Skip to content

las3_pub / predictable_parallel_patterns

Go to a project
Commit c6dd2fc0 by FritzFlorian
Options

WIP: Partly functional version. Stealing and continuation tarding works 'most' of the time. 

The main issue seems to still be the fact that we have a lock free protocol where a steal can be pending. We plan to remove this fact next by introducing a protocol that works on a single atomic update.
parent adf05e9a
Pipeline #1340 failed with stages
in 40 seconds
Inline Side-by-side
Showing with 98 additions and 137 deletions
app/benchmark_fft/main.cpp
......@@ -8,8 +8,6 @@ using namespace pls::internal::scheduling;
#include <vector>
#include <atomic>
std::atomic<unsigned long> count;
static constexpr int CUTOFF = 16;
static constexpr int INPUT_SIZE = 8192;
typedef std::vector<std::complex<double>> complex_vector;
......@@ -44,7 +42,6 @@ void combine(complex_vector::iterator data, int n) {
void fft_normal(complex_vector::iterator data, int n) {
if (n < 2) {
// count++;
return;
}
......@@ -56,14 +53,15 @@ void fft_normal(complex_vector::iterator data, int n) {
parallel_result<short> fft(complex_vector::iterator data, int n) {
if (n < 2) {
return 0;
return parallel_result<short>{0};
}
divide(data, n);
if (n <= CUTOFF) {
fft_normal(data, n / 2);
fft_normal(data + n / 2, n / 2);
return 0;
combine(data, n);
return parallel_result<short>{0};
} else {
return scheduler::par([=]() {
return fft(data, n / 2);
......@@ -71,7 +69,7 @@ parallel_result<short> fft(complex_vector::iterator data, int n) {
return fft(data + n / 2, n / 2);
}).then([=](int, int) {
combine(data, n);
return 0;
return parallel_result<short>{0};
});
}
}
......@@ -93,13 +91,13 @@ complex_vector prepare_input(int input_size) {
}
static constexpr int NUM_ITERATIONS = 1000;
constexpr size_t NUM_THREADS = 1;
constexpr size_t NUM_THREADS = 2;
constexpr size_t NUM_TASKS = 64;
constexpr size_t NUM_TASKS = 128;
constexpr size_t MAX_TASK_STACK_SIZE = 0;
constexpr size_t NUM_CONTS = 64;
constexpr size_t MAX_CONT_SIZE = 192;
constexpr size_t NUM_CONTS = 128;
constexpr size_t MAX_CONT_SIZE = 512;
int main() {
complex_vector initial_input = prepare_input(INPUT_SIZE);
......@@ -112,7 +110,6 @@ int main() {
scheduler scheduler{static_scheduler_memory, NUM_THREADS};
count.store(0);
auto start = std::chrono::steady_clock::now();
for (int i = 0; i < NUM_ITERATIONS; i++) {
complex_vector input_2(initial_input);
......@@ -122,23 +119,20 @@ int main() {
}, []() {
return parallel_result<int>{0};
}).then([](int, int) {
return 0;
return parallel_result<int>{0};
});
});
}
auto end = std::chrono::steady_clock::now();
std::cout << "Count: " << count.load() << std::endl;
std::cout << "Framework: " << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count()
<< std::endl;
count.store(0);
start = std::chrono::steady_clock::now();
for (int i = 0; i < NUM_ITERATIONS; i++) {
complex_vector input_1(initial_input);
fft_normal(input_1.begin(), INPUT_SIZE);
}
end = std::chrono::steady_clock::now();
std::cout << "Count: " << count.load() << std::endl;
std::cout << "Normal: " << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count()
<< std::endl;
......
