High-performance atomic increment using thread caching.

folly/ThreadCachedInt.h introduces a integer class designed for high performance increments from multiple threads simultaneously without loss of precision. It has two read modes, readFast gives a potentially stale value with one load, and readFull gives the exact value, but is much slower, as discussed below.


Increment performance is up to 10x greater than std::atomic_fetch_add in high contention environments. See folly/test/ThreadCachedIntTest.h for more comprehensive benchmarks.

readFast is as fast as a single load.

readFull, on the other hand, requires acquiring a mutex and iterating through a list to accumulate the values of all the thread local counters, so is significantly slower than readFast.


Create an instance and increment it with increment or the operator overloads. Read the value with readFast for quick, potentially stale data, or readFull for a more expensive but precise result. There are additional convenience functions as well, such as set.

    ThreadCachedInt<int64_t> val;
    EXPECT_EQ(0, val.readFast());
    ++val;                        // increment in thread local counter only
    EXPECT_EQ(0, val.readFast()); // increment has not been flushed
    EXPECT_EQ(1, val.readFull()); // accumulates all thread local counters
    EXPECT_EQ(2, val.readFast());
    EXPECT_EQ(2, val.readFull());


folly::ThreadCachedInt uses folly::ThreadLocal to store thread specific objects that each have a local counter. When incrementing, the thread local instance is incremented. If the local counter passes the cache size, the value is flushed to the global counter with an atomic increment. It is this global counter that is read with readFast via a simple load, but will not count any of the updates that haven't been flushed.

In order to read the exact value, ThreadCachedInt uses the extended readAllThreads() API of folly::ThreadLocal to iterate through all the references to all the associated thread local object instances. This currently requires acquiring a global mutex and iterating through the references, accumulating the counters along with the global counter. This also means that the first use of the object from a new thread will acquire the mutex in order to insert the thread local reference into the list. By default, there is one global mutex per integer type used in ThreadCachedInt. If you plan on using a lot of ThreadCachedInts in your application, considering breaking up the global mutex by introducing additional Tag template parameters.

set simply sets the global counter value, and marks all the thread local instances as needing to be reset. When iterating with readFull, thread local counters that have been marked as reset are skipped. When incrementing, thread local counters marked for reset are set to zero and unmarked for reset.

Upon destruction, thread local counters are flushed to the parent so that counts are not lost after increments in temporary threads. This requires grabbing the global mutex to make sure the parent itself wasn't destroyed in another thread already.

Alternate Implementations

There are of course many ways to skin a cat, and you may notice there is a partial alternate implementation in folly/test/ThreadCachedIntTest.cpp that provides similar performance. ShardedAtomicInt simply uses an array of std::atomic<int64_t>'s and hashes threads across them to do low-contention atomic increments, and readFull just sums up all the ints.

This sounds great, but in order to get the contention low enough to get similar performance as ThreadCachedInt with 24 threads, ShardedAtomicInt needs about 2000 ints to hash across. This uses about 20x more memory, and the lock-free readFull has to sum up all 2048 ints, which ends up being a about 50x slower than ThreadCachedInt in low contention situations, which is hopefully the common case since it's designed for high-write, low read access patterns. Performance of readFull is about the same speed as ThreadCachedInt in high contention environments.

Depending on the operating conditions, it may make more sense to use one implementation over the other. For example, a lower contention environment will probably be able to use a ShardedAtomicInt with a much smaller array without hurting performance, while improving memory consumption and perf of readFull.