This is an automated email from the ASF dual-hosted git repository.
gaodayue pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-doris.git
The following commit(s) were added to refs/heads/master by this push:
new 4d1e926b6c [feature][config] introduce a new BE config
storage_page_cache_shard_size (#9821)
4d1e926b6c is described below
commit 4d1e926b6c8b637c4707bbcbe37042d237c081db
Author: Dayue Gao <gaoda...@meituan.com>
AuthorDate: Sat May 28 10:17:09 2022 +0800
[feature][config] introduce a new BE config storage_page_cache_shard_size
(#9821)
Co-authored-by: gaodayue <gaoda...@bytedance.com>
---
be/src/common/config.h | 3 ++
be/src/olap/lru_cache.cpp | 47 +++++++++++++++++------------
be/src/olap/lru_cache.h | 21 +++++++------
be/src/olap/page_cache.cpp | 20 +++++++-----
be/src/olap/page_cache.h | 7 +++--
be/src/olap/segment_loader.cpp | 2 +-
be/src/runtime/exec_env_init.cpp | 5 +--
be/test/olap/page_cache_test.cpp | 8 +++--
docs/en/admin-manual/config/be-config.md | 6 ++++
docs/zh-CN/admin-manual/config/be-config.md | 6 ++++
10 files changed, 80 insertions(+), 45 deletions(-)
diff --git a/be/src/common/config.h b/be/src/common/config.h
index a73efcf24d..9bdd8ba3e4 100644
--- a/be/src/common/config.h
+++ b/be/src/common/config.h
@@ -226,6 +226,9 @@ CONF_Int64(index_stream_cache_capacity, "10737418240");
// Cache for storage page size
CONF_String(storage_page_cache_limit, "20%");
+// Shard size for page cache, the value must be power of two.
+// It's recommended to set it to a value close to the number of BE cores in
order to reduce lock contentions.
+CONF_Int32(storage_page_cache_shard_size, "16");
// Percentage for index page cache
// all storage page cache will be divided into data_page_cache and
index_page_cache
CONF_Int32(index_page_cache_percentage, "10");
diff --git a/be/src/olap/lru_cache.cpp b/be/src/olap/lru_cache.cpp
index 90b2e4ddf5..19d8b5d4e9 100644
--- a/be/src/olap/lru_cache.cpp
+++ b/be/src/olap/lru_cache.cpp
@@ -11,6 +11,7 @@
#include <sstream>
#include <string>
+#include "gutil/bits.h"
#include "olap/olap_common.h"
#include "olap/olap_define.h"
#include "olap/olap_index.h"
@@ -430,19 +431,25 @@ inline uint32_t ShardedLRUCache::_hash_slice(const
CacheKey& s) {
return s.hash(s.data(), s.size(), 0);
}
-uint32_t ShardedLRUCache::_shard(uint32_t hash) {
- return hash >> (32 - kNumShardBits);
-}
-
-ShardedLRUCache::ShardedLRUCache(const std::string& name, size_t
total_capacity, LRUCacheType type)
+ShardedLRUCache::ShardedLRUCache(const std::string& name, size_t
total_capacity, LRUCacheType type,
+ uint32_t num_shards)
: _name(name),
+ _num_shard_bits(Bits::FindLSBSetNonZero(num_shards)),
+ _num_shards(num_shards),
+ _shards(nullptr),
_last_id(1),
_mem_tracker(MemTracker::create_tracker(-1, name, nullptr,
MemTrackerLevel::OVERVIEW)) {
- const size_t per_shard = (total_capacity + (kNumShards - 1)) / kNumShards;
- for (int s = 0; s < kNumShards; s++) {
- _shards[s] = new LRUCache(type);
- _shards[s]->set_capacity(per_shard);
+ CHECK(num_shards > 0) << "num_shards cannot be 0";
+ CHECK_EQ((num_shards & (num_shards - 1)), 0)
+ << "num_shards should be power of two, but got " << num_shards;
+
+ const size_t per_shard = (total_capacity + (_num_shards - 1)) /
_num_shards;
+ LRUCache** shards = new (std::nothrow) LRUCache*[_num_shards];
+ for (int s = 0; s < _num_shards; s++) {
+ shards[s] = new LRUCache(type);
+ shards[s]->set_capacity(per_shard);
}
+ _shards = shards;
