Author_Institution :
Stanford Univ., Stanford, CA, USA
Abstract :
Summary form only given. Computing is now an essential tool for all aspects of human endeavor, including healthcare, education, science, commerce, government, and entertainment. We expect our computers, whether those hidden away in data-centers or those in a handheld form factor, to be capable of running sophisticated algorithms that process rapidly growing volumes of data. In other words, we expect our computers to have exponentially increasing performance at constant cost (energy and chip area). For decades, CMOS technology has been our ally, providing exponential improvements in both transistor density and energy consumption, which we turned into exponential improvements in system performance. Unfortunately, we are now in a phase where transistor cost and energy consumption are barely scaling, making it necessary to rethink the way we build scalable systems. In this talk, we will consider how to advance computer systems without technology progress. There are several promising directions that combined can provide improvements equivalent to several decades of Moore´s law. These directions include massive parallelism with locality awareness, specialization, removing the bloat from our infrastructure, increasing system utilization, and embracing approximate computing. We will review motivating results in these areas, establish that they require cross-layer optimizations across both hardware and software, and discuss the remaining challenges that systems researchers must address.
Keywords :
computer centres; parallel processing; performance evaluation; CMOS technology; Moore´s law; approximate computing; computer systems; cross-layer optimizations; data centers; energy consumption; exponential system performance improvement; handheld form factor; locality awareness; specialization; system utilization; transistor cost; transistor density; Computers; Educational institutions; Energy consumption; Hardware; Parallel processing; Software; Transistors;