On Performance: From Hardware up to Distributed Systems

Abstract

Nothing is easy nowadays: frequency of processors increased thousand times, system performance as a whole sometimes tripled. Complexity of the system became uncontrollable with zillions of processes and elements to juggle increased unconsciously, leaving for us some comfort but at an astronomical cost. What it means? We are doing something seriously wrong and doing it consistently and persistently. Thus, authors of this work have decided to put together our own discussions and estimations we did since 2002 up to now. We show that system performance depends on user, hardware, and software, structure or architecture of a system and its topology. We propose to see performance analysis a bit wider, thinking systematically what various zones of computer or distributed system can bring or contribute, including the role of processor, structure of system software and overvalued parallelization (try to eat and dance at the same time—it might be fun). We have introduced a kind of virtual architecture through which see instruction execution considering what is in there for us and what system requires for itself. The observation is rather pessimistic. We have briefly demonstrated what simplest architecture if carefully designed can give regarding performance, reliability and energy efficiency AT THE SAME TIME! Regarding distributed systems, we show that Amdahl Law is also very overoptimistic mostly serves to promote parallel architectures and distributed systems. Simple model that we have explained for kids from British primary school and even did field study with them so-called “fence model” made clear that the limit of performance or simply overall reasonably good design is unachievable until we start rethinking the whole architecture and its main element interaction—human, hardware and system software together, pursuing three nonfunctional requirements, performance, reliability, and energy efficiency in concert.