Enabling Opportunistic use of Transient Thin-Clients in Internet Suspend/Resume Ajay Surie The Internet Suspend/Resume (ISR) system today allows ubiquitous access to a user.s personal computing environment. A user.s environment is encapsulated in a parcel, which follows the user through the internet as she travels between locations. Since a parcel is represented in the system as a virtual machine image, which is comprised of a disk and memory image, the amount of data that needs to be transferred to the resume site can be on the order of gigabytes. One possible approach to reducing resume latency is to integrate the use of thin clients into the ISR system. In essence, when a user decides to resume over a weak link, the parcel can be checked out on the user.s behalf at a compute server, which is assumed to be well connected to the content servers which store user parcels. The compute server is then responsible for running the user.s parcel and providing the user with a thin client interface to the running parcel on the compute server. The user.s client machine can then, in the background, fetch the user.s parcel from the compute server, and when this process completes, transition from the possibly jittery thin client remote link to the (now) locally running parcel. My project will be to investigate the issues with this computing model and advance as possible towards creating an experimental prototype. A first step is to support thin client mode, the next step will be to integrate live migration. To accomplish this, I need to understand the ISR code base. Among questions I will work towards answering are the following: -When is it suitable for a user to request thin-client access? Should this process be automated? Initially, a simple bandwidth test could suffice. -Should all dirty state transactions be managed by the compute server? -Should the thin-client server (VNC) be run inside a user parcel or outside? Currently, only paravirtualized Xen supports live migration, so initially running the server inside the guest operating system might be easiest. -What is the best way to transition between thin and thick clients without disrupting the user? -How much does using a thin-client reduce resume latency? Are suspend times increased due to the distribution of dirty state between clients and compute servers?