SWiM QoS Lessons from Multimedia David Maier OGI School of Science & Engineering Oregon Health & Science University
SWiM Multimedia QoS Work with Richard Staehli and Jonathan Walpole Lessons If you are going to degrade, do so in preferred and thrifty manner: Least reduction in perceived quality for maximum reduction in resource usage Error is multifaceted: Different kinds of error more or less objectionable for different tasks, e.g., lower resolution vs. lower frame rate Software adaptivity is possible, but tricky to tune.
SWiM Our Model ContentViewPresentation Error = Ideal vs. Actual
SWiM Might Apply to Stream Queries ContentViewPresentation Error = Ideal vs. Actual Input Stream Query Output Stream
SWiM More Than One Way to Explain Error Amplitude Shift Lag Drift Quantization
SWiM Error Model Error model consists of one or more error components (e.g., amplitude, shift) An error component can be scaled by a coefficient (e.g., amount of shift) Error interpretation: expressing error between ideal and actual using error components E total = c 1 ·E 1 + c 2 ·E 2 + … + c n ·E n
SWiM Can Have More Than One Interpretation of an Error Amplitude Lag Amplitude Lag
SWiM Uses of Error Model Define combined quality bound 0.8*c amp + 0.2*c lag min over all interpretations Define limits on individual components State user preferences: degrade resolution before introducing lag Pre-compute effect of different load shedding options on error