Our large-scale computing systems have enjoyed tremendous growth in computational power over the last few decades. Today's supercomputers can perform 10^18 floating-point calculations per second — roughly equivalent to our brain's postulated computational power and incur more than a million metric tons of carbon per year, equivalent to over a quarter-million households' carbon footprint.

Unfortunately, the race for raw speed has come at a cost and posed new challenges. The systems have become increasingly complex and heterogeneous; therefore, extracting meaningful performance from these systems has led to a severe productivity challenge for programmers. This is both sad and concerning because ultimately, a productivity-challenged end-user is likely to experience slower performance, higher costs, and be unable to reason about the environmental implications of their computation.

If you have run an application on a cloud computing platform or own a high-performance cluster and were frustrated with slow speed and high costs or feel anxious about assessing the environmental impact of your computation, you are not alone -- you are the perfect audience for this talk! In this talk, I will share insights and new techniques to make complex applications run faster and cheaper in an increasingly heterogeneous computing environment and emerging computing models. Simply buying the fastest and most expensive computing device is neither always cost-effective nor carbon-footprint-effective — in fact, it can worsen the situation. Instead, I'll demonstrate how to devise strategies to "mix up" the old and new hardware and map applications intelligently to extract more performance at a lower cost and a lower carbon footprint.