kMAF

Abstract

One of the main challenges for parallel architectures is the increasing complexity of the memory hierarchy, which consists of several levels of private and shared caches, as well as interconnections between separate memories in NUMA machines. To make full use of this hierarchy, it is necessary to improve the locality of memory accesses by reducing accesses to remote caches and memories, and using local ones instead. Two techniques can be used to increase the memory access locality: executing threads and processes that access shared data close to each other in the memory hierarchy (thread affinity), and placing the memory pages they access on the NUMA node they are executing on (data affinity). Most related work in this area focuses on either thread or data affinity, but not both, which limits the improvements. Other mechanisms require expensive operations, such as memory access traces or binary analysis, require changes to hardware or work only on specific parallel APIs.In this paper, we introduce kMAF, a mechanism that automatically manages thread and data affinity on the kernel level. The memory access behavior of the running application is determined during its execution by analyzing its page faults. This information is used by kMAF to migrate threads and memory pages, such that the overall memory access locality is optimized. Extensive evaluation with 27 benchmarks from 4 benchmark suites shows substantial performance improvements, with results close to an oracle mechanism. Execution time was reduced by up to 35.7% (13.8% on average), while energy efficiency was improved by up to 34.6% (9.3% on average).

A Sharing-Aware Memory Management Unit for Online Mapping in Multi-core Architectures

“…The kMAF affinity framework is proposed in [11]. It performs both thread and data mapping and gather information from page faults.…”

“…We compare SAMMU to the following techniques: Random and Oracle mappings, the Marathe [17] data mapping mechanism, Autopin [14], the kMAF affinity framework [11] and NUMA Balancing [7]. For the random mapping, we randomly generated a thread and data mapping for each execution.…”

“…Sharing-aware thread and data mapping improve performance and energy efficiency of parallel applications by optimizing memory accesses [11]. Improvements happen for three main reasons.…”

Reducing Data Movement on Large Shared Memory Systems by Exploiting Computation Dependencies

“…In order to mitigate NUMA effects, techniques for migrating threads, memory pages or both already exist [15,16,39]. These techniques aim to move computation near to data or vice versa in order to reduce memory access time.…”

Graph partitioning applied to DAG scheduling to reduce NUMA effects

“…Techniques at the OS level for mitigating these effects include migrating threads, memory pages or both [2,3,8]. However, they do not exploit application-specific information to predict accesses to remotely allocated data before the software starts showing this behavior, they take action when the application is already suffering from remote memory accesses.…”