app/playground/main.cpp
......@@ -8,18 +8,15 @@
using namespace pls::internal;
constexpr size_t NUM_THREADS = 2;
constexpr size_t NUM_THREADS = 1;
constexpr size_t NUM_TASKS = 64;
constexpr size_t NUM_TASKS = 128;
constexpr size_t MAX_TASK_STACK_SIZE = 0;
constexpr size_t NUM_CONTS = 64;
constexpr size_t NUM_CONTS = 128;
constexpr size_t MAX_CONT_SIZE = 256;
std::atomic<int> count{0};
scheduling::parallel_result<int> fib(int n) {
base::this_thread::sleep(100);
// std::cout << "Fib(" << n << "): " << count++ << ", " << scheduling::thread_state::get().get_id() << std::endl;
int fib_normal(int n) {
if (n == 0) {
return 0;
}
......@@ -27,20 +24,11 @@ scheduling::parallel_result<int> fib(int n) {
return 1;
}
return scheduling::scheduler::par([=]() {
return fib(n - 1);
}, [=]() {
return fib(n - 2);
}).then([=](int a, int b) {
scheduling::parallel_result<int> result{a + b};
base::this_thread::sleep(100);
// std::cout << "Done Fib(" << n << "): " << (a + b) << ", " << scheduling::thread_state::get().get_id() << std::endl;
return result;
});
int result = fib_normal(n - 1) + fib_normal(n - 2);
return result;
}
int fib_normal(int n) {
// std::cout << "Fib(" << n << "): " << count++ << std::endl;
scheduling::parallel_result<int> fib(int n) {
if (n == 0) {
return 0;
}
......@@ -48,9 +36,13 @@ int fib_normal(int n) {
return 1;
}
int result = fib_normal(n - 1) + fib_normal(n - 2);
// std::cout << "Done Fib(" << n << "): " << result << std::endl;
return result;
return scheduling::scheduler::par([=]() {
return fib(n - 1);
}, [=]() {
return fib(n - 2);
}).then([=](int a, int b) {
return scheduling::parallel_result<int>{a + b};
});
}
int main() {
......@@ -63,27 +55,26 @@ int main() {
scheduling::scheduler scheduler{static_scheduler_memory, NUM_THREADS};
auto start = std::chrono::steady_clock::now();
// std::cout << "fib = " << fib_normal(10) << std::endl;
std::cout << "fib = " << fib_normal(30) << std::endl;
auto end = std::chrono::steady_clock::now();
std::cout << "Normal: " << std::chrono::duration_cast<std::chrono::microseconds>(end - start).count()
std::cout << "Normal: " << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count()
<< std::endl;
start = std::chrono::steady_clock::now();
scheduler.perform_work([]() {
// return scheduling::scheduler::par([]() {
// return scheduling::parallel_result<int>(0);
// }, []() {
// return fib(16);
// }).then([](int, int b) {
// std::cout << "fib = " << (b) << std::endl;
// return scheduling::parallel_result<int>{0};
// });
return fib(10);
return scheduling::scheduler::par([]() {
return scheduling::parallel_result<int>(0);
}, []() {
return fib(30);
}).then([](int, int b) {
std::cout << "fib = " << b << std::endl;
return scheduling::parallel_result<int>{0};
});
});
end = std::chrono::steady_clock::now();
std::cout << "Framework: " << std::chrono::duration_cast<std::chrono::microseconds>(end - start).count() << std::endl;
std::cout << "Framework: " << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() << std::endl;
return 0;
}
lib/pls/include/pls/internal/scheduling/cont.h
......@@ -77,7 +77,7 @@ class cont : public base_cont {
using BASE_RES_TYPE = typename std::remove_cv<typename std::remove_reference<RES_TYPE>::type>::type;
static void execute(cont &cont) {
parallel_result<BASE_RES_TYPE> result{cont.function_((*cont.result_1_).value(), (*cont.result_2_).value())};
parallel_result<BASE_RES_TYPE> result{cont.function_((*cont.left_result_).value(), (*cont.right_result_).value())};
if (result.fast_path() && cont.parent_ != nullptr) {
if (cont.is_right_child()) {
cont.parent_->store_right_result(std::move(result));
......@@ -90,7 +90,7 @@ class cont : public base_cont {
template<typename INNER_TYPE>
struct result_runner<parallel_result<INNER_TYPE>> {
static void execute(cont &cont) {
auto result = cont.function_((*cont.result_1_).value(), (*cont.result_2_).value());
auto result = cont.function_((*cont.left_result_).value(), (*cont.right_result_).value());
if (result.fast_path() && cont.parent_) {
if (cont.is_right_child()) {
cont.parent_->store_right_result(std::move(result));
......@@ -113,7 +113,7 @@ class cont : public base_cont {
task_{std::forward<T2ARGS>(task_2_args)..., this} {};
void execute() override {
using result_type = decltype(function_((*result_1_).value(), (*result_2_).value()));
using result_type = decltype(function_((*left_result_).value(), (*right_result_).value()));
result_runner<result_type>::execute(*this);
this->get_memory_block()->free_buffer();
this->get_memory_block()->reset_state();
......@@ -124,12 +124,11 @@ class cont : public base_cont {
task_.execute();
}
void *get_right_result_pointer() override {
return &result_1_;
}
void *get_left_result_pointer() override {
return &result_2_;
return &left_result_;
}
void *get_right_result_pointer() override {
return &right_result_;
}
T2 *get_task() {
......@@ -143,8 +142,8 @@ class cont : public base_cont {
// Some fields/actual values stay uninitialized (save time on the fast path if we don not need them).
