hcq/oceanbase
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[CP] fix io schedule when close resource_plan

Browse Source
This commit is contained in:
renju96
2024-02-07 02:13:28 +00:00
committed by ob-robot
parent bbd30977cd
commit 1be977d9d5
7 changed files with 255 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
src/share/io/io_schedule/ob_io_mclock.cpp
View File

@ -24,6 +24,7 @@ using namespace oceanbase::common;
ObMClock::ObMClock()
: is_inited_(false),
is_stopped_(false),
is_unlimited_(false),
reservation_clock_(),
limitation_clock_(),
proportion_clock_()
@ -52,6 +53,7 @@ int ObMClock::init(const int64_t min_iops, const int64_t max_iops, const int64_t
limitation_clock_.last_ns_ = 0;
proportion_clock_.iops_ = weight;
proportion_clock_.last_ns_ = proportion_ts * 1000L;
is_unlimited_ = false;
is_stopped_ = false;
is_inited_ = true;
}
@ -84,12 +86,14 @@ void ObMClock::start()
void ObMClock::stop()
{
is_stopped_ = true;
is_unlimited_ = true;
}
void ObMClock::destroy()
{
is_inited_ = false;
is_stopped_ = false;
is_unlimited_ = false;
reservation_clock_.reset();
limitation_clock_.reset();
proportion_clock_.reset();
@ -110,6 +114,11 @@ bool ObMClock::is_stop() const
return is_stopped_;
}
bool ObMClock::is_unlimited() const
{
return is_unlimited_;
}
int ObMClock::calc_phy_clock(const int64_t current_ts, const double iops_scale, const double weight_scale, ObPhyQueue *phy_queue)
{
int ret = OB_SUCCESS;
@ -121,6 +130,11 @@ int ObMClock::calc_phy_clock(const int64_t current_ts, const double iops_scale,
|| weight_scale <= std::numeric_limits<double>::epsilon())) {
ret = OB_INVALID_ARGUMENT;
LOG_WARN("invalid argument", K(ret), K(current_ts), K(iops_scale), K(weight_scale));
} else if (is_unlimited_) {
phy_queue->reservation_ts_ = current_ts;
phy_queue->group_limitation_ts_ = current_ts;
phy_queue->tenant_limitation_ts_ = current_ts;
phy_queue->proportion_ts_ = current_ts;
} else {
reservation_clock_.atom_update(current_ts, iops_scale, phy_queue->reservation_ts_);
limitation_clock_.atom_update(current_ts, iops_scale, phy_queue->group_limitation_ts_);
@ -489,6 +503,14 @@ int ObTenantIOClock::update_clock_unit_config(const ObTenantIOConfig &io_config)
return ret;
}
bool ObTenantIOClock::is_unlimited_config(const ObMClock &clock, const ObTenantIOConfig::GroupConfig &cur_config)
{
return clock.is_stop() ||
cur_config.deleted_ ||
cur_config.cleared_ ||
(cur_config.min_percent_ == INT64_MAX && cur_config.max_percent_ == INT64_MAX);
}
int ObTenantIOClock::update_io_clock(const int64_t index, const ObTenantIOConfig &io_config, const int64_t all_group_num)
{
int ret = OB_SUCCESS;
@ -518,8 +540,9 @@ int ObTenantIOClock::update_io_clock(const int64_t index, const ObTenantIOConfig
const ObTenantIOConfig::GroupConfig &cur_config = io_config.group_configs_.at(index);
if (!group_clocks_.at(index).is_inited()) {
LOG_WARN("clock is not init", K(ret), K(index), K(group_clocks_.at(index)));
} else if (group_clocks_.at(index).is_stop() || cur_config.deleted_ || cur_config.cleared_) {
// group has been deleted, ignore
} else if (is_unlimited_config(group_clocks_.at(index), cur_config)) {
group_clocks_.at(index).set_unlimited();
LOG_INFO("clock set unlimited", K(group_clocks_.at(index)), K(cur_config));
} else if (!cur_config.is_valid()) {
LOG_WARN("config is not valid", K(ret), K(index), K(cur_config), K(group_clocks_.at(index)));
// stop
@ -531,6 +554,9 @@ int ObTenantIOClock::update_io_clock(const int64_t index, const ObTenantIOConfig
LOG_WARN("update group io clock failed", K(ret), K(index), K(unit_config), K(cur_config));
} else {
group_clocks_.at(index).start();
if (!is_unlimited_config(group_clocks_.at(index), cur_config)) {
group_clocks_.at(index).set_limited();
}
LOG_INFO("update group clock success", K(index), K(unit_config), K(cur_config));
}
} else {

