Malleability in resource allocation for improved system efficiency in high-performance computing

高性能计算（HPC）系统的环境影响和二氧化碳排放的很大一部分可归因于其制造及其操作（包括运行空闲）。因此，一旦安装了系统，就必须尽可能接近满负荷使用，并始终最大限度地提高科学吞吐量，以便在系统的货币和碳成本方面获得最佳投资回报。然而，这个非常理想的100%利用率在实践中几乎不可能实现。系统的工作负载由其资源分配器管理，该资源分配器尝试从提交队列（用户不断向其添加新作业）中放置作业，以填补可用资源中的空白。并不总是能够达到完美的工作安排，因此，资源闲置。资源分配中的可塑性引入了资源的概念，用户在作业提交时已经请求的计算核或节点的数量（计算核或节点的数量，或者甚至系统的数量）不是固定的，并且如果该改变意味着作业可以被调度来运行，并且因此完成，MIRA（“Malleability In Resource Allocation for improved system efficiency in high-performance computing”）将研究单个系统内以及跨多个系统的计算资源分配的延展性概念，以提高整体系统利用率和科学吞吐量，从而最大化可以实现的“每焦耳科学”。

A significant part of the environmental impact and CO2 emissions of a high-performance computing (HPC) system can be attributed to its manufacturing as well as its operation (including running idle). Once a system has been installed, it is therefore imperative that it is used as close to full capacity as possible and that science throughput should be maximised at all times, in order to get the best return on investment on both the monetary and carbon cost of the system. This highly desirable 100% utilisation rate is however near impossible to achieve in practice. The workload of a system is managed by its resource allocator, which attempts to place jobs from a submission queue (that users continuously add new jobs to) to fill gaps in the available resources. It is not always possible to attain perfect job placement and as a result, resources sit idle.Malleability in resource allocation introduces the concept that the resources (the number of compute cores or nodes, or even the system) that have been requested by a user at job submission time are not fixed and can be changed if this change means a job can be scheduled to run, and thus complete, sooner.MIRA ("Malleability In Resource Allocation for improved system efficiency in high-performance computing") will investigate the concept of malleability in compute resource allocation within a single system as well as across multiple systems, to improve overall system utilisation and science throughput, thereby maximising the "science per Joule'' that can be achieved.