Instruction, thread, process level parallelism

Instruction-level parallelism (ILP)

Several instructions of a single program are overlapped or executed at the same time inside one processor. Hardware techniques include pipelining, superscalar execution (several execution units issue more than one instruction per clock), out-of-order execution, and VLIW. The limit comes from data and control dependences between instructions. The programmer need not do anything; the hardware or compiler extracts it.

Thread-level parallelism (TLP)

A program is split into several threads (lightweight units sharing the same address space and resources), and threads execute at the same time on different cores, or are interleaved on one core with hardware multithreading (such as Hyper-Threading). For example, a web browser uses one thread for the user interface and others for downloading. Threads communicate through shared memory and need synchronisation (locks).

Process-level parallelism

Different independent processes (separate address spaces) run simultaneously on multiple processors or machines (multiprocessors, clusters). Processes communicate by IPC (messages, pipes). It is coarse-grained and suits multi-user and distributed systems, for example running a compiler and a database server in parallel.