Concurrent computing is a form of computing in which several computations are executed during overlapping time periods —concurrently— instead of sequentially (one completing before the next starts). This is a property of a system —this may be an individual program, a computer, or a network— and there is a separate execution point or "thread of control" for each computation ("process"). A concurrent system is one where a computation can advance without waiting for all other computations to complete; where more than one computation can advance at the same time.
Concurrent computing is related to but distinct from parallel computing, though these concepts are frequently confused, and both can be described as "multiple processes executing during the same period of time". In parallel computing, execution occurs at the same physical instant, for example on separate processors of a multi-processor machine, with the goal of speeding up computations—parallel computing is impossible on a (one-core) single processor, as only one computation can occur at any instant (during any single clock cycle).[a] By contrast, concurrent computing consists of process lifetimes overlapping, but execution need not happen at the same instant. The goal here is to model processes in the outside world that happen concurrently, such as multiple clients accessing a server at the same time. Structuring software systems as composed of multiple concurrent, communicating parts can be useful for tackling complexity, regardless of whether the parts can be executed in parallel.:1
For example, concurrent processes can be executed on one core by interleaving the execution steps of each process via time-sharing slices: only one process runs at a time, and if it does not complete during its time slice, it is paused, another process begins or resumes, and then later the original process is resumed. In this way, multiple processes are part-way through execution at a single instant, but only one process is being executed at that instant.
Concurrent computations may be executed in parallel, for example by assigning each process to a separate processor or processor core, or distributing a computation across a network, but in general, the languages, tools and techniques for parallel programming may not be suitable for concurrent programming, and vice versa.
The exact timing of when tasks in a concurrent system are executed depend on the scheduling, and tasks need not always be executed concurrently. For example, given two tasks, T1 and T2:
T1 may be executed and finished before T2 or vice versa (serial and sequential);
T1 and T2 may be executed alternately (serial and concurrent);
T1 and T2 may be executed simultaneously at the same instant of time (parallel and concurrent).
The word "sequential" is used as an antonym for both "concurrent" and "parallel"; when these are explicitly distinguished, concurrent/sequential and parallel/serial are used as opposing pairs. A schedule in which tasks execute one at a time (serially, no parallelism), without interleaving (sequentually, no concurrency: no task begins until the prior task ends) is called a serial schedule. A set of tasks that can be scheduled serially is serializable, which simplifies concurrency control.
The main challenge in designing concurrent programs is concurrency control: ensuring the correct sequencing of the interactions or communications between different computational executions, and coordinating access to resources that are shared among executions. Potential problems include race conditions, deadlocks, and resource starvation. For example, consider the following algorithm to make withdrawals from a checking account represented by the shared resource balance:
Suppose balance = 500, and two concurrent threads make the calls withdraw(300) and withdraw(350). If line 3 in both operations executes before line 5 both operations will find that balance >= withdrawal evaluates to true, and execution will proceed to subtracting the withdrawal amount. However, since both processes perform their withdrawals, the total amount withdrawn will end up being more than the original balance. These sorts of problems with shared resources need the use of concurrency control, or non-blocking algorithms.
Because concurrent systems rely on the use of shared resources (including communication media), concurrent computing in general needs the use of some form of arbiter somewhere in the implementation to mediate access to these resources.
Unfortunately, while many solutions exist to the problem of a conflict over one resource, many of those "solutions" have their own concurrency problems such as deadlock when more than one resource is involved.
Increased program throughput—parallel execution of a concurrent program allows the number of tasks completed in a given time to increase.
High responsiveness for input/output—input/output-intensive programs mostly wait for input or output operations to complete. Concurrent programming allows the time that would be spent waiting to be used for another task.
More appropriate program structure—some problems and problem domains are well-suited to representation as concurrent tasks or processes.
A number of different methods can be used to implement concurrent programs, such as implementing each computational execution as an operating system process, or implementing the computational processes as a set of threads within a single operating system process.
In some concurrent computing systems, communication between the concurrent components is hidden from the programmer (e.g., by using futures), while in others it must be handled explicitly. Explicit communication can be divided into two classes:
Concurrent components communicate by altering the contents of shared memory locations (exemplified by Java and C#). This style of concurrent programming usually needs the use of some form of locking (e.g., mutexes, semaphores, or monitors) to coordinate between threads. A program that properly implements any of these is said to be thread-safe.
Concurrent components communicate by exchanging messages (exemplified by Scala, Erlang and occam). The exchange of messages may be carried out asynchronously, or may use a synchronous "rendezvous" style in which the sender blocks until the message is received. Asynchronous message passing may be reliable or unreliable (sometimes referred to as "send and pray"). Message-passing concurrency tends to be far easier to reason about than shared-memory concurrency, and is typically considered a more robust form of concurrent programming. A wide variety of mathematical theories to understand and analyze message-passing systems are available, including the actor model, and various process calculi. Message passing can be efficiently implemented via symmetric multiprocessing, with or without shared memory cache coherence.
Shared memory and message passing concurrency have different performance characteristics. Typically (although not always), the per-process memory overhead and task switching overhead is lower in a message passing system, but the overhead of message passing is greater than for a procedure call. These differences are often overwhelmed by other performance factors.
Concurrent computing developed out of earlier work on railroads and telegraphy, from the 19th and early 20th century, and some terms date to this period, such as semaphores. These arose to address the question of how to handle multiple trains on the same railroad system (avoiding collisions and maximizing efficiency) and how to handle multiple transmissions over a given set of wires (improving efficiency), such as via time-division multiplexing (1870s).
The academic study of concurrent algorithms started in the 1960s, with Dijkstra (1965) credited with being the first paper in this field, identifying and solving mutual exclusion.
Today, the most commonly used programming languages that have specific constructs for concurrency are Java and C#. Both of these languages fundamentally use a shared-memory concurrency model, with locking provided by monitors (although message-passing models can and have been implemented on top of the underlying shared-memory model). Of the languages that use a message-passing concurrency model, Erlang is probably the most widely used in industry at present.
Many concurrent programming languages have been developed more as research languages (e.g. Pict) rather than as languages for production use. However, languages such as Erlang, Limbo, and occam have seen industrial use at various times in the last 20 years. Languages in which concurrency plays an important role include:
Ada—general purpose, with native support for message passing and monitor based concurrency
Rust—for system programming, focus on massive concurrency, using message-passing with move semantics, shared immutable memory, and shared mutable memory that is provably free of race conditions.
SALSA—actor-based with token-passing, join, and first-class continuations for distributed computing over the Internet
Scala—general purpose, designed to express common programming patterns in a concise, elegant, and type-safe way
SequenceL—general purpose functional, main design objectives are ease of programming, code clarity-readability, and automatic parallelization for performance on multicore hardware, and provably free of race conditions
^This is discounting parallelism internal to a processor core, such as pipelining or vectorized instructions. A one-core, one-processor machine may be capable of some parallelism, such as with a coprocessor, but the processor alone is not.
Patterson, David A.; Hennessy, John L. (2013). Computer Organization and Design: The Hardware/Software Interface. The Morgan Kaufmann Series in Computer Architecture and Design (5 ed.). Morgan Kaufmann. ISBN978-0-12407886-4.