Threading Models Sample Clauses

Threading Models. In describing processors that support multiple threads of execution, various terms are used to specify how the individual threads can be controlled by the programmer. Both MPMD and SPMD (described below) are cases of MIMD (Multiple Instruction Multiple Data) execution, where each thread notionally executes its own instruction stream. Single Instruction Multiple Data (SIMD) SIMD is not a multi-threading model at all, but a way to extract more performance from a single thread. SIMD is an architectural feature that allows a single processor to process a given data set using fewer total operations. A processor supporting SIMD includes registers that can contain several discrete data values, and instructions that can simultaneously perform the same operation on all data values packed into registers. The number of values that can be stored and operated on in this way is termed the ”SIMD width” (in modern processors, typically four or eight 32-bit values). The maximum possible performance increases by a factor approaching the SIMD width (typically 4 or 8), but this improvement can only be seen when the workload is able to use the SIMD features effectively. For maximum performance every operation in the algorithm must be converted to use the SIMD operations, otherwise the non-SIMD operations become the bottleneck, which requires regularity in the algorithm. As the SIMD width increases, the amount of regularity required increases too; this limits the usefulness of increasing the width. In practice many algorithms do see a significant performance improvement when SIMD features are used, but there is relatively little scope for further improvement as increasing SIMD width has diminishing returns. Multiple Program Multiple Data (MPMD) In an MPMD model, each thread is independent of the others. Each thread executes its own program with its own data, so that a given thread can be doing a different task and processing different data from any other thread. MPMD systems support a fixed number of hardware threads. A larger number of software threads can be supported by switching each hardware thread amongst several software threads in the OS. This makes thread creation relatively expensive, as each thread must be individually created by software. This places constraints on how these software created threads can be used by programmers, as the useful work done by a thread has to justify the cost of creating it. It also requires at least some component of the OS to run on all ...