Skip to content

Commit

Permalink
feat(tianmu) Refactor the aggregation and The IN operator to multithr…
Browse files Browse the repository at this point in the history 
…ead stoneatom#472 (stoneatom#422 stoneatom#465 stoneatom#466) (stoneatom#472)

* Refactor the aggregation to multithread

Co-authored-by: hustjieke <gaoriyao1@gmail.com>
  • Loading branch information
@adofsauron @hustjieke
adofsauron and hustjieke authored Sep 17, 2022
1 parent e3a9902 commit 208f79d
Show file tree
Hide file tree
Showing 15 changed files with 445 additions and 138 deletions.
8 changes: 8 additions & 0 deletions sql/table.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3052,6 +3052,14 @@ inline bool is_perfschema_db(const char *name)
*/
inline bool belongs_to_p_s(TABLE_LIST *tl)
{
if (!tl->db) {
return false;
}

if (!tl->table_name) {
return false;
}

return (!strcmp("performance_schema", tl->db) &&
strcmp(tl->table_name, "threads") &&
strstr(tl->table_name, "setup_") == NULL);
Expand Down
172 changes: 97 additions & 75 deletions storage/tianmu/core/aggregation_algorithm.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -194,7 +194,7 @@ void AggregationAlgorithm::Aggregate(bool just_distinct, int64_t &limit, int64_t
}
} else {
int64_t local_limit = limit == -1 ? upper_approx_of_groups : limit;
MultiDimensionalGroupByScan(gbw, local_limit, offset, sender, limit_less_than_no_groups);
MultiDimensionalGroupByScan(gbw, local_limit, offset, sender, limit_less_than_no_groups, true);
if (limit != -1) limit = local_limit;
}
t->ClearMultiIndexP(); // cleanup (i.e. regarded as materialized,
Expand All @@ -205,7 +205,8 @@ void AggregationAlgorithm::Aggregate(bool just_distinct, int64_t &limit, int64_t
}

void AggregationAlgorithm::MultiDimensionalGroupByScan(GroupByWrapper &gbw, int64_t &limit, int64_t &offset,
ResultSender *sender, bool limit_less_than_no_groups) {
ResultSender *sender, bool limit_less_than_no_groups,
bool force_parall) {
MEASURE_FET("TempTable::MultiDimensionalGroupByScan(...)");
bool first_pass = true;
// tuples are numbered according to tuple_left filter (not used, if tuple_left
Expand Down Expand Up @@ -234,9 +235,19 @@ void AggregationAlgorithm::MultiDimensionalGroupByScan(GroupByWrapper &gbw, int6
}
gbw.SetDistinctTuples(mit.NumOfTuples());

auto get_thd_cnt = []() {
int hardware_concurrency = std::thread::hardware_concurrency();
// TODO: The original code was the number of CPU cores divided by 4, and the reason for that is to be traced further
return hardware_concurrency > 4 ? (hardware_concurrency / 4) : 1;
};

int thd_cnt = 1;
if (ParallelAllowed(gbw) && !limit_less_than_no_groups) {
thd_cnt = std::thread::hardware_concurrency() / 4; // For concurrence reason, don't swallow all cores once.
if (force_parall) {
thd_cnt = get_thd_cnt();
} else {
if (ParallelAllowed(gbw) && !limit_less_than_no_groups) {
thd_cnt = get_thd_cnt(); // For concurrence reason, don't swallow all cores once.
}
}

AggregationWorkerEnt ag_worker(gbw, mind, thd_cnt, this);
Expand Down Expand Up @@ -278,13 +289,13 @@ void AggregationAlgorithm::MultiDimensionalGroupByScan(GroupByWrapper &gbw, int6

