An Initial Investigation of a Multi-layered Approach for Optimizing and Parallelizing Real-Time Media and Audio Applications

Amdahl's law in the context of heterogeneous many‐core systems – a survey

“…This technique, using what may be regarded as a penalty term in the denominator of Amdahl's law (the CC‐ratio in case of the Li–Malek model), has been used on multiple occasions by different researchers, to cover additive overheads of all types [77, (p.167), 23, (p.42), 46, 51, 52]. However, overheads may be caused by very complicated effects of multiple factors, which lead this line of modelling to become more and more sophisticated, expanding the penalty term from a constant to various different functions.…”

“… for this fully balanced workload distribution is This is what Hill–Marty's asymmetric model assumes. This ideal case almost never happens, as the values are real numbers and workloads generally cannot be divided into infinitesimally small parts, and any dependency across threads would cause the critical path to be irreducible after some point [23, p. 42]. Some amount of waiting is therefore almost inevitable, even if such system software as load balancers may support the run‐time redistribution of partially executed threads across cores, through such techniques as task (thread) migration.…”

“…The method of adding a penalty term to the denominator of Amdahl's law to represent overheads, exemplified by Li–Malek's model (9) [50] and used in other work [77, p.167, 23, p.42, 51, 52], is also covered by the multi‐fraction model. The multi‐fraction model allows the more precise representation of qualitative and quantitative overhead effects as these can usually be precisely incorporated into task graph modifications: additional tasks, lengthening existing tasks, synchronising tasks, data overheads, instruction overheads, etc.…”

“…An extensive body of work has been concentrated on the modelling and analysis of the effects of computational parallelisation. The property of speedup is a popular topic in these studies [1–5, 13, 19, 22–24], whilst energy and power have not been neglected [3, 5, 25–30]. Various problems related to mapping parallelisable (mostly software) workloads onto M/MCP hardware platforms have also been intensely scrutinised.…”

“…This paper reviews the literature on workload parallelisation in the context of M/MCP speedup, particularly those related to Amdahl‐type speedup models such as Amdahl [1], Gustafson [22] and Sun and Ni [2]. Non‐Amdahl style speedup models relating to the concept of parallelism will also be comparatively studied [2, 13, 35, 36] and new research on extending Amdahl's law to cover non‐zero and non‐infinity parallelism will be highlighted [19, 23, 24, 37, 38]. The review especially covers model extensions dealing with HeMCP with different degrees of heterogeneity [4, 5, 24, 38].…”