1. Adaptive Offloading for Pervasive Computing
AmiN Saremi
7/6/2005

2. Introduction Pervasive Computing
Challenge: run complex applications on resource-constrained mobile device such as PDA.
Solutions
rewrite applications according to the resource capacity of each mobile device
application-based or system-based adaptations
Adaptive Offloading

3. Decision Making Problems for Adaptive Offloading
The offloading inference engine should trigger offloading at the right time and offload the right program objects to achieve low offloading overhead and efficient program execution.
adaptive offloading triggering
efficient application partitioning

4. Solution Overview Our assumptions
the application is written in an object-oriented languages
the user’s environment contains powerful surrogates and plentiful wireless bandwidth

6. Offloading inference engine does not require any prior knowledge about an application’s execution pattern or the runtime environment’s resource status
offloading inference engine employs the Fuzzy Control model as the basis for the offloading triggering inference module
selects an effective application partitioning from many possible partition plans
Memory constraint or CPU speed

7. Distributed Offloading Platform
application execution monitoring
Application execution graph
Each graph node represents a Java class
memory size, AccessFreq, Location, IsNative
Each graph edge represents the interactions between the objects of two classes
InteractionFreq, BandwidthRequirement

8. Candidate Partition Plan Generation
Resource Monitoring
mobile device, the surrogate, and the wireless network
available memory in the Java heap, wireless bandwidth and delay
Candidate Partition Plan Generation

9. Surrogate Discovery
Transparent RPC Platform

10. Adaptive Offloading Inference Engine
Overhead of offloading
transferring objects between the mobile device and the surrogate
performing remote data accesses and function invocations over a wireless network
Offloading Triggering Inference
examines the current resource and the available resources
Decides whether offloading should be triggered
decides what level of resource utilization

11. simple threshold-based approach
“if the current amount of free memory on the mobile device is less than 20% of its total memory, then trigger offloading and offload enough program objects to free up at least 40% of the mobile device’s memory.”
Fuzzy Control model
linguistic decision-making rules provided by system or application developers
membership functions
generic fuzzy inference engine based on fuzzy logic theory

12. offloading memory size
Offloading rules
if (AvailMem is low) and (AvailBW is high) then NewMemSize := low;
if (AvailMem is low) and (AvailBW is moderate) then NewMemSize := average;
if (AvailMem is high) and (AvailBW is low) then NewMemSize := high;
If any of these rules is matched by the current system conditions, the offloading inference engine triggers offloading
offloading memory size
current memory consumption - new memory utilization

13. Mappings between numerical and linguistic values for each linguistic variable

15. Application Partition Selection
considering the target memory utilization on the mobile device
multiple offloading requirements
minimizing wireless bandwidth overhead
minimizing average response time
minimizing total execution time
For each neighbor node Vk of Vi
B i,k to denote the total amount of data traffic transferred between Vi and Vk,
F i,k to define a total interaction number,
MS k to represent the memory size of Vk.

16. For cost metrics Ck and Cl : Ck >=Cl if and only if
Splitting Large Classes

17. Trace-Driven Simulation Experiments
For comparison
algorithm least recently used (LRU)
Split Class
Fuzzy Trigger
our approach

19. References
Xiaohui Gu, Alan Messer, Ira Greenberg, Dejan Milojicic, Klara Nahrstedt, “Adaptive Offloading for Pervasive Computing”, IEEE Pervasive Computing Magazine 2004.
X. Gu, K. Nahrstedt, A. Messer, I. Greenberg, and D. Milojicic, “Adaptive Offloading Inference for Delivering Applications in Pervasive Computing Environment”, Proc. of IEEE International Conference on Pervasive Computing and Communications (PerCom 2003), Dallas-Fort Worth, Texas, March 2003.