// Grouping on a packrow
int64_t packrow_length = mit.GetPackSizeLeft();
int grouping_result = AggregatePackrow(gbw, &mit, cur_tuple);
AggregaGroupingResult grouping_result = AggregatePackrow(gbw, &mit, cur_tuple);
if (sender) {
sender->SetAffectRows(gbw.NumOfGroups());
}
if (grouping_result == 2) throw common::KilledException();
if (grouping_result != 5) packrows_found++; // for statistics
if (grouping_result == 1) break; // end of the aggregation
if (grouping_result == AggregaGroupingResult::AGR_KILLED) throw common::KilledException();
if (grouping_result != AggregaGroupingResult::AGR_NO_LEFT) packrows_found++; // for statistics
if (grouping_result == AggregaGroupingResult::AGR_FINISH) break; // end of the aggregation
if (!gbw.IsFull() && gbw.MemoryBlocksLeft() == 0) {
gbw.SetAsFull();
}
Expand Down Expand Up @@ -361,7 +372,7 @@ void AggregationAlgorithm::MultiDimensionalGroupByScan(GroupByWrapper &gbw, int6
displayed_no_groups = t->NumOfObj();
if (t->NumOfObj() >= limit) break;
if (gbw.AnyTuplesLeft()) gbw.ClearUsed(); // prepare for the next pass, if needed
} while (gbw.AnyTuplesLeft()); // do the next pass, if anything left
} while (gbw.AnyTuplesLeft() && (1 == thd_cnt)); // do the next pass, if anything left
} catch (...) {
ag_worker.Commit(false);
throw;
Expand Down Expand Up @@ -468,11 +479,12 @@ void AggregationAlgorithm::MultiDimensionalDistinctScan(GroupByWrapper &gbw, MII
}
}

int AggregationAlgorithm::AggregatePackrow(GroupByWrapper &gbw, MIIterator *mit, int64_t cur_tuple) {
AggregaGroupingResult AggregationAlgorithm::AggregatePackrow(GroupByWrapper &gbw, MIIterator *mit, int64_t cur_tuple) {
std::scoped_lock guard(mtx);
int64_t packrow_length = mit->GetPackSizeLeft();
if (!gbw.AnyTuplesLeft(cur_tuple, cur_tuple + packrow_length - 1)) {
mit->NextPackrow();
return 5;
return AggregaGroupingResult::AGR_NO_LEFT;
}
int64_t uniform_pos = common::NULL_VALUE_64;
bool skip_packrow = false;
Expand Down Expand Up @@ -509,7 +521,7 @@ int AggregationAlgorithm::AggregatePackrow(GroupByWrapper &gbw, MIIterator *mit,
if (gbw.NumOfGroupingAttrs() == gbw.NumOfAttrs() // just DISTINCT without grouping
|| stop_all) { // or aggregation already done on rough level
gbw.TuplesResetAll(); // no more rows needed, just produce output
return 1; // aggregation finished
return AggregaGroupingResult::AGR_FINISH; // aggregation finished
}
}
if (skip_packrow)
Expand All @@ -522,15 +534,15 @@ int AggregationAlgorithm::AggregatePackrow(GroupByWrapper &gbw, MIIterator *mit,

if (packrow_done || skip_packrow) {
mit->NextPackrow();
return 0; // success - roughly omitted
return AggregaGroupingResult::AGR_OK; // success - roughly omitted
}

// bool require_locking_ag = true; // a new packrow,
// so locking will be needed bool require_locking_gr = (uniform_pos ==
// common::NULL_VALUE_64); // do not lock if the grouping row is uniform

while (mit->IsValid()) { // becomes invalid on pack end
if (m_conn->Killed()) return 2; // killed
if (m_conn->Killed()) return AggregaGroupingResult::AGR_KILLED; // killed
if (gbw.TuplesGet(cur_tuple)) {
if (require_locking_gr) {
for (int gr_a = 0; gr_a < gbw.NumOfGroupingAttrs(); gr_a++)
Expand Down Expand Up @@ -560,7 +572,7 @@ int AggregationAlgorithm::AggregatePackrow(GroupByWrapper &gbw, MIIterator *mit,
gbw.SetAllGroupsFound();
if (gbw.NumOfGroupingAttrs() == gbw.NumOfAttrs()) { // just DISTINCT without grouping
gbw.TuplesResetAll(); // no more rows needed, just produce output
return 1; // aggregation finished
return AggregaGroupingResult::AGR_FINISH; // aggregation finished
}
}
}
Expand All @@ -577,7 +589,9 @@ int AggregationAlgorithm::AggregatePackrow(GroupByWrapper &gbw, MIIterator *mit,
for (int gr_a = gbw.NumOfGroupingAttrs(); gr_a < gbw.NumOfAttrs(); gr_a++)
if (gbw.ColumnNotOmitted(gr_a)) {
bool value_successfully_aggregated = gbw.PutAggregatedValue(gr_a, pos, *mit, factor);
if (!value_successfully_aggregated) gbw.DistinctlyOmitted(gr_a, cur_tuple);
if (!value_successfully_aggregated) {
gbw.DistinctlyOmitted(gr_a, cur_tuple);
}
}
}
}
Expand All @@ -587,7 +601,7 @@ int AggregationAlgorithm::AggregatePackrow(GroupByWrapper &gbw, MIIterator *mit,
if (mit->PackrowStarted()) break;
}
gbw.CommitResets();
return 0; // success
return AggregaGroupingResult::AGR_OK; // success
}

