Leveraging disaggregated accelerators and non-volatile memories to improve the efficiency of modern datacenters
(English) Traditional data centers consist of computing nodes that possess all the resources physically attached. When there was the need to deal with more significant demands, the solution has been to either add more nodes (scaling out) or increase the capacity of existing ones (scaling-up). Worklo...
| 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:2117/380809 |
| Acceso en línea: | https://hdl.handle.net/2117/380809 https://dx.doi.org/10.5821/dissertation-2117-380809 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Informàtica |
| Sumario: | (English) Traditional data centers consist of computing nodes that possess all the resources physically attached. When there was the need to deal with more significant demands, the solution has been to either add more nodes (scaling out) or increase the capacity of existing ones (scaling-up). Workload requirements are traditionally fulfilled by selecting compute platforms from pools that better satisfy their average or maximum resource requirements depending on the price that the user is willing to pay. The amount of processor, memory, storage, and network bandwidth of a selected platform needs to meet or exceed the platform requirements of the workload. Beyond those explicitly required by the workload, additional resources are considered stranded resources (if not used) or bonus resources (if used). Meanwhile, workloads in all market segments have evolved significantly during the last decades. Today, workloads have a larger variety of requirements in terms of characteristics related to the computing platforms. Those workload new requirements include new technologies such as GPU, FPGA, NVMe, etc. These new technologies are more expensive and thus become more limited. It is no longer feasible to increase the number of resources according to potential peak demands, as this significantly raises the total cost of ownership. Software-Defined-Infrastructures (SDI), a new concept for the data center architecture, is being developed to address those issues. The main SDI proposition is to disaggregate all the resources over the fabric to enable the required flexibility. On SDI, instead of pools of computational nodes, the pools consist of individual units of resources (CPU, memory, FPGA, NVMe, GPU, etc.). When an application needs to be executed, SDI identifies the computational requirements and assembles all the resources required, creating a composite node. Resource disaggregation brings new challenges and opportunities that this thesis will explore. This thesis demonstrates that resource disaggregation brings opportunities to increase the efficiency of modern data centers. This thesis demonstrates that resource disaggregation may increase workloads' performance when sharing a single resource. Thus, needing fewer resources to achieve similar results. On the other hand, this thesis demonstrates how through disaggregation, aggregation of resources can be made, increasing a workload's performance. However, to take maximum advantage of those characteristics and flexibility, orchestrators must be aware of them. This thesis demonstrates how workload-aware techniques applied at the resource management level allow for improved quality of service leveraging resource disaggregation. Enabling resource disaggregation, this thesis demonstrates a reduction of up to 49% missed deadlines compared to a traditional schema. This reduction can rise up to 100% when enabling workload awareness. Moreover, this thesis demonstrates that GPU partitioning and disaggregation further enhances the data center flexibility. This increased flexibility can achieve the same results with half the resources. That is, with a single physical GPU partitioned and disaggregated, the same results can be achieved with 2 GPU disaggregated but not partitioned. Finally, this thesis demonstrates that resource fragmentation becomes key when having a limited set of heterogeneous resources, namely NVMe and GPU. For the case of an heterogeneous set of resources, and specifically when some of those resources are highly demanded but limited in quantity. That is, the situation where the demand for a resource is unexpectedly high, this thesis proposes a technique to minimize fragmentation that reduces deadlines missed compared to a disaggregation-aware policy of up to 86%. |
|---|