extra_space/slab_size + quant; + + int num_overflow = num_colors - 1 - (extra_space >> CACHE_LINE_SCALE); + if (num_overflow <= 0) + return (double)extra_space/slab_size + quant; + + double coloring = (double)num_overflow * block_size / num_colors; + return ((double)extra_space + coloring)/slab_size + quant; +} + +// size classes for various alignments, max 6% expected internal fragmentation + +// 2B-128B blocks, 4k slab +static uint8_t A1_4kB[] = { 2, 3, 5, 7, 9, 11, 14, 17, 20, 24, 28, 33, 39, 46, 53, 62, 70, 80, 91, 105, 120, 128 }; +static uint8_t A2_4kB[] = { 2, 4, 6, 8, 10, 14, 18, 22, 28, 34, 40, 48, 56, 66, 74, 84, 94, 104, 120, 128 }; +static uint8_t A4_4kB[] = { 4, 8, 12, 16, 20, 24, 32, 40, 48, 56, 68, 80, 92, 104, 120, 128 }; +static uint8_t A8_4kB[] = { 8, 16, 24, 32, 40, 48, 64, 80, 96, 112, 120, 128 }; +static uint8_t A16_4kB[] = { 16, 32, 48, 64, 80, 96, 112, 128 }; + +// 128B-1kB blocks, 32k slab +static uint16_t A1_32kB[] = { 137, 156, 178, 201, 227, 256, 288, 323, 361, 402, 447, 494, 545, 598, 654, 712, 771, 832, 895, 958, 1022 }; +static uint16_t A8_32kB[] = { 144, 168, 192, 224, 256, 296, 336, 376, 424, 472, 528, 584, 640, 704, 768, 832, 896, 960, 1024 }; +static uint16_t A16_32kB[] = { 144, 176, 208, 240, 272, 320, 368, 416, 464, 512, 576, 640, 704, 768, 832, 896, 960, 1024 }; + +// 1kB-8kB blocks, 256k slab +static uint16_t A1_256kB[] = { 1152, 1297, 1458, 1636, 1832, 2048, 2284, 2541, 2820, 3124, 3550, 3904, 4280, 4676, 5092, 5525, 5974, 6435, 6906, 7380, 7856 }; +static uint16_t A8_256kB[] = { 1152, 1288, 1440, 1608, 1792, 2000, 2224, 2472, 2744, 3032, 3344, 3680, 4040, 4416, 4816, 5232, 5664, 6112, 6568, 7032, 7504, 7976 }; +static uint16_t A64_256kB[] = { 1152, 1280, 1408, 1536, 1664, 1856, 2048, 2240, 2432, 2688, 2944, 3200, 3520, 3840, 4160, 4544, 4928, 5312, 5696, 6144, 6592, 7040, 7488, 7936 }; + +// 8kB-100kB blocks, 2MB slab +static uint32_t A64_2MB[] = { + 8896, 9984, 11200, 12544, 14016, 15616, 17408, 19328, 21440, 23744, 26176, 28800, 31616, 34624, 37760, 41024, + 44416, 47936, 51584, 55296, 59008, 62784, 66496, 70208, 73856, 77376, 80832, 84160, 87360, 90368, 93248, 95936, + 98496, 100864 +}; + +int main (void) { + + double x = 100864; + int n; + for (n = 0; n < 40 && x < (1 << 21); ++n) { + x *= 1.1; + x = (uint32_t)x & ~63; + printf("%u, ", (uint32_t)x); + } + printf("\n%d\n", n); + return 0; + const int start1 = 120832; + const int start2 = 1408; + const int alignment = 64; +#define ischosen(x) \ + (x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || \ + x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || \ + x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || \ + x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || \ + x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || \ + x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || \ + x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || \ + x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || \ + x == 0 || x == 0 || x == 0 || x == 0 || x == 0 || x == 0) + + const int slab_size = 1 << 21; + const double thresh = .06; + int block_size; + int i = 0; + for (block_size = start1; i < 87 && block_size < (slab_size >> 3); ++i, block_size += alignment) { + printf("%5d ", block_size); + + int d; + double min = 1; + int ch = block_size + alignment; + for (d = block_size; d >= alignment; d-=alignment) { + int x = block_size - d; + if (ischosen(x)) { + double f = calc_frag(slab_size, block_size, d); + if (f < thresh && f < min) { min = f; ch = d; } + } + } + + for (d = start2; d > start2 - 1024; d-=alignment) { + if (d <= block_size && d <= ch) { + double f = calc_frag(slab_size, block_size, d); + if (f < thresh) { + if (d == ch) { + printf(" *%3.1f%% ", f*100); + } else { + printf(" %4.1f%% ", f*100); + } + continue; + } + } + if (d-1 <= block_size && d-alignment <= ch && calc_frag(slab_size, block_size, d - alignment) < thresh) { + printf("%6d ", block_size); + continue; + } + printf(" "); + } + + if (ischosen(block_size)) { + printf("%5d*", block_size); + } else { + printf("%5d", block_size); + } + printf("\n"); + } + return 0; +} diff --git a/runtime/rcu.c b/runtime/rcu.c index 0c49133..8120713 100644 --- a/runtime/rcu.c +++ b/runtime/rcu.c @@ -31,7 +31,7 @@ static fifo_t *pending_[MAX_NUM_THREADS] = {}; static int num_threads_ = 0; static fifo_t *fifo_alloc(int scale) { - fifo_t *q = (fifo_t *)nbd_malloc(sizeof(fifo_t) + (1 << scale) * sizeof(void *)); + fifo_t *q = (fifo_t *)nbd_malloc(sizeof(fifo_t) + (1ULL << scale) * sizeof(void *)); memset(q, 0, sizeof(fifo_t)); q->scale = scale; q->head = 0; @@ -51,6 +51,7 @@ void rcu_update (void) { LOCALIZE_THREAD_LOCAL(tid_, int); assert(tid_ < num_threads_); int next_thread_id = (tid_ + 1) % num_threads_; + TRACE("r1", "rcu_update: updating thread %llu", next_thread_id, 0); int i; for (i = 0; i < num_threads_; ++i) { if (i == tid_) @@ -62,6 +63,7 @@ void rcu_update (void) { uint64_t x = rcu_[tid_][i]; rcu_[next_thread_id][i] = rcu_last_posted_[tid_][i] = x; + TRACE("r2", "rcu_update: posted updated value (%llu) for thread %llu", x, i); } // free diff --git a/test/haz_test.c b/test/haz_test.c index b427fda..1514efa 100644 --- a/test/haz_test.c +++ b/test/haz_test.c @@ -17,7 +17,6 @@ #include "hazard.h" #define NUM_ITERATIONS 10000000 -#define MAX_NUM_THREADS 4 typedef struct node { struct node *next; @@ -77,7 +76,7 @@ void *worker (void *arg) { int main (int argc, char **argv) { //lwt_set_trace_level("m0r0"); - int num_threads = 2; + int num_threads = MAX_NUM_THREADS; if (argc == 2) { errno = 0; diff --git a/test/map_test1.c b/test/map_test1.c index 9d45812..f05fa3d 100644 --- a/test/map_test1.c +++ b/test/map_test1.c @@ -66,7 +66,7 @@ int main (int argc, char **argv) { return -1; } - num_threads_ = 2; + num_threads_ = MAX_NUM_THREADS; if (argc == 2) { errno = 0; diff --git a/test/map_test2.c b/test/map_test2.c index 8ab2b6f..52438ea 100644 --- a/test/map_test2.c +++ b/test/map_test2.c @@ -136,7 +136,7 @@ void *add_remove_worker (void *arg) { map_t *map = wd->map; CuTest* tc = wd->tc; int d = wd->id; - int iters = 10000; + int iters = (map_type_ == &MAP_IMPL_LL ? 