void AggregationAlgorithm::AggregateFillOutput(GroupByWrapper &gbw, int64_t gt_pos, int64_t &omit_by_offset) {
Expand Down Expand Up @@ -813,23 +827,27 @@ void AggregationAlgorithm::TaskFillOutput(GroupByWrapper *gbw, Transaction *ci,
}
}

void AggregationWorkerEnt::TaskAggrePacks(MIUpdatingIterator *taskIterator, [[maybe_unused]] DimensionVector *dims,
[[maybe_unused]] MIIterator *mit, [[maybe_unused]] int pstart,
[[maybe_unused]] int pend, int tuple, GroupByWrapper *gbw, Transaction *ci) {
int i = 0;
int64_t cur_tuple = tuple;
common::SetMySQLTHD(ci->Thd());
current_txn_ = ci;
void AggregationWorkerEnt::TaskAggrePacks(MIIterator *taskIterator, [[maybe_unused]] DimensionVector *dims,
[[maybe_unused]] MIIterator *mit, [[maybe_unused]] CTask *task,
GroupByWrapper *gbw, Transaction *ci) {
taskIterator->Rewind();
int task_pack_num = 0;
while (taskIterator->IsValid()) {
int64_t packrow_length = taskIterator->GetPackSizeLeft();
int grouping_result = aa->AggregatePackrow(*gbw, taskIterator, cur_tuple);
if (grouping_result != 5) i++;
if (grouping_result == 1) break;
if (grouping_result == 2) throw common::KilledException();
if (grouping_result == 3 || grouping_result == 4) throw common::NotImplementedException("Aggregation overflow.");
cur_tuple += packrow_length;
}
TIANMU_LOG(LogCtl_Level::DEBUG, "TaskAggrePacks routine ends. Task id %d", taskIterator->GetTaskNum());
if ((task_pack_num >= task->dwStartPackno) && (task_pack_num <= task->dwEndPackno)) {

int cur_tuple = (*task->dwPack2cur)[task_pack_num];
MIInpackIterator mii(*taskIterator);
AggregaGroupingResult grouping_result = aa->AggregatePackrow(*gbw, &mii, cur_tuple);
if (grouping_result == AggregaGroupingResult::AGR_FINISH) break;
if (grouping_result == AggregaGroupingResult::AGR_KILLED) throw common::KilledException();
if (grouping_result == AggregaGroupingResult::AGR_OVERFLOW ||
grouping_result == AggregaGroupingResult::AGR_OTHER_ERROR)
throw common::NotImplementedException("Aggregation overflow.");
}

taskIterator->NextPackrow();
++task_pack_num;
}
}

void AggregationWorkerEnt::PrepShardingCopy(MIIterator *mit, GroupByWrapper *gb_sharding,
Expand All @@ -852,83 +870,87 @@ void AggregationWorkerEnt::DistributeAggreTaskAverage(MIIterator &mit) {
std::vector<std::unique_ptr<GroupByWrapper>> vGBW;
vGBW.reserve(m_threads);
vTask.reserve(m_threads);
if (rc_control_.isOn()) rc_control_.lock(conn->GetThreadID()) << "Prepare data for parallel aggreation" << system::unlock;
if (rc_control_.isOn())
rc_control_.lock(conn->GetThreadID()) << "Prepare data for parallel aggreation" << system::unlock;

int packnum = 0;
int curtuple_index = 0;
std::unordered_map<int, int> pack2cur;
while (mit.IsValid()) {
pack2cur.emplace(std::pair<int, int>(packnum, curtuple_index));