// More fields untouched on the fast path is good, but for ease of an implementation we only keep some for now.
delayed_init<R1> result_1_;
delayed_init<R2> result_2_;
delayed_init<R1> left_result_;
delayed_init<R2> right_result_;
};
}
......
lib/pls/include/pls/internal/scheduling/cont_manager.h
......@@ -25,12 +25,12 @@ class cont_manager {
: max_cont_size_{MAX_CONT_SIZE}, num_conts_{NUM_CONTS} {
// First node is currently active and our local start
start_node_ = active_node_ = init_memory_block<MAX_CONT_SIZE>(cont_storage, nullptr, nullptr, 0);
active_node_ = init_memory_block<MAX_CONT_SIZE>(cont_storage, nullptr, 0);
// Build up chain after it
memory_block *current_node = start_node_;
memory_block *current_node = active_node_;
for (size_t i = 1; i < NUM_CONTS; i++) {
memory_block *next_node = init_memory_block<MAX_CONT_SIZE>(cont_storage, start_node_, current_node, i);
memory_block *next_node = init_memory_block<MAX_CONT_SIZE>(cont_storage, current_node, i);
current_node->set_next(next_node);
current_node = next_node;
}
......@@ -38,16 +38,30 @@ class cont_manager {
// Aquire and release memory blocks...
memory_block *get_next_memory_block() {
active_node_->set_start(start_node_);
auto result = active_node_;
active_node_ = active_node_->get_next();
return result;
}
void return_memory_block() {
active_node_ = active_node_->get_prev();
}
void move_active_node(int depth) {
if (depth < 0) {
for (int i = 0; i < (depth * -1); i++) {
active_node_ = active_node_->get_prev();
}
} else {
for (int i = 0; i < depth; i++) {
active_node_ = active_node_->get_next();
}
}
}
void move_active_node_to_start() {
move_active_node(-1 * active_node_->get_depth());
}
memory_block *get_active_node() {
return active_node_;
}
// Manage the fall through behaviour/slow path behaviour
bool falling_through() const {
......@@ -60,44 +74,11 @@ class cont_manager {
}
void aquire_memory_chain(memory_block *target_chain) {
auto *our_next_node = get_node(target_chain->get_depth() + 1);
our_next_node->set_prev(target_chain);
target_chain->set_next(our_next_node);
start_node_ = target_chain->get_start();
active_node_ = target_chain;
}
memory_block *get_node(unsigned int depth) {
// TODO: Remove this O(n) factor to avoid the
// T_1/P + (T_lim + D) stealing time bound.
memory_block *current = start_node_;
for (unsigned int i = 0; i < depth; i++) {
current->set_start(start_node_);
current = current->get_next();
}
return current;
}
void check_clean_chain() {
memory_block *current = start_node_;
for (unsigned int i = 0; i < num_conts_; i++) {
bool buffer_used = current->is_buffer_used();
auto state_value = current->get_state().load().value;
if (state_value != memory_block::initialized || buffer_used || current->get_depth() != i) {
PLS_ASSERT(false,
"Must always steal with a clean chain!");
}
current->set_start(start_node_);
current = current->get_next();
}
}
PLS_ASSERT(active_node_->get_depth() == target_chain->get_depth() + 1,
"Can only steal aquire chain parts with correct depth.");
void set_active_depth(unsigned int depth) {
active_node_ = get_node(depth);
active_node_->set_prev(target_chain);
target_chain->set_next(active_node_);
}
void execute_fall_through_code() {
......@@ -109,8 +90,8 @@ class cont_manager {
auto *notified_cont = fall_through_cont_;
bool notifier_is_right_child = fall_through_child_right;
std::cout << "Notifying Cont on core " << my_state.get_id() << " and depth "
<< notified_cont->get_memory_block()->get_depth() << std::endl;
// std::cout << "Notifying Cont on core " << my_state.get_id() << " and depth "
// << notified_cont->get_memory_block()->get_depth() << std::endl;
fall_through_cont_ = nullptr;
fall_through_ = false;
......@@ -158,17 +139,18 @@ class cont_manager {
// ... we finished the continuation.
