Switch Assisted Peer To Peer
Masters Thesis, University of Utah. August 2020.
Cloudlab provides a flexible testbed for research and education of future cloud
computing. On receiving a request for a research environment, Cloudlab identifies the
resources satisfying the constraints and assigns it to the user. Our thesis focuses on the
crucial step of loading the user's choice of disk image to the machine.
Frisbee is the system that facilitates the image loading task. A Frisbee server
process is spawned upon request for an image, Frisbee client running in target node
downloads the image and installs it. The Frisbee server process can serve multiple clients
simultaneously using multicast.
In this thesis, we study in detail the performance of the existing Frisbee system
and its shortcomings. We propose a new Switch Assisted Peer to Peer (SWP2P) transfer
system that addresses the shortcomings by significantly reducing the load on the Frisbee
server and the traffic on trunk links in Cloudlab. SWP2P builds an image availability
database on the switch to keep track of the availability among clients connected to it. The
system then uses the database to redirect requests to other clients instead of the Frisbee
server. We further evaluate the performance of the new and old system, where we show
that SWP2P transfer reduces the server and trunk load by a minimum of 21% to a
maximum of 96% depending on the level of concurrency among the clients.