int64_t packrow_length = mit.GetPackSizeLeft();
curtuple_index += packrow_length;
packnum++;
mit.NextPackrow();
}

int loopcnt = (packnum < m_threads) ? packnum : m_threads;
pack2cur.emplace(std::pair<int, int>(packnum, curtuple_index));

int loopcnt = (packnum < m_threads) ? packnum : m_threads;
int mod = packnum % loopcnt;
int num = packnum / loopcnt;

utils::result_set<void> res;
for (int i = 0; i < loopcnt; ++i) {
res.insert(ha_rcengine_->query_thread_pool.add_task(&AggregationWorkerEnt::PrepShardingCopy, this, &mit, gb_main, &vGBW));
res.insert(
ha_rcengine_->query_thread_pool.add_task(&AggregationWorkerEnt::PrepShardingCopy, this, &mit, gb_main, &vGBW));

int pack_start = i * num;
int pack_end = 0;
int dwPackNum = 0;
if (i == (loopcnt - 1)) {
pack_end = packnum;
dwPackNum = packnum;
} else {
pack_end = (i + 1) * num - 1;
dwPackNum = pack_end + 1;
}

int cur_start = pack2cur[pack_start];
int cur_end = pack2cur[pack_end] - 1;

CTask tmp;
tmp.dwTaskId = i;
tmp.dwPackNum = num + mod + i * num;
tmp.dwPackNum = dwPackNum;
tmp.dwStartPackno = pack_start;
tmp.dwEndPackno = pack_end;
tmp.dwStartTuple = cur_start;
tmp.dwEndTuple = cur_end;
tmp.dwTuple = cur_start;
tmp.dwPack2cur = &pack2cur;

vTask.push_back(tmp);
}
res.get_all_with_except();

if (rc_control_.isOn())
rc_control_.lock(conn->GetThreadID()) << "Prepare data for parallel aggreation done. Total packnum " << packnum
<< system::unlock;
packnum = 0;
mit.Rewind();
int curtuple = 0;
while (mit.IsValid()) {
int64_t packrow_length = mit.GetPackSizeLeft();
packnum++;
curtuple += packrow_length;
for (auto &it : vTask) {
if (packnum == it.dwPackNum) {
it.dwEndPackno = mit.GetCurPackrow(0);
it.dwTuple = curtuple;
}
}

if (rc_control_.isOn())
rc_control_.lock(conn->GetThreadID()) << " GetCurPackrow: " << mit.GetCurPackrow(0) << " packnum: " << packnum
<< " cur_tuple: " << curtuple << system::unlock;

mit.NextPackrow();
}

std::vector<MultiIndex> mis;
mis.reserve(vTask.size());

std::vector<MIUpdatingIterator> taskIterator;
std::vector<MIIterator> taskIterator;
taskIterator.reserve(vTask.size());

utils::result_set<void> res1;
for (uint i = 0; i < vTask.size(); ++i) {
auto &mi = mis.emplace_back(*mind, true);
if (dims.NoDimsUsed() == 0) dims.SetAll();
auto &mii = taskIterator.emplace_back(&mi, dims);
auto &mii = taskIterator.emplace_back(mit, true);
mii.SetTaskNum(vTask.size());
mii.SetTaskId(i);
mii.SetNoPacksToGo(vTask[i].dwEndPackno);
mii.RewindToPack((i == 0) ? 0 : vTask[i - 1].dwEndPackno + 1);
}

res1.insert(ha_rcengine_->query_thread_pool.add_task(&AggregationWorkerEnt::TaskAggrePacks, this, &taskIterator[0], &dims,
&mit, 0, vTask[0].dwEndPackno, 0, gb_main, conn));
for (size_t i = 1; i < vTask.size(); ++i) {
res1.insert(ha_rcengine_->query_thread_pool.add_task(&AggregationWorkerEnt::TaskAggrePacks, this, &taskIterator[i], &dims,
&mit, vTask[i - 1].dwEndPackno + 1, vTask[i].dwEndPackno,
vTask[i - 1].dwTuple, vGBW[i].get(), conn));
for (size_t i = 0; i < vTask.size(); ++i) {
GroupByWrapper *gbw = i == 0 ? gb_main : vGBW[i].get();
res1.insert(ha_rcengine_->query_thread_pool.add_task(&AggregationWorkerEnt::TaskAggrePacks, this, &taskIterator[i],
&dims, &mit, &vTask[i], gbw, conn));
}
res1.get_all_with_except();

for (size_t i = 0; i < vTask.size(); ++i) {
// Merge aggreation data together
if (i != 0) gb_main->Merge(*(vGBW[i]));
if (i != 0) {
aa->MultiDimensionalDistinctScan(*(vGBW[i]), mit);
gb_main->Merge(*(vGBW[i]));
}
}
}

