SNARF

Heungsik Eom, Pierre St. Juste, Renato Figueiredo, Omesh Tickoo, Ramesh Illikkal, Ravishankar K. Iyer2012 · DOI

TL;DR

This paper introduces SNARF, a novel framework that enables mobile devices to seamlessly offload computation to remote resources like personal computers or cloud services. SNARF addresses the limitations of mobile device processing power and battery life by providing a unified API and an intelligent runtime that dynamically discovers and utilizes available remote capabilities, improving performance and energy efficiency.

What Problem Does This Paper Address?

Mobile devices are becoming increasingly powerful, but their inherent limitations in size, battery life, and processing power restrict the complexity of applications they can run efficiently. To overcome this, researchers have explored "computation offloading," where computationally intensive tasks are sent to more powerful remote resources. However, a significant challenge lies in architecting a system that allows mobile applications to easily discover these remote services, communicate with them securely, and abstract away the complexities of network communication and resource management. Existing solutions often assume static resource discovery or require significant application-level modifications. SNARF aims to bridge this gap by providing a flexible, dynamic, and user-friendly framework for opportunistic offloading.

Methods

SNARF is built upon a layered architecture. At its core is a virtual networking substrate that leverages SocialVPN [3], a decentralized, self-configuring VPN service. SocialVPN provides a unified IP network layer that supports standard TCP/IP communication and enables the use of widely-used resource discovery protocols like Universal Plug-and-Play (UPnP). This virtual network allows devices to discover each other regardless of their physical location or network topology.

On top of this, SNARF implements an offload infrastructure and a set of APIs. The infrastructure manages the aggregated platform resources, tracks dynamic resources, and handles optimal offload scheduling based on performance requirements and available power. The APIs provide a seamless interface for applications, abstracting away the complexities of remote interaction. SNARF categorizes application assistance for offloading into four levels: minimum, maximum, medium, and API-assisted, allowing developers to choose the level of integration. For computation offload itself, SNARF uses a REST-like RPC system, encapsulating method calls and data within HTTP headers and bodies for efficient transfer.

Datasets

The paper does not explicitly mention specific datasets for training. However, for evaluation, they used a set of JPEG images of varying sizes (26KB, 674KB, and 1,743KB) to test the performance and energy consumption of an image processing workload (Sobelfilter).

Key Insights and Ideas

  • SocialVPN as a Foundation: The integration of SocialVPN is a key enabler, providing a robust and flexible virtual networking layer that supports both local and wide-area discovery and communication, crucial for dynamic offloading.
  • UPnP for Seamless Discovery: Leveraging UPnP for resource discovery, tunneled over the SocialVPN, allows for unmodified UPnP applications and middleware to be used, simplifying integration.
  • Opportunistic Offloading: The framework's intelligence lies in its ability to monitor the execution environment and opportunistically offload tasks based on performance needs and power availability, rather than requiring explicit user intervention for every offload decision.
  • API Abstraction: The well-defined API layers are critical for hiding the complexities of remote resource management from application developers, promoting portability and ease of use.
  • Performance vs. Energy Trade-offs: The experiments highlight that offloading is not always a clear win for both performance and energy. For small workloads, local processing might be more energy-efficient due to the overhead of setting up remote connections. However, for larger, compute-intensive tasks, offloading offers significant benefits.

Key Results

  • Service Discovery Time: SNARF demonstrated efficient UPnP service discovery across wide-area networks (U. Chicago, UC San Diego, U. Texas) with discovery times ranging from 27ms to 57ms. It also showed that Wi-Fi connections generally have longer service discovery times than wired LAN connections.
  • Image Processing Performance: For larger JPEG images (674KB and 1,743KB), offloading the Sobelfilter task to a local workstation or an Amazon EC2 GPU node significantly reduced execution time compared to local processing on an Android tablet. For the smallest image size, local processing was faster.
  • Energy Consumption: For larger image sizes, offloading to remote resources consumed less energy than local processing on the mobile device. However, offloading to the EC2 GPU cluster consumed more energy than offloading to the local workstation, attributed to Wi-Fi power characteristics and increased latency.

Limitations and Future Work

  • Remoting Overhead: The authors acknowledge that remoting overhead exists and is a significant factor, especially for smaller workloads. Future work includes further characterizing this overhead and investigating architectural enhancements to reduce it.
  • SocialVPN Cost: The paper mentions that the cost of secure connections for mobile devices (in terms of energy consumption, bandwidth, and latency) using SocialVPN is an area for future investigation.
  • Dynamic Offloading and Scheduling: While SNARF provides opportunistic offloading, further work is needed to dynamically discover remote capabilities and intelligently offload to cloud resources across various application and connectivity scenarios.
  • Workload Characterization: The authors plan to continue characterizing SNARF's benefits using different workloads relevant to tablet and smartphone use cases.

Summary updated

Abstract

The diminishing size and battery requirements of mobile devices restrict the scope of computations possible on such devices and motivate approaches that support the selective offloading of computations to remote resources. With a variety of resources available to potentially host offloaded computations -- such as cloud-provisioned resources, and devices within a user's personal or social network -- a key challenge lies in architecting a framework that enables applications to seamlessly discover available services, effectively and securely communicate with them, and be presented with API interfaces that hide the complexities associated with managing the interactions with a remote device from applications and present the abstraction of a local device. In this paper, we outline a framework that addresses these challenges by layering APIs and an offload infrastructure upon a virtual networking substrate that supports TCP/IP networking and widely-used resource discovery protocols. An intelligent runtime scheduling layer monitors the execution environment and provides opportunistic remote offloads based on the performance requirements, offload benefits and expendable power. We demonstrate the feasibility of the approach through experiments that evaluate end-to-end application execution times and energy consumption in offloaded mobile devices, as well as the ability to support universal plug-and-play (UPnP) resource discovery in both local- and wide-area environments.
BibTeX
@misc{eom2012snarf,
  author = {Heungsik Eom and Pierre St. Juste and Renato Figueiredo and Omesh Tickoo and Ramesh Illikkal and Ravishankar K. Iyer},
  title = {{SNARF}},
  year = {2012},
  doi = {10.1145/2342509.2342516},
}
Figure