Non uniform memory access | EPFL Graph Search

En informatique, un système NUMA (pour non uniform memory access ou non uniform memory architecture, signifiant respectivement accès mémoire non uniforme et architecture mémoire non uniforme) est un système multiprocesseur dans lequel les zones mémoire sont séparées et placées en différents endroits (et sur différents bus). Vis-à-vis de chaque processeur, les temps d'accès diffèrent donc suivant la zone mémoire accédée. Le système NUMA est conçu pour pallier les limites de l'architecture SMP (symmetric multiprocessing) dans laquelle tout l'espace mémoire est certes accessible par un unique bus, mais rend de ce fait inefficaces, par encombrement, les accès concurrents par les différents processeurs. Une architecture plus adaptée devient donc nécessaire pour les systèmes ayant de nombreux processeurs. NUMA représente une position médiane entre le SMP et le clustering (diverses machines). ccNUMA est acronyme de « cache coherent NUMA ». Par abus de langage et d'usage, les deux sont de

Scaling Up Concurrent Analytical Workloads on Multi-Core Servers

Iraklis Psaroudakis

Today, an ever-increasing number of researchers, businesses, and data scientists collect and analyze massive amounts of data in database systems. The database system needs to process the resulting highly concurrent analytical workloads by exploiting modern multi-socket multi-core processor systems with non-uniform memory access (NUMA) architectures and increasing memory sizes. Conventional execution engines, however, are not designed for many cores, and neither scale nor perform efficiently on modern multi-core NUMA architectures. Firstly, their query-centric approach, where each query is optimized and evaluated independently, can result in unnecessary contention for hardware resources due to redundant work found across queries in highly concurrent workloads. Secondly, they are unaware of the non-uniform memory access costs and the underlying hardware topology, incurring unnecessarily expensive memory accesses and bandwidth saturation. In this thesis, we show how these scalability and performance impediments can be solved by exploiting sharing among concurrent queries and incorporating NUMA-aware adaptive task scheduling and data placement strategies in the execution engine. Regarding sharing, we identify and categorize state-of-the-art techniques for sharing data and work across concurrent queries at run-time into two categories: reactive sharing, which shares intermediate results across common query sub-plans, and proactive sharing, which builds a global query plan with shared operators to evaluate queries. We integrate the original research prototypes that introduce reactive and proactive sharing, perform a sensitivity analysis, and show how and when each technique benefits performance. Our most significant finding is that reactive and proactive sharing can be combined to exploit the advantages of both sharing techniques for highly concurrent analytical workloads. Regarding NUMA-awareness, we identify, implement, and compare various combinations of task scheduling and data placement strategies under a diverse set of highly concurrent analytical workloads. We develop a prototype based on a commercial main-memory column-store database system. Our most significant finding is that there is no single strategy for task scheduling and data placement that is best for all workloads. In specific, inter-socket stealing of memory-intensive tasks can hurt overall performance, and unnecessary partitioning of data across sockets involves an overhead. For this reason, we implement algorithms that adapt task scheduling and data placement to the workload at run-time. Our experiments show that both sharing and NUMA-awareness can significantly improve the performance and scalability of highly concurrent analytical workloads on modern multi-core servers. Thus, we argue that sharing and NUMA-awareness are key factors for supporting faster processing of big data analytical applications, fully exploiting the hardware resources of modern multi-core servers, and for more responsive user experience.

EPFL2016

High Performance Transaction Processing on Non-Uniform Hardware Topologies

Danica Porobic

Transaction processing is a mission critical enterprise application that runs on high-end servers. Traditionally, transaction processing systems have been designed for uniform core-to-core communication latencies. In the past decade, with the emergence of multisocket multicores, for the first time we have Islands, i.e., groups of cores that communicate fast among themselves and slower with other groups. In current mainstream servers, each multicore processor corresponds to an Island. As the number of cores on a chip increases, however, we expect that multiple Islands will form within a single processor in the nearby future. In addition, the access latencies to the local memory and to the memory of another server over fast interconnect are converging, thus creating a hierarchy of Islands within a group of servers. Non-uniform hardware topologies pose a significant challenge to the scalability and the predictability of performance of transaction processing systems. Distributed transaction processing systems can alleviate this problem; however, no single deployment configuration is optimal for all workloads and hardware topologies. In order to fully utilize the available processing power, a transaction processing system needs to adapt to the underlying hardware topology and tune its configuration to the current workload. More specifically, the system should be able to detect any changes to the workload and hardware topology, and adapt accordingly without disrupting the processing. In this thesis, we first systematically quantify the impact of hardware Islands on deployment configurations of distributed transaction processing systems. We show that none of these configurations is optimal for all workloads, and the choice of the optimal configuration depends on the combination of the workload and hardware topology. In the cluster setting, on the other hand, the choice of optimal configuration additionally depends on the properties of the communication channel between the servers. We address this challenge by designing a dynamic shared-everything system that adapts its data structures automatically to hardware Islands. To ensure good performance in the presence of shifting workload patterns, we use a lightweight partitioning and placement mechanism to balance the load and minimize the synchronization overheads across Islands. Overall, we show that masking the non-uniformity of inter-core communication is critical for achieving predictably high performance for latency-sensitive applications, such as transaction processing. With clusters of a handful of multicore chips with large main memories replacing high-end many-socket servers, the deployment rules of thumb identified in our analysis have a potential to significantly reduce the synchronization and communication costs of transaction processing. As workloads become more dynamic and diverse, while still running on partitioned infrastructure, the lightweight monitoring and adaptive repartitioning mechanisms proposed in this thesis will be applicable to a wide range of designs for which traditional offline schemes are impractical.

EPFL2016

ATraPos: Adaptive Transaction Processing on Hardware Islands

Anastasia Ailamaki, Erietta Liarou, Danica Porobic, Pinar Tözün

Nowadays, high-performance transaction processing applications increasingly run on multisocket multicore servers. Such architectures exhibit non-uniform memory access latency as well as non-uniform thread communication costs. Unfortunately, traditional shared-everything database management systems are designed for uniform inter-core communication speeds. This causes unpredictable access latencies in the critical path. While lack of data locality may be a minor nuisance on systems with fewer than 4 processors, it becomes a serious scalability limitation on larger systems due to accesses to centralized data structures. In this paper, we propose ATraPos. a storage manager design that is aware of the non-uniform access latencies of multisocket systems. ATraPos achieves good data locality by carefully partitioning the data as well as internal data structures (e.g., state information) to the available processors and by assigning threads to specific partitions. Furthermore, ATraPos dynamically adapts to the workload characteristics, i.e., when the workload changes, ATraPos detects the change and automatically revises the data partitioning and thread placement to fit the current access patterns and hardware topology. We prototype ATraPos on top of an open-source storage manager Shore-MT and we present a detailed experimental analysis with both synthetic and standard (TPC-C and TATP) benchmarks. We show that ATraPos exhibits performance improvements of a factor ranging from 1.4 to 6.7x for a wide collection of transactional workloads. In addition, we show that the adaptive monitoring and partitioning scheme of ATraPos poses a negligible cost, while it allows the system to dynamically and gracefully adapt when the workload changes.

2014

Scaling Up Concurrent Analytical Workloads on Multi-Core Servers

Iraklis Psaroudakis

EPFL2016

High Performance Transaction Processing on Non-Uniform Hardware Topologies

Danica Porobic

EPFL2016