NumaTBB

Threading Building Blocks with gpu and numa

Use dynamic splitting algorithms for data-driven parallel tasks

Reference from:

Self-Tuning Query Scheduling for Analytical Workloads

https://15721.courses.cs.cmu.edu/spring2024/papers/08-scheduling/wagner-sigmod21.pdf

How to build

use sh build.sh to build and run unit test

use sh build.sh r to build with release

How to use

1. make a task set

we got some situation here:

• the input's iterator is known

// define input
std::vector<int> input(1000, 1);

// operator
auto process_func = [](std::vector<int>::iterator begin,
                       std::vector<int>::iterator end,
                       int multiplier) -> int {
        // ...
};

// finalizer
auto finalize_func =
    [](std::vector<int>&& partial_results) -> int {
        // ...
};

auto task_set =
    make_task_set(data.begin(), data.end(), process_func, finalize_func, /* params */ 2);

// we can get a result by future type
auto future = task_set->get_future();

// submit to scheduler...

int result = future.get();

• the input is a future type (eg. from the task set before)

auto input_future = before_task_set->get_future();

// operator
auto process_func = [](std::vector<int>::iterator begin,
                       std::vector<int>::iterator end,
                       int multiplier) -> int {
        // ...
};

// finalizer
auto finalize_func =
    [](std::vector<int>&& partial_results) -> int {
        // ...
};

auto task_set =
    make_task_set(std::move(input_future), process_func, finalize_func, /* params */ 2);

// we can get a result by future type
auto future = task_set->get_future();

// submit to scheduler...

int result = future.get();

• need to pass begin or end of the input, if the input is already defined, just pass the value, else if the input is a future type, need to define like this:

auto input_future = before_task_set->get_future();

// operator
auto process_func = [](std::vector<int>::iterator begin,
                       std::vector<int>::iterator end,
                       std::pair<std::vector<int>::iterator, std::vector<int>::iterator> context,
                       int multiplier) -> int {
        // ...
        // now can calculate offset by the context
};

// finalizer
auto finalize_func =
    [](std::vector<int>&& partial_results) -> int {
        // ...
};

// the nullptr will be replace by the _begin / _end when the func is invoked
auto task_set =
    make_task_set(std::move(input_future), process_func, finalize_func, /* params */ {nullptr, nullptr}, 2);

// we can get a result by future type
auto future = task_set->get_future();

// submit to scheduler...

int result = future.get();

• if the func return void, you can use VOID_FINALIZE as a finalizer func, and still can get a future(always return true) to judge whether the task set is finished

• some params is future type or shared future type, just pass them, the inner will call future.get() to get the value