_entity = DorisMetrics::instance()->metric_registry()->register_entity(
std::string("lru_cache:") + name, {{"name", name}});
@@ -456,8 +463,11 @@ ShardedLRUCache::ShardedLRUCache(const std::string& name,
size_t total_capacity,
}
ShardedLRUCache::~ShardedLRUCache() {
- for (int s = 0; s < kNumShards; s++) {
- delete _shards[s];
+ if (_shards) {
+ for (int s = 0; s < _num_shards; s++) {
+ delete _shards[s];
+ }
+ delete[] _shards;
}
_entity->deregister_hook(_name);
DorisMetrics::instance()->metric_registry()->deregister_entity(_entity);
@@ -501,7 +511,7 @@ uint64_t ShardedLRUCache::new_id() {
int64_t ShardedLRUCache::prune() {
int64_t num_prune = 0;
- for (int s = 0; s < kNumShards; s++) {
+ for (int s = 0; s < _num_shards; s++) {
num_prune += _shards[s]->prune();
}
return num_prune;
@@ -509,7 +519,7 @@ int64_t ShardedLRUCache::prune() {
int64_t ShardedLRUCache::prune_if(CacheValuePredicate pred) {
int64_t num_prune = 0;
- for (int s = 0; s < kNumShards; s++) {
+ for (int s = 0; s < _num_shards; s++) {
num_prune += _shards[s]->prune_if(pred);
}
return num_prune;
@@ -520,7 +530,7 @@ void ShardedLRUCache::update_cache_metrics() const {
size_t total_usage = 0;
size_t total_lookup_count = 0;
size_t total_hit_count = 0;
- for (int i = 0; i < kNumShards; i++) {
+ for (int i = 0; i < _num_shards; i++) {
total_capacity += _shards[i]->get_capacity();
total_usage += _shards[i]->get_usage();
total_lookup_count += _shards[i]->get_lookup_count();
@@ -536,12 +546,9 @@ void ShardedLRUCache::update_cache_metrics() const {
: ((double)total_hit_count /
total_lookup_count));
}
-Cache* new_lru_cache(const std::string& name, size_t capacity) {
- return new ShardedLRUCache(name, capacity, LRUCacheType::SIZE);
-}
-
-Cache* new_typed_lru_cache(const std::string& name, size_t capacity,
LRUCacheType type) {
- return new ShardedLRUCache(name, capacity, type);
+Cache* new_lru_cache(const std::string& name, size_t capacity, LRUCacheType
type,
+ uint32_t num_shards) {
+ return new ShardedLRUCache(name, capacity, type, num_shards);
}
} // namespace doris
diff --git a/be/src/olap/lru_cache.h b/be/src/olap/lru_cache.h
index 6fb87744f4..4bac173fad 100644
--- a/be/src/olap/lru_cache.h
+++ b/be/src/olap/lru_cache.h
@@ -53,11 +53,10 @@ enum LRUCacheType {
NUMBER // The capacity of cache is based on the number of cache entry.
};
-// Create a new cache with a specified name and a fixed SIZE capacity.
+// Create a new cache with a specified name and capacity.
// This implementation of Cache uses a least-recently-used eviction policy.
-extern Cache* new_lru_cache(const std::string& name, size_t capacity);
-
-extern Cache* new_typed_lru_cache(const std::string& name, size_t capacity,
LRUCacheType type);
+extern Cache* new_lru_cache(const std::string& name, size_t capacity,
+ LRUCacheType type = LRUCacheType::SIZE, uint32_t
num_shards = 16);
class CacheKey {
public:
@@ -356,12 +355,10 @@ private:
uint64_t _hit_count = 0; // 命中cache的总次数
};
-static const int kNumShardBits = 4;
-static const int kNumShards = 1 << kNumShardBits;
-
class ShardedLRUCache : public Cache {
public:
- explicit ShardedLRUCache(const std::string& name, size_t total_capacity,
LRUCacheType type);
+ explicit ShardedLRUCache(const std::string& name, size_t total_capacity,
LRUCacheType type,
+ uint32_t num_shards);
// TODO(fdy): 析构时清除所有cache元素
virtual ~ShardedLRUCache();
virtual Handle* insert(const CacheKey& key, void* value, size_t charge,
@@ -381,10 +378,14 @@ private:
private:
static uint32_t _hash_slice(const CacheKey& s);
- static uint32_t _shard(uint32_t hash);
+ uint32_t _shard(uint32_t hash) {
+ return _num_shard_bits > 0 ? (hash >> (32 - _num_shard_bits)) : 0;
+ }
std::string _name;
- LRUCache* _shards[kNumShards];
+ const int _num_shard_bits;
+ const uint32_t _num_shards;
+ LRUCache** _shards;
std::atomic<uint64_t> _last_id;
std::shared_ptr<MemTracker> _mem_tracker;
diff --git a/be/src/olap/page_cache.cpp b/be/src/olap/page_cache.cpp
index c1b0aac1d8..2ec540b384 100644
--- a/be/src/olap/page_cache.cpp
+++ b/be/src/olap/page_cache.cpp
@@ -23,26 +23,32 @@ namespace doris {
StoragePageCache* StoragePageCache::_s_instance = nullptr;
-void StoragePageCache::create_global_cache(size_t capacity, int32_t
index_cache_percentage) {
+void StoragePageCache::create_global_cache(size_t capacity, int32_t
index_cache_percentage,
+ uint32_t num_shards) {
DCHECK(_s_instance == nullptr);
- static StoragePageCache instance(capacity, index_cache_percentage);
+ static StoragePageCache instance(capacity, index_cache_percentage,
num_shards);
_s_instance = &instance;
}
-StoragePageCache::StoragePageCache(size_t capacity, int32_t
index_cache_percentage)
+StoragePageCache::StoragePageCache(size_t capacity, int32_t
index_cache_percentage,
+ uint32_t num_shards)
: _index_cache_percentage(index_cache_percentage),
_mem_tracker(MemTracker::create_tracker(capacity,
"StoragePageCache", nullptr,
MemTrackerLevel::OVERVIEW)) {
SCOPED_SWITCH_THREAD_LOCAL_MEM_TRACKER(_mem_tracker);
if (index_cache_percentage == 0) {
- _data_page_cache =
std::unique_ptr<Cache>(new_lru_cache("DataPageCache", capacity));
+ _data_page_cache = std::unique_ptr<Cache>(
+ new_lru_cache("DataPageCache", capacity, LRUCacheType::SIZE,
num_shards));
} else if (index_cache_percentage == 100) {
- _index_page_cache =
std::unique_ptr<Cache>(new_lru_cache("IndexPageCache", capacity));
+ _index_page_cache = std::unique_ptr<Cache>(
+ new_lru_cache("IndexPageCache", capacity, LRUCacheType::SIZE,
num_shards));
} else if (index_cache_percentage > 0 && index_cache_percentage < 100) {
_data_page_cache = std::unique_ptr<Cache>(
- new_lru_cache("DataPageCache", capacity * (100 -
index_cache_percentage) / 100));
+ new_lru_cache("DataPageCache", capacity * (100 -
index_cache_percentage) / 100,
+ LRUCacheType::SIZE, num_shards));
_index_page_cache = std::unique_ptr<Cache>(
- new_lru_cache("IndexPageCache", capacity *
index_cache_percentage / 100));
+ new_lru_cache("IndexPageCache", capacity *
index_cache_percentage / 100,
+ LRUCacheType::SIZE, num_shards));
} else {
CHECK(false) << "invalid index page cache percentage";
}
diff --git a/be/src/olap/page_cache.h b/be/src/olap/page_cache.h
index be5a3ffd71..f03f50bd5a 100644
--- a/be/src/olap/page_cache.h
+++ b/be/src/olap/page_cache.h
@@ -54,14 +54,17 @@ public:
}
};
+ static constexpr uint32_t kDefaultNumShards = 16;
+
// Create global instance of this class
- static void create_global_cache(size_t capacity, int32_t
index_cache_percentage);
+ static void create_global_cache(size_t capacity, int32_t
index_cache_percentage,
+ uint32_t num_shards = kDefaultNumShards);
// Return global instance.
// Client should call create_global_cache before.
static StoragePageCache* instance() { return _s_instance; }
- StoragePageCache(size_t capacity, int32_t index_cache_percentage);
+ StoragePageCache(size_t capacity, int32_t index_cache_percentage, uint32_t
num_shards);
// Lookup the given page in the cache.