7
src/share/io/io_schedule/ob_io_mclock.h
View File

@ -33,18 +33,22 @@ public:
void start();
void stop();
void destroy();
void set_unlimited() { is_unlimited_ = true; }
void set_limited() { is_unlimited_ = false; }
bool is_inited() const;
bool is_valid() const;
bool is_stop() const;
bool is_unlimited() const;
int calc_phy_clock(const int64_t current_ts, const double iops_scale, const double weight_scale, ObPhyQueue *phy_queue);
int dial_back_reservation_clock(const double iops_scale);
int time_out_dial_back(const double iops_scale, const double weight_scale);
int dial_back_proportion_clock(const int64_t delta_us);
int64_t get_proportion_ts() const;
TO_STRING_KV(K(is_inited_), K(is_stopped_), K_(reservation_clock), K_(limitation_clock), K_(proportion_clock));
TO_STRING_KV(K(is_inited_), K(is_stopped_), K_(reservation_clock), K_(is_unlimited), K_(limitation_clock), K_(proportion_clock));
private:
bool is_inited_;
bool is_stopped_;
bool is_unlimited_; // use this flag to send io_req in useless_queue out ASAP
ObAtomIOClock reservation_clock_;
ObAtomIOClock limitation_clock_;
ObAtomIOClock proportion_clock_;
@ -68,6 +72,7 @@ public:
int update_io_clock(const int64_t index, const ObTenantIOConfig &io_config, const int64_t all_group_num);
int update_clock_unit_config(const ObTenantIOConfig &io_config);
int64_t get_min_proportion_ts();
bool is_unlimited_config(const ObMClock &clock, const ObTenantIOConfig::GroupConfig &cur_config);
void stop_clock(const uint64_t index);
TO_STRING_KV(K(is_inited_), "group_clocks", group_clocks_, "other_clock", other_group_clock_,
K_(unit_clock), K(io_config_), K(io_usage_));

4
src/share/io/ob_io_calibration.cpp
View File

@ -702,10 +702,6 @@ int ObIOCalibration::read_from_table()
if (OB_FAIL(update_io_ability(tmp_ability))) {
LOG_WARN("update io ability failed", K(ret), K(tmp_ability));
}
} else {
if (OB_FAIL(reset_io_ability())) {
LOG_WARN("reset io ability failed", K(ret));
}
}
return ret;
}