// We are now in charge continuing to execute the above continuation chain.
if (get_node(notified_cont->get_memory_block()->get_depth()) != notified_cont->get_memory_block()) {
my_state.cont_manager_.check_clean_chain();
PLS_ASSERT(active_node_->get_prev()->get_depth() == notified_cont->get_memory_block()->get_depth(),
"We must hold the system invariant to be in the correct depth.")
if (active_node_->get_prev() != notified_cont->get_memory_block()) {
// We do not own the thing we will execute.
// Own it by swapping the chain belonging to it in.
aquire_memory_chain(notified_cont->get_memory_block());
std::cout << "Now in charge of memory chain on core " << my_state.get_id() << std::endl;
// std::cout << "Now in charge of memory chain on core " << my_state.get_id() << std::endl;
}
my_state.parent_cont_ = notified_cont->get_parent();
my_state.right_spawn_ = notified_cont->is_right_child();
active_node_ = notified_cont->get_memory_block();
std::cout << "Execute cont on core " << my_state.get_id() << std::endl;
// std::cout << "Execute cont on core " << my_state.get_id() << std::endl;
notified_cont->execute();
if (!falling_through() && notified_cont->get_parent() != nullptr) {
fall_through_and_notify_cont(notified_cont->get_parent(), notified_cont->is_right_child());
......@@ -178,14 +160,16 @@ class cont_manager {
// ... we did not finish the last continuation.
// We are no longer in charge of executing the above continuation chain.
if (get_node(notified_cont->get_memory_block()->get_depth()) == notified_cont->get_memory_block()) {
PLS_ASSERT(active_node_->get_prev()->get_depth() == notified_cont->get_memory_block()->get_depth(),
"We must hold the system invariant to be in the correct depth.")
if (active_node_->get_prev() == notified_cont->get_memory_block()) {
// We own the thing we are not allowed to execute.
// Get rid of the ownership by using the offered chain.
aquire_memory_chain(target_chain);
std::cout << "No longer in charge of chain above on core " << my_state.get_id() << std::endl;
// std::cout << "No longer in charge of chain above on core " << my_state.get_id() << std::endl;
}
my_state.cont_manager_.check_clean_chain();
move_active_node_to_start();
// We are done here...nothing more to execute
return;
}
......@@ -194,7 +178,6 @@ class cont_manager {
private:
template<size_t MAX_CONT_SIZE>
static memory_block *init_memory_block(data_structures::aligned_stack &cont_storage,
memory_block *memory_chain_start,
memory_block *prev,
unsigned long depth) {
// Represents one cont_node and its corresponding memory buffer (as one continuous block of memory).
......@@ -202,7 +185,7 @@ class cont_manager {
char *memory_block_ptr = cont_storage.push_bytes<memory_block>();
char *memory_block_buffer_ptr = cont_storage.push_bytes(buffer_size);
return new(memory_block_ptr) memory_block(memory_block_buffer_ptr, buffer_size, memory_chain_start, prev, depth);
return new(memory_block_ptr) memory_block(memory_block_buffer_ptr, buffer_size, prev, depth);
}
private:
......@@ -212,7 +195,6 @@ class cont_manager {
/**
* Managing the continuation chain.
*/
memory_block *start_node_;
memory_block *active_node_;
/**
......
lib/pls/include/pls/internal/scheduling/memory_block.h
......@@ -18,11 +18,9 @@ class memory_block {
public:
memory_block(char *memory_buffer,
size_t memory_buffer_size,
memory_block *memory_chain_start,
memory_block *prev,
unsigned int depth)
: memory_chain_start_{memory_chain_start},
prev_{prev},
: prev_{prev},
next_{nullptr},
offered_chain_{nullptr},
state_{{initialized}},
......@@ -70,12 +68,6 @@ class memory_block {
void set_next(memory_block *next) {
next_ = next;
}
memory_block *get_start() {
return memory_chain_start_;
}
void set_start(memory_block *start) {
memory_chain_start_ = start;
}
enum state { initialized, execute_local, stealing, stolen, invalid };
using stamped_state = data_structures::stamped_integer;
......@@ -107,7 +99,6 @@ class memory_block {
// Linked list property of memory blocks (a complete list represents a threads currently owned memory).