//
diff --git a/be/src/olap/segment_loader.cpp b/be/src/olap/segment_loader.cpp
index 5b62c92e24..a105336f6a 100644
--- a/be/src/olap/segment_loader.cpp
+++ b/be/src/olap/segment_loader.cpp
@@ -32,7 +32,7 @@ void SegmentLoader::create_global_instance(size_t capacity) {
SegmentLoader::SegmentLoader(size_t capacity) {
_cache = std::unique_ptr<Cache>(
- new_typed_lru_cache("SegmentLoader:SegmentCache", capacity,
LRUCacheType::NUMBER));
+ new_lru_cache("SegmentLoader:SegmentCache", capacity,
LRUCacheType::NUMBER));
}
bool SegmentLoader::_lookup(const SegmentLoader::CacheKey& key,
SegmentCacheHandle* handle) {
diff --git a/be/src/runtime/exec_env_init.cpp b/be/src/runtime/exec_env_init.cpp
index ff5b847810..a1cdf32b63 100644
--- a/be/src/runtime/exec_env_init.cpp
+++ b/be/src/runtime/exec_env_init.cpp
@@ -250,8 +250,9 @@ Status ExecEnv::_init_mem_tracker() {
// Reason same as buffer_pool_limit
storage_cache_limit = storage_cache_limit / 2;
}
- int32_t index_page_cache_percentage = config::index_page_cache_percentage;
- StoragePageCache::create_global_cache(storage_cache_limit,
index_page_cache_percentage);
+ int32_t index_percentage = config::index_page_cache_percentage;
+ uint32_t num_shards = config::storage_page_cache_shard_size;
+ StoragePageCache::create_global_cache(storage_cache_limit,
index_percentage, num_shards);
LOG(INFO) << "Storage page cache memory limit: "
<< PrettyPrinter::print(storage_cache_limit, TUnit::BYTES)
<< ", origin config value: " << config::storage_page_cache_limit;
diff --git a/be/test/olap/page_cache_test.cpp b/be/test/olap/page_cache_test.cpp
index 1eca896010..73c11c39b8 100644
--- a/be/test/olap/page_cache_test.cpp
+++ b/be/test/olap/page_cache_test.cpp
@@ -21,6 +21,8 @@
namespace doris {
+static int kNumShards = StoragePageCache::kDefaultNumShards;
+
class StoragePageCacheTest : public testing::Test {
public:
StoragePageCacheTest() {}
@@ -29,7 +31,7 @@ public:
// All cache space is allocated to data pages
TEST(StoragePageCacheTest, data_page_only) {
- StoragePageCache cache(kNumShards * 2048, 0);
+ StoragePageCache cache(kNumShards * 2048, 0, kNumShards);
StoragePageCache::CacheKey key("abc", 0);
StoragePageCache::CacheKey memory_key("mem", 0);
@@ -89,7 +91,7 @@ TEST(StoragePageCacheTest, data_page_only) {
// All cache space is allocated to index pages
TEST(StoragePageCacheTest, index_page_only) {
- StoragePageCache cache(kNumShards * 2048, 100);
+ StoragePageCache cache(kNumShards * 2048, 100, kNumShards);
StoragePageCache::CacheKey key("abc", 0);
StoragePageCache::CacheKey memory_key("mem", 0);
@@ -149,7 +151,7 @@ TEST(StoragePageCacheTest, index_page_only) {
// Cache space is allocated by index_page_cache_ratio
TEST(StoragePageCacheTest, mixed_pages) {
- StoragePageCache cache(kNumShards * 2048, 10);
+ StoragePageCache cache(kNumShards * 2048, 10, kNumShards);
StoragePageCache::CacheKey data_key("data", 0);
StoragePageCache::CacheKey index_key("index", 0);
diff --git a/docs/en/admin-manual/config/be-config.md
b/docs/en/admin-manual/config/be-config.md
index 4bddb44111..ddb0e31f83 100644
--- a/docs/en/admin-manual/config/be-config.md
+++ b/docs/en/admin-manual/config/be-config.md
@@ -1136,6 +1136,12 @@ Default: 20%
Cache for storage page size
+### `storage_page_cache_shard_size`
+
+Default: 16
+
+Shard size of StoragePageCache, the value must be power of two. It's
recommended to set it to a value close to the number of BE cores in order to
reduce lock contentions.
+
### `index_page_cache_percentage`
* Type: int32
* Description: Index page cache as a percentage of total storage page cache,
value range is [0, 100]
diff --git a/docs/zh-CN/admin-manual/config/be-config.md
b/docs/zh-CN/admin-manual/config/be-config.md
index 873ef1d2b8..2d6599a284 100644
--- a/docs/zh-CN/admin-manual/config/be-config.md
+++ b/docs/zh-CN/admin-manual/config/be-config.md
@@ -1143,6 +1143,12 @@
storage_flood_stage_usage_percent和storage_flood_stage_left_capacity_bytes两
缓存存储页大小
+### `storage_page_cache_shard_size`
+
+默认值: 16
+
+StoragePageCache的分片大小，值为 2^n (n=0,1,2,...)。建议设置为接近BE
CPU核数的值，可减少StoragePageCache的锁竞争。
+
### `index_page_cache_percentage`
* 类型：int32
* 描述：索引页缓存占总页面缓存的百分比，取值为[0, 100]。
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@doris.apache.org
For additional commands, e-mail: commits-h...@doris.apache.org
