Many Arrow C++ operations distribute work across multiple threads to take
advantage of underlying hardware parallelism. For example, when
Parquet file we can decode each
column in parallel. To achieve this we submit tasks to an executor of some kind.
Within Arrow C++ we use thread pools for parallel scheduling and an event loop when the user has requested serial execution. It is possible for users to provide their own custom implementation, though that is an advanced concept and not covered here.
CPU vs. I/O¶
In order to minimize the overhead of context switches our default thread pool for CPU-intensive tasks has a fixed size, defaulting to std::thread::hardware_concurrency. This means that CPU tasks should never block for long periods of time because this will result in under-utilization of the CPU. To achieve this we have a separate thread pool which should be used for tasks that need to block. Since these tasks are usually associated with I/O operations we call this the I/O thread pool. This model is often associated with asynchronous computation.
The size of the I/O thread pool currently defaults to 8 threads and should
be sized according to the parallel capabilities of the I/O hardware. For example,
if most reads and writes occur on a typical HDD then the default of 8 will probably
be sufficient. On the other hand, when most reads and writes occur on a remote
filesystem such as S3, it is often possible to benefit from many concurrent reads
and it may be possible to increase I/O performance by increasing the size of the
I/O thread pool. The size of the default I/O thread pool can be managed with
ARROW_IO_THREADS environment variable or
Increasing the size of the CPU thread pool is not likely to have any benefit. In
some cases it may make sense to decrease the size of the CPU thread pool in order
to reduce the impact that Arrow C++ has on hardware shared with other processes or user
threads. The size of the default CPU thread pool can be managed with the
OMP_NUM_THREADS environment variable or with the
Operations in Arrow C++ that may use threads can usually be configured to run serially via some kind of parameter. In this case we typically replace the CPU executor with an event loop operated by the calling thread. However, many operations will continue to use the I/O thread pool. This means that some parallelism may still occur even when serial execution is requested.
Jemalloc Background Threads¶
When using the jemalloc allocator a small number of background threads will be created by jemalloc to manage the pool. These threads should have minimal impact but can show up as a memory leak when running analysis tools like Valgrind. This is harmless and can be safely suppressed or Arrow C++ can be compiled without jemalloc.
Arrow C++ uses
arrow::Future to communicate results between threads. Typically
arrow::Future will be created when an operation needs to perform some kind
of long running task that will block for some period of time.
objects are mainly meant for internal use and any method that returns an
arrow::Future will usually have a synchronous variant as well.