The majority of today's scientific and data analytics workloads are still run on relatively energy inefficient, heavyweight, general-purpose processing cores, often referred to in the literature as latency-oriented architectures. The flexibility of these architectures and the programmer aids included (e.g. large and deep cache hierarchies, branch prediction logic, pre-fetch logic) makes them flexible enough to run a wide range of applications fast. However, we have started to see growth in the use of lightweight, simpler, energy-efficient, and functionally constrained cores. These architectures are commonly referred to as throughput-oriented.

Within each shared memory node, the computational backbone of future throughput-oriented HPC machines will consist of large pools of lightweight cores. The first wave of throughput-oriented computing came in the mid 2000's with the use of GPUs for general-purpose and scientific computing. Today we are entering the second wave of throughput-oriented computing, with the introduction of NVIDIA Pascal GPUs, Intel Knights Landing Xeon Phi processors, the Epiphany Co-Processor, the Sunway MPP, and other throughput-oriented architectures that enable pre-exascale computing. However, while the majority of the FLOPS in designs for future HPC systems come from throughput-oriented architectures, they are still commonly paired with latency-oriented cores which handle management functions and lightweight/un-parallelizable computational kernels. Hence, most future HPC machines will be heterogeneous in their processing cores.

However, the heterogeneity of future machines will not be limited to the processing elements. Indeed, heterogeneity will also exist in the storage, networking, memory, and software stacks of future supercomputers. As a result, it will be necessary to combine many different programming models and libraries in a single application. How to do so in a programmable and well-performing manner is an open research question. This thesis addresses this question using two approaches.

First, we explore using managed runtimes on HPC platforms. As a result of their high-level programming models, these managed runtimes have a long history of supporting data analytics workloads on commodity hardware, but often come with overheads which make them less common in the HPC domain. Managed runtimes are also not supported natively on throughput-oriented architectures.

Second, we explore the use of a modular programming model and work-stealing runtime to compose the programming and scheduling of multiple third-party HPC libraries. This approach leverages existing investment in HPC libraries, unifies the scheduling of work on a platform, and is designed to quickly support new programming model and runtime extensions.

In support of these two approaches, this thesis also makes novel contributions in tooling for future supercomputers. We demonstrate the value of checkpoints as a software development tool on current and future HPC machines, and present novel techniques in performance prediction across heterogeneous cores.