Dynamic VM Synthesis for Cloudlet -ISTC Retreat Poster- Kiryong Ha, Padmanabhan S Pillai, Mahadev Satyanarayanan

Motivation 1.Rich applications using video and voice are arising in mobile context, but they are too expensive to run on the client alone 2.And it cannot be handled in current Cloud computing –Cloud computing is highly centralized approach, so that it is too far from you  High latency and limited bandwidth across WAN 3.Let’s bring the cloud to the mobile users  Cloudlet! Olympus Mobile Eye Trek Wearable Computer Handtalk Wearable Glove Tablet Smartphone Coffee shop Cloudlet WAN to distant cloud on Internet

Key Challenges –Management Decentralization requires extreme standardization for minimal system management –Personalization Much customization needed, even for commercial apps Ex) personal preferences, speech tuning, domain-specific vocabulary. Out Approach –Use a personal VM and take it with you everywhere

Dynamic VM Synthesis Problems in VM : too big to carry, transfer, and launch Solution: Dynamic VM Synthesis –Base VM: Cloudlet pre-fetches large, widely-used VM –VM overlay: Mobile device delivers small patch just before use –Cloudlet discards VM after use/Or caches for future reuse Mobile Device Nearby Computing Resource 1) Transfer Overlay VM 3) Launched VM 2-2) Synthesis with Base VM 2-1) Decompress Application 4) Cloudlet Service

Prototype System Intel Xeon 6 8GB Main Memory n Wireless Smartphone

Prototype Applications Object-Recognition (MOPED) –Object recognition and pose estimation program (CMU RI) Face Detection and Recognition –OpenCV Application characteristics –Requires high computation and interactivity –Future Candidates Indoor navigation Language translation …

Experiment Overlay Size Application Program size (MB) OS Base Disk (GB) * Base Memory (MB) ** Compressed Ov erlay (MB) Gimp55.2Ubunutu MOPED27.5Ubunutu FACE17.66Windows XP Null0Ubunutu * 8GB Disk with qcow2 ** Memory snapshot of 2GB Main Memory

Experiment VM Synthesis Time Application Overlay Transfer (s) Decompressio n (s) Apply Delta (s)Run KVM (s)Total (s) Gimp MOPED FACE Null * * Null case is for comparison and does not install anything

Work in progress Goal: Reducing VM synthesis time –Discovery, transfer, and launch to happen within 5 seconds Work in progress –Pipe-lining Overlapping multiple steps with a pipelined fashion –Exploiting parallelism –Multi-layer overlay Dividing overlay into multi-layer and caching some for next run. –Launch before completing synthesis Use base VM information to start VM even before getting all overlay.