20
src/share/io/ob_io_define.cpp
View File

@ -1770,7 +1770,15 @@ int64_t ObTenantIOConfig::to_string(char* buf, const int64_t buf_len) const
bool need_comma = false;
for (int64_t i = 0; i < group_configs_.count(); ++i) {
if (group_configs_.at(i).deleted_ || group_configs_.at(i).cleared_) {
continue;
if (need_comma) {
J_COMMA();
}
BUF_PRINTF("group_id = ");
char group_id[8];
snprintf(group_id, sizeof(group_id), "%ld", group_ids_.at(i));
J_KV(group_id, group_configs_.at(i).deleted_);
J_KV(" cleared", group_configs_.at(i).cleared_);
need_comma = true;
} else if (!self_valid || group_configs_.at(i).is_valid()) {
if (need_comma) {
J_COMMA();
@ -2087,6 +2095,16 @@ int ObMClockQueue::pop_with_ready_queue(const int64_t current_ts, ObIORequest *&
ret = OB_EAGAIN;
if (!r_heap_.empty() && !r_heap_.top()->req_list_.is_empty()) {
ObPhyQueue *next_tmp_phy_queue = r_heap_.top();
if (OB_UNLIKELY(next_tmp_phy_queue->reservation_ts_ == INT64_MAX &&
(next_tmp_phy_queue->group_limitation_ts_ == INT64_MAX || next_tmp_phy_queue->tenant_limitation_ts_== INT64_MAX) &&
next_tmp_phy_queue->proportion_ts_== INT64_MAX)) {
// 对应min = max = 0的极端场景
const int64_t current_ts = ObTimeUtility::fast_current_time();
next_tmp_phy_queue->reservation_ts_ = current_ts;
next_tmp_phy_queue->group_limitation_ts_ = current_ts;
next_tmp_phy_queue->tenant_limitation_ts_ = current_ts;
next_tmp_phy_queue->proportion_ts_ = current_ts;
}
if (0 == deadline_ts) {
deadline_ts = next_tmp_phy_queue->reservation_ts_;
} else {

28
src/share/io/ob_io_manager.cpp
View File

@ -563,6 +563,28 @@ int ObIOManager::refresh_tenant_io_config(const uint64_t tenant_id, const ObTena
return ret;
}
// for unittest
int ObIOManager::modify_group_io_config(const uint64_t tenant_id,
const uint64_t index,
const int64_t min_percent,
const int64_t max_percent,
const int64_t weight_percent)
{
int ret = OB_SUCCESS;
ObRefHolder<ObTenantIOManager> tenant_holder;
if (OB_UNLIKELY(!is_inited_)) {
ret = OB_NOT_INIT;
LOG_WARN("not init", K(ret), K(is_inited_));
} else if (OB_FAIL(get_tenant_io_manager(tenant_id, tenant_holder))) {
LOG_WARN("get tenant io manager failed", K(ret), K(tenant_id));
} else if (OB_FAIL(tenant_holder.get_ptr()->modify_group_io_config(index, min_percent, max_percent, weight_percent, false, false))) {
LOG_WARN("update tenant io config failed", K(ret), K(tenant_id), K(min_percent), K(max_percent), K(weight_percent));
} else if (OB_FAIL(tenant_holder.get_ptr()->refresh_group_io_config())) {
LOG_WARN("fail to refresh group config", K(ret));
}
return ret;
}
int ObIOManager::get_tenant_io_manager(const uint64_t tenant_id, ObRefHolder<ObTenantIOManager> &tenant_holder)
{
int ret = OB_SUCCESS;
@ -1368,7 +1390,7 @@ int ObTenantIOManager::reset_all_group_config()
for (int64_t i = 0; i < io_config_.group_num_; ++i) {
if(io_config_.group_configs_.at(i).deleted_) {
//do nothing
} else if (OB_FAIL(modify_group_io_config(i, 0, 0, 0, false, true/*cleared*/))) {
} else if (OB_FAIL(modify_group_io_config(i, 0, 100, 0, false, true/*cleared*/))) {
LOG_WARN("modify group io config failed", K(ret), K(i));
}
}
@ -1409,7 +1431,7 @@ int ObTenantIOManager::reset_consumer_group_config(const int64_t group_id)
} else if (OB_UNLIKELY(index == INT64_MAX || index < 0)) {
ret = OB_ERR_UNEXPECTED;
LOG_WARN("invalid index, maybe try to reset OTHER_GROUPS or deleted_groups", K(ret), K(index), K(group_id));
} else if (OB_FAIL(modify_group_io_config(index, 0, 0, 0, false, true/*cleared*/))) {
} else if (OB_FAIL(modify_group_io_config(index, 0, 100, 0, false, true/*cleared*/))) {
LOG_WARN("modify group io config failed", K(ret), K(tenant_id_), K(index));
} else {
LOG_INFO ("stop group io control success when delete directive", K(tenant_id_), K(group_id), K(index), K(io_config_));
@ -1458,7 +1480,7 @@ int ObTenantIOManager::delete_consumer_group_config(const int64_t group_id)
} else {
if (OB_FAIL(group_id_index_map_.set_refactored(group_id, INT64_MAX, 1))) { //使用非法值覆盖
LOG_WARN("stop phy queues failed", K(ret), K(tenant_id_), K(index));
} else if (OB_FAIL(modify_group_io_config(index, 0, 0, 0, true/*deleted*/, false))) {
} else if (OB_FAIL(modify_group_io_config(index, 0, 100, 0, true/*deleted*/, false))) {
LOG_WARN("modify group io config failed", K(ret), K(tenant_id_), K(index));
}
}

