Reducing interference in memory hierarchy resources using application aware management.

摘要

Aggressive technology scaling has resulted in an increase in number of cores being integrated on-chip. While on-chip cores are increasing at a fast rate, the memory hierarchy resources are scaling at a much slower pace. The memory resources, such as different levels of on-chip cache and the off-chip memory bandwidth, are costly and are often shared across multiple on-chip cores. This leads to multiple applications contending for access to these common resources. In the process, these applications can harmfully interfere with one another, and, this interference can result in significant degradation of both system throughput and individual application performance. Therefore, intelligently managing the shared memory resources by mitigating inter-application interference is vitally important in emerging multicore systems.;This dissertation makes three key contributions towards addressing the above problem of interference in shared memory resources. First, this dissertation considers the last-level shared cache and the off-chip memory as instances of shared memory resources, and, studies the causes and different ways in which applications interfere with one another while contending for a resource. Second, this dissertation studies the negative impact of resource contention on application and system performances. Third, this dissertation proposes novel schemes to mitigate inter-application interference and thereby improve system and application performance. These schemes aim to efficiently manage the resources in an application aware manner with the goal of mitigating the overall inter-application interference. An application aware resource management scheme considers the memory access characteristics of all the contending applications and uses this information to manage the shared resource. The resource management decisions are based on two key principles: 1) isolating applications/threads that harmfully interfere from each other by partitioning the resources between the interfering applications, and, 2) deciding the size of the resource partition that an application gets based on its memory access characteristics and requirements.;The trend of integrating increasing number of cores on a single chip is projected to continue into the future. This continued scaling is propelling the parallel computation capability of emerging multicore systems. Efficient management of shared memory hierarchy resources will become ever more important in the future if we are to ensure that applications extract the maximum possible parallelism from these multicore systems. This dissertation takes an important step towards addressing this problem by proposing novel schemes to efficiently manage multiple memory hierarchy resources. These schemes are very effective in practice, improving both system performance and individual application performance.