The shift of commodity hardware from single- to multi-core processors in the early 2000s compelled software developers to take advantage of the available parallelism of multi-cores. Unfortunately, only few---so-called embarrassingly parallel---applications can leverage this available parallelism in a straightforward manner. The remaining---non-embarrassingly parallel---applications require that their processes coordinate their possibly interleaved executions to ensure overall correctness---they require synchronization. Synchronization is achieved by constraining or even prohibiting parallel execution. Thus, per Amdahl' s law, synchronization limits software scalability. In this dissertation, we explore how to minimize the effects of synchronization on software scalability. We show that scalability of synchronization is mainly a property of the underlying hardware. This means that synchronization directly hampers the cross-platform performance portability of concurrent software. Nevertheless, we can achieve portability without sacrificing performance, by creating design patterns and abstractions, which implicitly leverage hardware details without exposing them to software developers. We first perform an exhaustive analysis of the performance behavior of synchronization on several modern platforms. This analysis clearly shows that the performance and scalability of synchronization are highly dependent on the characteristics of the underlying platform. We then focus on lock-based synchronization and analyze the energy/performance trade-offs of various waiting techniques. We show that the performance and the energy efficiency of locks go hand in hand on modern x86 multi-cores. This correlation is again due to the characteristics of the hardware that does not provide practical tools for reducing the power consumption of locks without sacrificing throughput. We then propose two approaches for developing portable and scalable concurrent software, hence hiding the limitations that the underlying multi-cores impose. First, we introduce OPTIK, a new practical design pattern for designing and implementing fast and scalable concurrent data structures. We illustrate the power of our OPTIK pattern by devising five new algorithms and by optimizing four state-of-the-art algorithms for linked lists, skip lists, hash tables, and queues. Second, we introduce MCTOP, a multi-core topology abstraction which includes low-level information, such as memory bandwidths. MCTOP enables developers to accurately and portably define high-level optimization policies. We illustrate several such policies through four examples, including automated backoff schemes for locks, and illustrate the performance and portability of these policies on five platforms.
EPFL_TH7246.pdf
openaccess
5.11 MB
Adobe PDF
286a155e63b6015341581f6c1fdb70a9