5
src/share/io/ob_io_manager.h
View File

@ -73,6 +73,11 @@ public:
int refresh_tenant_io_config(const uint64_t tenant_id, const ObTenantIOConfig &tenant_io_config);
int get_tenant_io_manager(const uint64_t tenant_id, ObRefHolder<ObTenantIOManager> &tenant_holder);
int get_tenant_ids(ObIArray<uint64_t> &tenant_ids);
int modify_group_io_config(const uint64_t tenant_id,
const uint64_t index,
const int64_t min_percent,
const int64_t max_percent,
const int64_t weight_percent);
ObIOScheduler *get_scheduler();
private:

174
unittest/storage/test_io_manager.cpp
View File

@ -1181,6 +1181,25 @@ public:
ObIOFd fd_;
};
class IOGroupModify : public ThreadPool
{
public:
IOGroupModify()
: modify_init_ts_(0)
{}
int init(int64_t modify_init_ts, int64_t modify_delay_ts, const IOPerfTenant &curr_tenant);
void destroy();
virtual void run1() override;
TO_STRING_KV(K(load_), K(modify_delay_ts_), K(fd_), K(curr_tenant_));
public:
int64_t modify_init_ts_;
int64_t modify_delay_ts_;
IOPerfTenant curr_tenant_;
ObConcurrentFIFOAllocator allocator_;
IOPerfLoad load_;
ObIOFd fd_;
};
class IOTracerSwitch : public ThreadPool
{
public:
@ -1695,6 +1714,97 @@ TEST_F(TestIOManager, ModifyCallbackThread)
LOG_INFO("modify callback thread finished");
}
TEST_F(TestIOManager, ModifyGroupIO)
{
// use multi thread to do modify group_io_config
bool is_perf_config_exist = false;
ASSERT_SUCC(FileDirectoryUtils::is_exists(GROUP_PERF_CONFIG_FILE, is_perf_config_exist));
if (!is_perf_config_exist) {
write_group_perf_config();
}
// parse configs
IOPerfScheduler scheduler_config;
ObArray<IOPerfDevice> perf_devices;
ObArray<IOPerfTenant> perf_tenants;
ObArray<IOPerfLoad> perf_loads;
ASSERT_SUCC(parse_group_perf_config(GROUP_PERF_CONFIG_FILE, scheduler_config, perf_devices, perf_tenants, perf_loads));
ASSERT_TRUE(perf_devices.count() > 0);
ASSERT_TRUE(perf_tenants.count() > 0);
ASSERT_TRUE(perf_loads.count() > 0);
ObIOManager::get_instance().destroy();
const int64_t memory_limit = 30L * 1024L * 1024L * 1024L; // 30GB
const int64_t queue_depth = 100L;
ASSERT_SUCC(ObIOManager::get_instance().init(memory_limit, queue_depth, scheduler_config.sender_count_, scheduler_config.schedule_media_id_));
ASSERT_SUCC(ObIOManager::get_instance().start());
// prepare devices and files
char *device_buf = (char *)malloc(sizeof(ObLocalDevice) * perf_devices.count());
ASSERT_TRUE(nullptr != device_buf);
for (int64_t i = 0; i < perf_devices.count(); ++i) {
IOPerfDevice &curr_config = perf_devices.at(i);
ASSERT_SUCC(prepare_file(curr_config.file_path_, curr_config.file_size_, curr_config.fd_));
ObLocalDevice *device = new (device_buf + sizeof(ObLocalDevice) * i) ObLocalDevice;
ASSERT_SUCC(init_device(curr_config.media_id_, *device));
