Action Nom du fichier Description Taille Accès Licence Version de la ressource
Montrer tous les fichiers...


The conventional wisdom is that aggressive networking requirements, such as high packet rates for small messages and microsecond-scale tail latency, are best addressed outside the kernel, in a user-level networking stack. In particular, dataplanes borrow design elements from network middleboxes to run tasks to completion in tight loops. In its basic form, the dataplane design leverages sweeping simplifications such as the elimination of any resource management and any task scheduling to improve throughput and lower latency. As a result, dataplanes perform best when the request rate is predictable (since there is no resource management) and the service time of each task has a low execution time and a low dispersion. On the other hand, they exhibit poor energy proportionality and workload consolidation, and suffer from head-of-line blocking. This thesis proposes the introduction of resource management to dataplanes. Current dataplanes decrease latency by constantly polling for incoming network packets. This approach trades energy usage for latency. We argue that it is possible to introduce a control plane, which manages the resources in the most optimal way in terms of power usage without affecting the performance of the dataplane. Additionally, this thesis proposes the introduction of scheduling to dataplanes. Current designs operate in a strict FIFO and run-to-completion manner. This method is effective only when the incoming request requires a minimal amount of processing in the order of a few microseconds. When the processing time of requests is (a) longer or (b) follows a distribution with higher dispersion, the transient load imbalances and head-of-line blocking deteriorate the performance of the dataplane. We claim that it is possible to introduce a scheduler to dataplanes, which routes requests to the appropriate core and effectively reduce the tail latency of the system while at the same time support a wider range of workloads.