10000 : 100000); (void)SYNC_ADD(wd->wait, -1); do { } while (*wd->wait); // wait for all workers to be ready @@ -180,8 +180,7 @@ void concurrent_add_remove_test (CuTest* tc) { pthread_t thread[2]; worker_data_t wd[2]; - static const int num_threads = 2; - volatile int wait = num_threads; + volatile int wait = 2; #ifdef TEST_STRING_KEYS map_t *map = map_alloc(map_type_, &DATATYPE_NSTRING); #else @@ -193,7 +192,7 @@ void concurrent_add_remove_test (CuTest* tc) { // In 2 threads, add & remove even & odd elements concurrently int i; - for (i = 0; i < num_threads; ++i) { + for (i = 0; i < 2; ++i) { wd[i].id = i; wd[i].tc = tc; wd[i].map = map; @@ -202,14 +201,14 @@ void concurrent_add_remove_test (CuTest* tc) { if (rc != 0) { perror("nbd_thread_create"); return; } } - for (i = 0; i < num_threads; ++i) { + for (i = 0; i < 2; ++i) { pthread_join(thread[i], NULL); } gettimeofday(&tv2, NULL); int ms = (int)(1000000*(tv2.tv_sec - tv1.tv_sec) + tv2.tv_usec - tv1.tv_usec) / 1000; map_print(map); - printf("Th:%d Time:%dms\n", num_threads, ms); + printf("Time:%dms\n", ms); fflush(stdout); // In the end, all members should be removed @@ -316,7 +315,7 @@ void big_iteration_test (CuTest* tc) { } int main (void) { - lwt_set_trace_level("H3m3l2t0"); + lwt_set_trace_level("r0m3l2t0"); static const map_impl_t *map_types[] = { &MAP_IMPL_LL, &MAP_IMPL_SL, &MAP_IMPL_HT }; for (int i = 0; i < sizeof(map_types)/sizeof(*map_types); ++i) { @@ -327,9 +326,9 @@ int main (void) { CuSuite* suite = CuSuiteNew(); SUITE_ADD_TEST(suite, concurrent_add_remove_test); - SUITE_ADD_TEST(suite, basic_test); - SUITE_ADD_TEST(suite, basic_iteration_test); - SUITE_ADD_TEST(suite, big_iteration_test); +// SUITE_ADD_TEST(suite, basic_test); +// SUITE_ADD_TEST(suite, basic_iteration_test); +// SUITE_ADD_TEST(suite, big_iteration_test); CuSuiteRun(suite); CuSuiteDetails(suite, output); diff --git a/test/perf_test.c b/test/perf_test.c index ace96d0..cc55ce0 100644 --- a/test/perf_test.c +++ b/test/perf_test.c @@ -18,15 +18,17 @@ static volatile int wait_; static volatile int stop_; -static long num_threads_; +static int num_threads_; +static int duration_; static map_t *map_; static int get_range_; static int put_range_; -static int num_keys_; +static size_t num_keys_; static map_key_t *keys_ = NULL; -static uint64_t times_[MAX_NUM_THREADS] = {}; static int ops_[MAX_NUM_THREADS] = {}; +#define FOO (1ULL << 20) + void *worker (void *arg) { int tid = (int)(size_t)arg; uint64_t s = nbd_rand_seed(tid); @@ -36,15 +38,14 @@ void *worker (void *arg) { (void)SYNC_ADD(&wait_, -1); do {} while (wait_); - uint64_t t1 = rdtsc(); - while (!stop_) { - int r = nbd_next_rand(&s); - int x = r & ( (1 << 20) - 1 ); - int i = nbd_next_rand(&s) & (num_keys_ - 1); - map_key_t key = keys_[i]; + map_key_t key = keys_[ nbd_next_rand(&s) & (num_keys_ - 1) ]; + uint32_t x = nbd_next_rand(&s) & (FOO - 1); if (x < get_range_) { - map_val_t val = map_get(map_, key); +#ifndef NDEBUG + map_val_t val = +#endif + map_get(map_, key); #ifdef TEST_STRING_KEYS ASSERT(val == DOES_NOT_EXIST || ns_cmp((nstring_t *)key, (nstring_t *)val) == 0); #else @@ -61,13 +62,13 @@ void *worker (void *arg) { rcu_update(); } - times_[tid] = rdtsc() - t1; ops_[tid] = get_ops + put_ops + del_ops; return NULL; } -void run_test (void) { +int run_test (void) { + int ops; wait_ = num_threads_ + 1; // Quicky sanity check @@ -95,24 +96,29 @@ void run_test (void) { do { /* nothing */ } while (wait_ != 1); wait_ = 0; - sleep(2); + sleep(duration_); stop_ = 1; for (int i = 0; i < num_threads_; ++i) { pthread_join(thread[i], NULL); } + ops = 0; + for (int i = 0; i < num_threads_; ++i) { + ops += ops_[i]; + } + return ops; } int main (int argc, char **argv) { char* program_name = argv[0]; - if (argc > 2) { + if (argc > 3) { fprintf(stderr, "Usage: %s num_threads\n", program_name); return -1; } num_threads_ = 2; - if (argc == 2) + if (argc > 1) { errno = 0; num_threads_ = strtol(argv[1], NULL, 10); @@ -130,13 +136,27 @@ int main (int argc, char **argv) { } } + int table_scale = 12; + if (argc > 2) { + table_scale = strtol(argv[2], NULL, 10); + if (errno) { + fprintf(stderr, "%s: Invalid argument for the scale of the collection\n", program_name); + return -1; + } + table_scale = strtol(argv[2], NULL, 10); + if (table_scale < 0 || table_scale > 31) { + fprintf(stderr, "%s: The scale of the collection must be between 0 and 31\n", program_name); + return -1; + } + } + - int table_scale = 10; - int read_ratio = 95; - get_range_ = (read_ratio << 20) / 100; - put_range_ = (((1 << 20) - get_range_) >> 1) + get_range_; + int read_ratio = 90; + int put_ratio = 50; + get_range_ = (int)((double)FOO / 100 * read_ratio); + put_range_ = get_range_ + (int)(((double)FOO - get_range_) / 100 * put_ratio); - static const map_impl_t *map_types[] = { &MAP_IMPL_HT }; + static const map_impl_t *map_types[] = { &MAP_IMPL_SL }; for (int i = 0; i < sizeof(map_types)/sizeof(*map_types); ++i) { #ifdef TEST_STRING_KEYS map_ = map_alloc(map_types[i], &DATATYPE_NSTRING); @@ -145,9 +165,10 @@ int main (int argc, char **argv) { #endif // Do some warmup - num_keys_ = 1 << table_scale; + num_keys_ = 1ULL << table_scale; keys_ = nbd_malloc(sizeof(map_key_t) * num_keys_); - for (int j = 0; j < num_keys_; ++j) { + ASSERT(keys_ != NULL); + for (uint64_t j = 0; j < num_keys_; ++j) { #ifdef TEST_STRING_KEYS char tmp[64]; snprintf(tmp, sizeof(tmp), "%dabc%d", j, j*17+123); @@ -159,18 +180,17 @@ int main (int argc, char **argv) { #endif } - struct timeval tv1, tv2; - gettimeofday(&tv1, NULL); - + duration_ = 10; int num_trials = 1; + int ops = 0; for (int i = 0; i < num_trials; ++i) { - run_test(); + ops += run_test(); } + double ops_per_sec = ops / num_trials / duration_; - gettimeofday(&tv2, NULL); - int ms = (int)(1000000*(tv2.tv_sec - tv1.tv_sec) + tv2.tv_usec - tv1.tv_usec) / 1000; - map_print(map_); - printf("Th:%ld Time:%dms\n\n", num_threads_, ms); + //map_print(map_); + printf("Threads:%-2d Size:2^%-2d Mops/Sec:%-4.3g per-thread:%-4.3g\n\n", + num_threads_, table_scale, ops_per_sec/1000000, ops_per_sec/num_threads_/1000000); fflush(stdout); map_free(map_); diff --git a/test/rcu_test.c b/test/rcu_test.c index 14134b9..b2dedaa 100644 --- a/test/rcu_test.c +++ b/test/rcu_test.c @@ -80,7 +80,7 @@ void *worker (void *arg) { int main (int argc, char **argv) { lwt_set_trace_level("m3r3"); - int num_threads = 2; + int num_threads = MAX_NUM_THREADS; if (argc == 2) { errno = 0; -- 2.40.0