Mastering std::async in Modern C++

· c++, concurrency, async, futures

std::async is the standard library’s shortest path to running work asynchronously and receiving the result as a std::future. Used well, it hides thread management, propagates exceptions, and helps structure CPU-bound pipelines. Misunderstood, it quietly serializes work or launches threads you never join. Let’s walk through how to make the most of it.

You may want to review Understanding Futures and Promises in Modern C++ first, then follow up with Composing Futures in Modern C++. For a deep dive into the move semantics that make futures efficient, see Understanding Reference Types in Modern C++.

What std::async Actually Does

#include <future>

std::future<int> future = std::async(std::launch::async, [] {
    return heavy_computation();
});

// later
int answer = future.get();

std::async wraps the callable in a packaged task backed by a promise/future pair. When you call get(), you either receive the return value or the exception the callable threw. No explicit threads, no manual promise, just the result.

Launch Policies Matter

The first template argument controls when and where work runs:

  • std::launch::async: start immediately on a new thread.
  • std::launch::deferred: delay execution until the first .get() or .wait(). Runs synchronously in that call.
  • Default (no policy): implementation may pick either. You must not depend on one behavior.
auto maybe_async = std::async([] { return compute(); });
auto definitely_async = std::async(std::launch::async, [] { return compute(); });
auto deferred = std::async(std::launch::deferred, [] { return compute(); });

Use std::launch::async explicitly when latency overlaps matter. Reserve deferred for expensive work that might not be needed.

Lifetime and Blocking Rules

  • The destructor of the returned std::future blocks if the task was launched with std::launch::async and you haven’t called get() or wait(). Always consume the future before it goes out of scope.
  • If the task was std::launch::deferred, the destructor does nothing—execution happens during get() instead.
  • Copying is illegal. Move the future if you need to transfer ownership.
std::future<void> fire_and_wait() {
    auto fut = std::async(std::launch::async, [] { work(); });
    // do other stuff
    fut.wait(); // ensures task finished before returning
    return fut; // UB: returning moved-from future. Instead, return after join
}

Avoid returning by value unless you’re happy with move-semantics and the caller takes ownership.

Exception Propagation

If the callable throws, the exception is stored and rethrown by future.get():

auto fut = std::async(std::launch::async, []() -> int {
    throw std::runtime_error("bad");
});

try {
    fut.get();
} catch (const std::runtime_error& e) {
    handle_error(e);
}

Because std::async already manages the promise, you don’t need explicit try/catch in the lambda unless you want to translate errors.

Example: Parallel Accumulate

template <typename It>
int parallel_sum(It begin, It end) {
    auto length = std::distance(begin, end);
    if (length < 1'000) {
        return std::accumulate(begin, end, 0);
    }

    It mid = begin;
    std::advance(mid, length / 2);

    auto lower = std::async(std::launch::async, parallel_sum<It>, begin, mid);
    int upper = parallel_sum(mid, end);
    return lower.get() + upper;
}
  • One branch recurses asynchronously.
  • The other runs inline.
  • The recursive get() ensures all child tasks complete before returning.

This pattern keeps the task tree bounded and avoids exhausting the thread implementation.

Example: Overlapping CPU and I/O

auto cpu_future = std::async(std::launch::async, run_simulation);
auto io_future  = std::async(std::launch::async, read_disk_snapshot);

simulate_ui();

auto state = io_future.get();   // wait for I/O
auto result = cpu_future.get(); // wait for simulation

Launching both tasks with std::launch::async overlaps CPU work with disk or networking. Be sure to call get() in all paths to avoid leaving background threads alive in destructors.

Handling Timeouts

auto fut = std::async(std::launch::async, run_rpc);
if (fut.wait_for(std::chrono::milliseconds(100)) == std::future_status::timeout) {
    cancel_rpc();
    // Future still needs to be consumed to avoid blocking destructor
    try {
        fut.get(); // likely throws or blocks until completion
    } catch (...) {
        // handle cancellation result
    }
}

You cannot truly cancel the underlying task with standard std::async, but you can structure your code so the worker checks a shared atomic flag.

Composition Strategies

std::async returns std::future. To combine results:

  • For simple chains, call .get() and feed results to the next std::async.
  • For fan-out, hand the futures to std::when_all (C++20) or helpers from the compound futures article.
  • When you need to schedule follow-up work, wrap the call in a helper that launches another std::async after .get() completes.

Common Pitfalls

  • Ignoring the return future: letting it go out of scope blocks (async) or leaves work undone (deferred).
  • Mixing policies accidentally: implementations may choose deferred by default. Specify the policy explicitly.
  • Long-running CPU loops: saturating std::async with hot loops may spawn more threads than your system can handle. Consider using a thread pool or task scheduler.
  • Shared state: std::async doesn’t eliminate data races. Protect shared data with synchronization or confine ownership to the task scope.

Wrapping Up

std::async shines when you want convenient, exception-safe task launching without writing a custom thread wrapper. Couple it with strong future handling practices—from basic promise/future contracts to the compound patterns we explored—and you get clean, maintainable concurrency for both small utilities and production code. Grow from here by experimenting with executors (std::jthread, C++23 continuations) and dedicated task frameworks when you need more control.

Keep Reading