Sharing-aware Cloud-based Mobile Outsourcing

Motivation:

Smart phones and tablets have become an important part of human life and many applications can now be performed on these devices. However, battery life and computational power of these devices is constrained and therefore cannot be used to execute resource intensive applications. Computation outsourcing to external resources address this problem. So far, mobile outsourcing considered single application optimization and outsourcing to fixed local resources, as outsourcing to cloud or wide-area seemed to incur high latency. The paper suggests that the relation between different applications can be leveraged to increase the outsourcing performance. Also, cloud can be used for outsourcing the applications as its a scalable and elastic resource. The paper has proposed to use data mining techniques to detect data sharing across multiple applications. 


Main Points:

• There are two options for offloading the resource-intensive parts of applications. One is a local, fixed group of servers and the other is third party providers such as cloud. But cloud is not preferred because of the wide-area latency issues. But many applications on the phone can tolerate these latencies. 
• The paper focuses on outsourcing the computation to cloud rather than the existing work which focuses on local resources. Few reasons for this being, the resource richness, ease of data-sharing, and dynamic and scalable properties of cloud. 
• From the experiments conducted on image processing, local execution consumes more power and CPU usage than remote execution, due to the high computation requirement. Wide-area outsourcing is therefore feasible for certain compute-intensive applications. 
• Local static servers cannot handle the network overhead due to multiple requests(of same type) and are also not scalable. The paper presents a cloud-based offloading platform that overcomes these issues by having a provision for dynamic resource allocation and also achieves high performance by computation collocation. 
• The offloading client makes the offloading decisions based on previous executions and current network state. The offloading backend system has the code, server to offload and a manager.
• For many applications, most of the code and data is shared. Therefore, this shared data can be cached, so that multiple applications can use them, thereby reducing the communication overhead. 
• An ideal scheduling algorithm ensures high performance while achieving load balance and high resource utilization across offloading servers. By intelligently assigning the offloading request, the network overhead due to lack of data sharing among applications is reduced. Similarly, by dynamically provisioning the offloading servers, poor scalability problem due to static resource allocation is also addressed. 
• There are different approaches to sharing data like local, no-sharing, storage, communication and intra-server. The paper also talks about three different component placement strategies in order to reduce the communication overhead,which are: user-centric, app-centric, co-location(hybrid of both). Data mining techniques like Sequence Mining is used to find the temporal relationships in data. Dynamic provisioning in cloud consists of server creation and server merging.         

Trade-offs:

• Identifying shared data in a code is difficult. Using techniques like Sequence Mining, would only identify temporal relations in data. 
• The scheduling algorithm used to identify the server would select the least offloaded server, but that server might be far away, which would increase the latency. So other parameters should also be taken into consideration.