} // namespace core
} // namespace Tianmu
22 changes: 16 additions & 6 deletions storage/tianmu/core/aggregation_algorithm.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,9 +23,19 @@
#include "core/mi_iterator.h"
#include "core/query.h"
#include "core/temp_table.h"
#include "core/ctask.h"

namespace Tianmu {
namespace core {
enum class AggregaGroupingResult {
AGR_OK = 0, // success
AGR_FINISH = 1, // finish
AGR_KILLED = 2, // killed
AGR_OVERFLOW = 3, // overflow
AGR_OTHER_ERROR = 4, // other error
AGR_NO_LEFT = 5 // pack already aggregated
};

class AggregationAlgorithm {
public:
AggregationAlgorithm(TempTable *tt)
Expand All @@ -37,13 +47,13 @@ class AggregationAlgorithm {
bool &ag_not_changeabe, bool &stop_all, int64_t &uniform_pos, int64_t rows_in_pack,
int64_t local_factor, int just_one_aggr = -1);
void MultiDimensionalGroupByScan(GroupByWrapper &gbw, int64_t &limit, int64_t &offset, ResultSender *sender,
bool limit_less_than_no_groups);
bool limit_less_than_no_groups, bool force_parall);
void MultiDimensionalDistinctScan(GroupByWrapper &gbw, MIIterator &mit);
void AggregateFillOutput(GroupByWrapper &gbw, int64_t gt_pos, int64_t &omit_by_offset);

// Return code for AggregatePackrow: 0 - success, 1 - stop aggregation
// (finished), 5 - pack already aggregated (skip)
int AggregatePackrow(GroupByWrapper &gbw, MIIterator *mit, int64_t cur_tuple);
AggregaGroupingResult AggregatePackrow(GroupByWrapper &gbw, MIIterator *mit, int64_t cur_tuple);

// No parallel for subquery/join/distinct cases
bool ParallelAllowed(GroupByWrapper &gbw) {
Expand Down Expand Up @@ -77,18 +87,18 @@ class AggregationWorkerEnt {
// Return code for AggregatePackrow: 0 - success, 1 - stop aggregation
// (finished), 2 - killed, 3
// - overflow, 4 - other error, 5 - pack already aggregated (skip)
int AggregatePackrow(MIUpdatingIterator &lmit, int64_t cur_tuple) {
AggregaGroupingResult AggregatePackrow(MIIterator &lmit, int64_t cur_tuple) {
return aa->AggregatePackrow(*gb_main, &lmit, cur_tuple);
}
int AggregatePackrow(MIInpackIterator &lmit, int64_t cur_tuple) {
AggregaGroupingResult AggregatePackrow(MIInpackIterator &lmit, int64_t cur_tuple) {
return aa->AggregatePackrow(*gb_main, &lmit, cur_tuple);
}
void Commit([[maybe_unused]] bool do_merge = true) { gb_main->CommitResets(); }
void ReevaluateNumberOfThreads([[maybe_unused]] MIIterator &mit) {}
int ThreadsUsed() { return m_threads; }
void Barrier() {}
void TaskAggrePacks(MIUpdatingIterator *taskIterator, DimensionVector *dims, MIIterator *mit, int pstart, int pend,
int tuple, GroupByWrapper *gbw, Transaction *ci);
void TaskAggrePacks(MIIterator *taskIterator, DimensionVector *dims, MIIterator *mit, CTask *task,
GroupByWrapper *gbw, Transaction *ci);
void DistributeAggreTaskAverage(MIIterator &mit);
void PrepShardingCopy(MIIterator *mit, GroupByWrapper *gb_sharding,
std::vector<std::unique_ptr<GroupByWrapper>> *vGBW);
Expand Down
Loading

0 comments on commit 208f79d

Please sign in to comment.