Mobile App Acceleration via Fine-Grain Offloading to the Cloud

MOTIVATION :

• Offloading of computations to the cloud is one of the important measures to alleviate the power constraints of mobile devices. This paper proposes “Compressive Offloading”, that uses  compressive sensing in a distributed shared memory setting.

• Applications that are statically partitioned into device specific and cloud specific components make runtime optimization infeasible. Usage of DSM(Distributed Shared Memory) system with continuous memory replication, solves this problem. But implementation of DSM is hard due to power, latency and computational overhead constraints. The paper proposes “UpShift”, a platform that uses compressive sensing and implements DSM.

KEY POINTS :

• Offloading is based on Compressive Sensing, a sparse random sampling technique for acquiring and reconstructing a signal. Compressive replication supported by UpShift compressively samples memory I/O signals. 

• The UpShift shim layer serves two purposes. It acts as replication agent and also redirects methods to the cloud. On the cloud, a daemon serves two purposes namely decoding/updating replicated memory pages and services the redirected object method invocations.

• Offloading depends on various factors like device’s battery levels, network conditions, cloud service provider’s capacity. End user can also decide whether to allow or disallow offloading. 

• The paper implements prototype for ios ecosystem in iPad3 and uses Amazon EC2 as cloud server. 

• OFFLOADING : Ios was chosen because its Objective C allows efficient memory management. The UpShift shim layer implements memory manager and does replication. The daemon provides all required class definitions for the objects in the server.

• REPLICATION : Parameters should be determined in such a way that replication latency is minimized. The partial DCT matrix is seen to be the best choice for sampling matrix. The shortest decoding time is provided by AIHT decoding algorithm.

TRADE OFFS

• Latency minimization is difficult because it is based on factors like encoding/compression time, decoding/decompression time, network transmission time which are dependent.

• The advantage of using Compressive Replication is that it uses lesser bandwidth and produces more stable traffic stream.

• According to the prototype implemented in the paper, 60% longer battery life was seen. This result is for a particular app on a particular os platform. It may vary for other devices and apps.

• There is strict restriction to use only Objective C objects in prototype.

• No clear discussion regarding offloading of security intensive applications.