// Each block knows its chain start to allow stealing a whole chain in O(1)
// without the need to traverse back to the chain start.
memory_block *memory_chain_start_;
memory_block *prev_, *next_;
// When blocked on this chain element, we need to know what other chain of memory we
......
lib/pls/include/pls/internal/scheduling/scheduler_impl.h
......@@ -76,8 +76,7 @@ struct scheduler::starter {
current_cont->store_left_result(std::move(result_1));
auto old_state = current_cont->get_memory_block()->get_state().load();
current_cont->get_memory_block()->get_state().store({old_state.stamp + 1, memory_block::invalid});
PLS_ASSERT(current_cont->get_memory_block()->get_results_missing().fetch_add(-1) != 1,
"We fall through, meaning we 'block' an cont above, thus this can never happen!");
current_cont->get_memory_block()->get_results_missing().fetch_add(-1);
// Unwind stack...
return result_type{};
}
......@@ -132,7 +131,6 @@ class scheduler::init_function_impl : public init_function {
explicit init_function_impl(F &function) : function_{function} {}
void run() override {
scheduler::par([]() {
std::cout << "Dummy Strain, " << scheduling::thread_state::get().get_id() << std::endl;
return parallel_result<int>{0};
}, [=]() {
return function_();
......
lib/pls/include/pls/internal/scheduling/task_manager.h
......@@ -59,11 +59,17 @@ class task_manager {
auto stolen_task_handle = task_deque_.pop_top();
if (stolen_task_handle) {
base_task *stolen_task = (*stolen_task_handle).task_;
std::cout << "Nearly stole on core " << thread_state::get().get_id() << " task with depth "
<< stolen_task->get_cont()->get_memory_block()->get_depth() << std::endl;
auto &atomic_state = (*stolen_task_handle).task_memory_block_->get_state();
auto &atomic_offered_chain = (*stolen_task_handle).task_memory_block_->get_offered_chain();
auto offered_chain = stealing_cont_manager.get_node((*stolen_task_handle).task_memory_block_->get_depth());
memory_block *stolen_task_memory = (*stolen_task_handle).task_memory_block_;
auto stolen_task_depth = stolen_task_memory->get_depth();
auto &atomic_state = stolen_task_memory->get_state();
auto &atomic_offered_chain = stolen_task_memory->get_offered_chain();
// std::cout << "Nearly stole on core " << thread_state::get().get_id() << " task with depth "
// << stolen_task_depth << std::endl;
// Move our chain forward for stealing...
stealing_cont_manager.move_active_node(stolen_task_depth);
auto offered_chain = stealing_cont_manager.get_active_node();
if (offered_chain == (*stolen_task_handle).task_memory_block_) {
PLS_ASSERT(false, "How would we offer our own chain? We only offer when stealing!");
......@@ -71,6 +77,7 @@ class task_manager {
auto last_state = atomic_state.load();
if (last_state.value != memory_block::initialized) {
stealing_cont_manager.move_active_node(-stolen_task_depth);
return nullptr;
}
......@@ -86,12 +93,14 @@ class task_manager {
last_offered_chain = atomic_offered_chain.load();
last_state = atomic_state.load();
if (last_state != loop_state) {
stealing_cont_manager.move_active_node(-stolen_task_depth);
return nullptr;
}
}
if (atomic_state.compare_exchange_strong(loop_state, {loop_state.stamp + 1, memory_block::stolen})) {
std::cout << "Steal!" << std::endl;
// std::cout << "Steal!" << std::endl;
stealing_cont_manager.move_active_node(1);
return stolen_task;
} else {
return nullptr;
......
lib/pls/src/internal/scheduling/scheduler.cpp
......@@ -69,7 +69,6 @@ void scheduler::work_thread_work_section() {
while (my_cont_manager.falling_through()) {
my_cont_manager.execute_fall_through_code();
}
my_cont_manager.check_clean_chain();
// Steal Routine (will be continuously executed when there are no more fall through's).
// TODO: move into separate function
......@@ -83,13 +82,11 @@ void scheduler::work_thread_work_section() {
auto &target_state = my_state.get_scheduler().thread_state_for(target);
my_cont_manager.check_clean_chain();
PLS_ASSERT(my_cont_manager.get_active_node()->get_depth() == 0, "Only steal with clean chain!");
auto *stolen_task = target_state.get_task_manager().steal_remote_task(my_cont_manager);
if (stolen_task != nullptr) {
my_state.parent_cont_ = stolen_task->get_cont();
my_state.right_spawn_ = true;
my_cont_manager.set_active_depth(stolen_task->get_cont()->get_memory_block()->get_depth() + 1);
my_cont_manager.check_clean_chain();
stolen_task->execute();
if (my_cont_manager.falling_through()) {
break;
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or sign in to comment