Mobile App Acceleration via Fine-Grain Offloading to the Cloud

Motivation:

In the present-day, the relatively slow development of mobile device hardware becomes the limitation of complex programs & applications. Computation offloading is one strategy for this scenario, and this strategy is already manifest in some systems, such as Siri and Google Now. However, shifting any arbitrary workloads can’t be implemented in these services at run time, while DSM, a solution to this problem, has computation and network overhead. This paper focus on Upshift, a platform which not only can takes advantage of DSM, but also can gain benefits for both mobile devices and servers, especially for ARM-based ones.

Main points:

• Fine granularity partitioning is more stable and efficient than coarse one, because it incurs less disruption and reduces latency which helps improve server’s utilization.
• Compressive offloading is based on compressive sensing & replica, which avoid communication cost for recovering during decoding. Also, this method shifts computation complexity from encoding to decoding, which is beneficial for both mobile devices and servers.
• A shim layer one the mobile device likes a replication agent, and intercepts and redirects object method to server on cloud. A daemon on cloud server will process decoding and updating, and service what received from mobile device. It is fast for these two Upshift components control the entire replication and offloading at both local and remote end-points, and it can pass pointers and perform address translation wherever appropriate.
• For the prototype that described in this paper, to entirely control memory management is important, thus IOS is a good choice with its lingua franca and Objective-C.
• Also, to minimize replication latency is important, since it is not worth implementing replication if the total response time is still large due to high latency.
• Encoding/compression time, network transmission time and decoding/decompression time are not independent, e.g. a quick encoding time can give less compression ratio, increasing the network cost, and this may also influence decoding time.
• An example that performs handwriting recognition of Chinese characters, uses input size, complexity, delta and number of samples as parameters, showing the gain of response time and battery efficiency.

Trade-off:

• A novel application, which takes advantage of DSM, gains benefits for both mobile systems and servers by compressive sensing, which reduced the data need to be sent.
• This prototype gives us a new method for offloading, but it didn’t determine when and how to decide offloading.
• Compressive sensing & replication is a useful way to shift computation, but more experiments should be done for different data types at both encoding and decoding sides, which will make these algorithms & method more convincing.
• Different platform & systems should be tested to make this method universal.
• Both stability and bandwidths in different network should be taken into consideration.