Introducing heterogeneous farms in the CMS framework

The High Luminosity upgrade scheduled for 2026 will greatly increase the number of events per collision. Moore’s law will optimistically get a factor 4 performance gain, not enough to handle the luminosity increase without incurring significant costs and logistic issues as the actual system is already deeply optimized.
A possible solution is the creation of a heterogeneous cluster, a cluster in which different nodes have different kinds of co-processors available. This study tackles the problem of moving efficiently a task that can be executed on a co-processor to a different node in the cluster. As a first step, a standalone application to study the communication protocol between the three kinds of nodes has been implemented. Then, the protocol has been introduced into the framework by implementing new kinds of EDProducers and EDAnalyzers. Finally, the code has been benchmarked and the results analyzed.