Should Amdahl's Law be repealed?

Abstract

An appropriate observation made by Gene Amdahl in 1967 about data management housekeeping overhead was subsequently construed by his commentators as embodying a fundamental limitation to parallel computation and was elevated to to the rank of “Amdahl's Law”. It is argued that Amdahl's Law, as formulated, has no fundamental character, but refers to specific technological choices, such as programs and input/output modes. Indeed, algorithmic research has shown that most problems are parallelizable. It is also argued here, on the basis of VLSI's area-time theory, that the I/O bandwidth is not a basic physical limitation. The seemingly inherently serial P-complete problems, as characterized by algorithmic research, constitute the only class to whose programs Amdahl's Law trivially applies.

Moreover, in a realistic computational model that fully accounts for the finiteness of the speed of light, it can be existentially shown that, with respect to multiprocessors, uniprocessors incur not only the obvious slowdown due to loss of parallelism but also a more subtle slowdown due to loss of locality. This makes the case for the multiprocessor and for the repeal of Amdahl's Law.