Exploiting data locality in cache-coherent NUMA systems
The end of Dennard scaling has caused a stagnation of the clock frequency in computers.To overcome this issue, in the last two decades vendors have been integrating larger numbers of processing elements in the systems, interconnecting many nodes, including multiple chips in the nodes and increasing...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2022 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/367546 |
| Acceso en línea: | https://hdl.handle.net/2117/367546 https://dx.doi.org/10.5821/dissertation-2117-367546 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Informàtica |
| Sumario: | The end of Dennard scaling has caused a stagnation of the clock frequency in computers.To overcome this issue, in the last two decades vendors have been integrating larger numbers of processing elements in the systems, interconnecting many nodes, including multiple chips in the nodes and increasing the number of cores in each chip. The speed of main memory has not evolved at the same rate as processors, it is much slower and there is a need to provide more total bandwidth to the processors, especially with the increase in the number of cores and chips. Still keeping a shared address space, where all processors can access the whole memory, solutions have come by integrating more memories: by using newer technologies like high-bandwidth memories (HBM) and non-volatile memories (NVM), by giving groups cores (like sockets, for example) faster access to some subset of the DRAM, or by combining many of these solutions. This has caused some heterogeneity in the access speed to main memory, depending on the CPU requesting access to a memory address and the actual physical location of that address, causing non-uniform memory access (NUMA) behaviours. Moreover, many of these systems are cache-coherent (ccNUMA), meaning that changes in the memory done from one CPU must be visible by the other CPUs and transparent for the programmer. These NUMA behaviours reduce the performance of applications and can pose a challenge to the programmers. To tackle this issue, this thesis proposes solutions, at the software and hardware levels, to improve the data locality in NUMA systems and, therefore, the performance of applications in these computer systems. The first contribution shows how considering hardware prefetching simultaneously with thread and data placement in NUMA systems can find configurations with better performance than considering these aspects separately. The performance results combined with performance counters are then used to build a performance model to predict, both offline and online, the best configuration for new applications not in the model. The evaluation is done using two different high performance NUMA systems, and the performance counters collected in one machine are used to predict the best configurations in the other machine. The second contribution builds on the idea that prefetching can have a strong effect in NUMA systems and proposes a NUMA-aware hardware prefetching scheme. This scheme is generic and can be applied to multiple hardware prefetchers with a low hardware cost but giving very good results. The evaluation is done using a cycle-accurate architectural simulator and provides detailed results of the performance, the data transfer reduction and the energy costs. Finally, the third and last contribution consists in scheduling algorithms for task-based programming models. These programming models help improve the programmability of applications in parallel systems and also provide useful information to the underlying runtime system. This information is used to build a task dependency graph (TDG), a directed acyclic graph that models the application where the nodes are sequential pieces of code known as tasks and the edges are the data dependencies between the different tasks. The proposed scheduling algorithms use graph partitioning techniques and provide a scheduling for the tasks in the TDG that minimises the data transfers between the different NUMA regions of the system. The results have been evaluated in real ccNUMA systems with multiple NUMA regions. |
|---|