Mobile App Acceleration via Fine-Grain Offloading to the Cloud

Motivation:

    Distributed shared memory known as DSM has long history in distributed system field however, never been introduced to mobile device due to latency, network bandwidth, power and computation overhead constrains. In this paper, the author purposes compressive offloading, which takes advantage of compressive sensing algorithm and therefore makes DSM possible between cloud servers and mobile devices.

Main point:

• The traditional DSM normally runs on cluster server with generic architecture, and requires overhead computation and communication to synchronize when replicating memory. Such limitation of traditional method would dramatically decrease the performance of traditional DSM running on mobile device.

• Compressive sensing algorithm uses random sampling technique of memory I/O signal and therefore compressively replicates the memory of mobile device. Such approach compared to traditional DSM has following advantages:

•  Reduce computational complexity encoding on mobile device and while the decoder side on cloud server has higher complexity

• Eliminate the need of communication to determine the delta of changed memory page, since such information is recovered using greedy algorithm and minimum-distance decoding.

• The offloading mechanism imitates the RPC system, however uses pointer passing and address translations instead of object marshaling.

• A shim layer is introduced into the running time environment on mobile device that functions as memory manager to replicate required memory data and redirect object method invocations to the cloud server. The layer takes control of memory management through its own alloc call.

• As the counterpart on the server side, whenever the application on mobile device is compiled, it is cross-compiled on the server. The server is on-call whenever is invocated.

• The performance of the method relies on the granularity of the work offloaded. Also, the sampling matrix on mobile side and the decoding algorithm server side has impact on the performance too in term of encoding and decoding time. 

• Finally, The offloading decision is made by shim layer on mobile device depends on the current battery level and network condition.

Trade-offs

• Introduce a traditional idea on the cloud computing field.

• DSM has fine partitioning of work and therefore incurs less disruption to the user when disconnected.

• Some compression algorithms can minimize computation on mobile device and also utilize the cloud server.

• DSM heavily relies on machine architecture; mobile device must be byte-order compilable with cloud server. 

• The paper does not give us how to determine when to offload, and how much work should be offloaded. 

• The performance of compression algorithm may vary from application to application.

• The tested application is not typical.