ASSERT_SUCC(OB_IO_MANAGER.add_device_channel(device, curr_config.async_channel_count_, curr_config.sync_channel_count_, curr_config.max_io_depth_));
curr_config.device_handle_ = device;
}
// prepare tenant io manager
for (int64_t i = 0; i < perf_tenants.count(); ++i) {
IOPerfTenant &curr_config = perf_tenants.at(i);
if (curr_config.tenant_id_ == 1002) {
LOG_INFO("qilu: tenant config", K(curr_config), K(i));
ASSERT_SUCC(OB_IO_MANAGER.add_tenant_io_manager(curr_config.tenant_id_, curr_config.config_));
ObRefHolder<ObTenantIOManager> tenant_holder;
ASSERT_SUCC(OB_IO_MANAGER.get_tenant_io_manager(curr_config.tenant_id_, tenant_holder));
ASSERT_SUCC(tenant_holder.get_ptr()->refresh_group_io_config());
}
}
// prepare perf runners
char *runner_buf = (char *)malloc(perf_loads.count() * sizeof(IOPerfRunner));
char *modifyer_buf = (char *)malloc(perf_loads.count() * sizeof(IOGroupModify));
ObArray<IOPerfRunner *> runners;
ObArray<IOGroupModify *> modifyers;
const int64_t start_ts = ObTimeUtility::current_time() + 10000L;
for (int64_t i = 0; i < perf_loads.count(); ++i) {
IOPerfRunner *runner = new (runner_buf + i * sizeof(IOPerfRunner)) IOPerfRunner();
const IOPerfLoad &cur_load = perf_loads.at(i);
if (cur_load.tenant_id_ == 1002) {
ASSERT_SUCC(runner->init(start_ts, cur_load));
ASSERT_SUCC(runners.push_back(runner));
LOG_INFO("runner start now");
}
}
//prepare modifyer
for (int64_t i = 0; i < perf_tenants.count(); ++i) {
IOPerfTenant &curr_tenant = perf_tenants.at(i);
if (curr_tenant.tenant_id_ == 1002) {
IOGroupModify *modifyer=new (modifyer_buf + i * sizeof(IOGroupModify)) IOGroupModify();
int64_t modify_init_ts = start_ts;
int64_t modify_delay_ts = 3000000L; //3s后开始修改
ASSERT_SUCC(modifyer->init(modify_init_ts, modify_delay_ts, curr_tenant));
ASSERT_SUCC(modifyers.push_back(modifyer));
}
}
// wait perf finished
for (int64_t i = 0; i < runners.count(); ++i) {
IOPerfRunner *runner = runners.at(i);
runner->wait();
ASSERT_SUCC(runner->print_result());
runner->destroy();
}
free(runner_buf);
free(modifyer_buf);
ObIOManager::get_instance().stop();
ObIOManager::get_instance().destroy();
for (int64_t i = 0; i < perf_devices.count(); ++i) {
ObLocalDevice *device_handle = perf_devices.at(i).device_handle_;
// ASSERT_SUCC(OB_IO_MANAGER.remove_device_channel(device_handle));
device_handle->destroy();
}
free(device_buf);
LOG_INFO("qilu: modify group finished");
}
TEST_F(TestIOManager, abnormal)
{
// simulate submit failure
@ -1783,13 +1893,14 @@ void write_group_perf_config()
"\n"
"tenant_id min_iops max_iops weight group\n"
"1001 5000 100000 700 10001: testgroup1: 80, 100, 60; 10002: testgroup2: 10, 60, 30; 0: OTHER_GROUPS: 10, 100, 10;\n"
"1002 1000 50000 1000 0: testgroup1: 100, 100, 100;\n"
