Caladan: a distributed meta-OS for data center disaggregation
File(s)
Published version
Author(s)
Type
Conference Paper
Abstract
Data center resource disaggregation promises cost savings by pool-ing compute, storage and memory resources into separate, net-worked nodes. The benefits of this model are clear, but a closer lookshows that its full performance and efficiency potential cannot beeasily realized. Existing systems use CPUs pervasively to interface ar-bitrary devices with the network and to orchestrate communicationamong them, reducing the benefits of disaggregation.In this paper we presentCaladan, a novel system with a trusteduni-versal resource fabricthat interconnects all resources and efficientlyoffloads the system and application control planes to SmartNICs,freeing server CPUs to execute application logic. Caladan offersthree core services: capability-driven distributed name space, virtualdevices, and direct inter-device communications. These servicesare implemented in a trustedmeta-kernelthat executes in per-nodeSmartNICs. Low-level device drivers running on the commodity hostOS are used for setting up accelerators and I/O devices, and exposingthem to Caladan. Applications run in a distributed fashion acrossCPUs and multiple accelerators, which in turn can directly performI/O, i.e., access files, other accelerators or host services. Our dis-tributed dataflow runtime runs on top of this substrate. It orchestratesthe distributed execution, connecting disaggregated resources usingdata transfers and inter-device communication, while eliminatingthe performance bottlenecks of the traditional CPU-centric design.
Date Issued
2020-04-27
Date Acceptance
2020-03-18
Citation
2020
Copyright Statement
© 2020 The Author(s).
Identifier
https://www.youtube.com/watch?v=KsDJE3NB6zE&feature=youtu.be
Source
Systems for Post-Moore Architectures (SPMA) 2020
Publication Status
Published
Start Date
2020-04-27