"1002 1000 50000 1000 12345: testgroup1: 50, 50, 50; 0: OTHER_GROUPS: 50, 50, 50;\n"
"\n"
"tenant_id device_id group io_mode io_size_byte io_depth perf_mode target_iops thread_count is_sequence start_s stop_s\n"
"1001 1 0 r 16384 10 rolling 0 16 0 0 8\n"
"1001 1 10001 r 16384 10 rolling 0 16 0 2 7\n"
"1001 1 10002 r 16384 10 rolling 0 16 0 0 6\n"
"1002 1 0 r 16384 10 rolling 0 16 0 0 8\n"
"1002 1 0 r 16384 100 rolling 0 16 0 0 10\n"
"1002 1 12345 r 16384 100 rolling 0 16 0 0 10\n"
;
const int64_t file_len = strlen(file_buf);
int write_ret = ::write(fd, file_buf, file_len);
@ -2344,6 +2455,65 @@ int IOConfModify::modify_tenant_io( const int64_t min_iops,
return ret;
}
int IOGroupModify::init(int64_t modify_init_ts, int64_t modify_delay_ts, const IOPerfTenant &curr_tenant)
{
int ret = OB_SUCCESS;
if (OB_UNLIKELY(!curr_tenant.is_valid())) {
ret = OB_INVALID_ARGUMENT;
LOG_WARN("invalid argument", K(ret), K(curr_tenant));
} else if (OB_FAIL(allocator_.init(OB_MALLOC_BIG_BLOCK_SIZE, "group modifier", OB_SERVER_TENANT_ID, 1024L * 1024L * 1024L * 10L))) {
LOG_WARN("init allocator failed", K(ret));
} else {
curr_tenant_ = curr_tenant;
modify_init_ts_ = modify_init_ts;
modify_delay_ts_ = modify_delay_ts;
}
if (OB_SUCC(ret)) {
if (OB_FAIL(set_thread_count(load_.thread_count_ + 1))) {
LOG_WARN("set thread count failed", K(ret), K(modify_init_ts_), K(curr_tenant_));
} else if (OB_FAIL(start())) {
LOG_WARN("start thread failed", K(ret), K(modify_init_ts_), K(curr_tenant_));
}
}
if (OB_FAIL(ret)) {
destroy();
}
return ret;
}
void IOGroupModify::destroy()
{
stop();
wait();
curr_tenant_ = IOPerfTenant();
}
void IOGroupModify::run1()
{
int ret = OB_SUCCESS;
const int64_t thread_idx = get_thread_idx();
LOG_INFO("modify thread start");
//change 1
int64_t current_ts = ObTimeUtility::current_time();
if (modify_init_ts_ + modify_delay_ts_ > current_ts) {
usleep(modify_init_ts_ + modify_delay_ts_ - current_ts);
}
if (OB_FAIL(OB_IO_MANAGER.modify_group_io_config(curr_tenant_.tenant_id_, INT64_MAX, 100, 100, 100))) {
LOG_WARN("fail to modify group config", K(ret));
} else if (OB_FAIL(OB_IO_MANAGER.modify_group_io_config(curr_tenant_.tenant_id_, 0, 0, 0, 0))) {
LOG_WARN("fail to modify group config", K(ret));
}
//change 2
usleep(3000L * 1000L); // sleep 3s
if (OB_FAIL(OB_IO_MANAGER.modify_group_io_config(curr_tenant_.tenant_id_, INT64_MAX, 40, 40, 40))) {
LOG_WARN("fail to modify group config", K(ret));
} else if (OB_FAIL(OB_IO_MANAGER.modify_group_io_config(curr_tenant_.tenant_id_, 0, 60, 60, 60))) {
LOG_WARN("fail to modify group config", K(ret));
}
}
int IOTracerSwitch::init(int64_t switch_init_ts, int64_t switch_delay_ts, const IOPerfTenant &curr_tenant)
{
int ret = OB_SUCCESS;