Energy-Optimal Sampling for Edge Computing Feedback Systems: Aperiodic Case Vishnu Narayanan Moothedath, Division of Information Science and Engineering KTH Royal Institute of Technology, Sweden Contents • Introduction and applicability • The sampling challenge • Related works and our work on optimum periodic sampling • System Model • Aperiodic Sampling: • Results 2023-05-12 Introduction • Edge Systems: Faster than local nearer than cloud • Reasons to go for edge – – – – Low processing power at the device power sensitive device Size of device Required node-cooperation • Relevance in modern systems Eg. URLLC – Needs ms/sub-ms latency 2023-05-12 Edge-Based Event Detection with Feedback • Wearable Cognitive Assistants WCA Terminal Edge Device Cloud Human User • Requirement: Capture Essential Events ASAP WCA Task Instructions 2023-05-12 The Sampling Challenge • High frequency sampling * Communication * Processing ⇒ Unnecessary communication, processing ⇒ Large amount of data in the pipeline • Low frequency sampling ⇒ Reduced system responsiveness ⇒ Increase in task execution time.* ⇒ Idle energy wastage • Energy/Responsiveness suffers on both extremes • Aim: Find the optimum sampling interval. * M. Olguín Muñoz, R. Klatzky, J. Wang, P. Pillai, M. Satyanarayanan and J. Gross. "Impact of delayed response on wearable cognitive assistance", PLOS ONE, vol. 16, no. 3, p. e0248690, 2021. 2023-05-12 * Detection Delay * Idle Energy Sampling Interval Related Works • Full/ Partial Offloading offload energy usage • Event detection works with energy consideration 1. Optimising sensor technologies 2. Node cooperation 3. Selective sensor activation Our work -- optimal sample generation that reduces total amount of data in the pipeline. 2023-05-12 Periodic Sampling: A quick look • V. N. Moothedath, J. P. V. Champati and J. Gross, "Energy Efficient Sampling Policies for Edge Computing Feedback Systems," in IEEE Transactions on Mobile Computing. • V. N. Moothedath, J. P. Champati and J. Gross, "Energy-Optimal Sampling of Edge-Based Feedback Systems," 2021 IEEE International Conference on Communications Workshops. Parameters chosen from previous experimental data, works on WCA and/or Google Glass. 2023-05-12 System Model : Aperiodic Sampling • TTE: Time to Event • TTF: Time to Feedback • TTE is Random. – What distribution? • Deterministic communication, processing 2023-05-12 Objective: Energy Minimisation • Energy is a function of 1. Number of samples 2. Wait time • Relevant part is called Energy Penalty • Aim: Minimise energy penalty for a given distribution 2023-05-12 TTE Distribution: Logic behind the choices • VAS (Video Analytics system) – Rayleigh/Exp-Gaussian/Log-Normal – E.g.: WCA, video surveillance TTE Distribution of a WCA Image courtesy: M. Olguín Muñoz, R. Klatzky, J. Wang, P. Pillai, M. Satyanarayanan and J. Gross. "Impact of delayed response on wearable cognitive assistance", PLOS ONE, vol. 16, no. 3, p. e0248690, 2021. 2023-05-12 Aperiodic Sampling • Conditions of a valid sampling interval 1. Positive 2. Decreasing - Condition1 - Condition2 • Recursive solution with a given first sampling interval • How to find optimum first sampling interval? 2023-05-12 Optimum t1 : A visual intuition {tn} vs. n for different values of t1. Two of the sequences that are valid upto n=15 are highlighted. The rest are invalid 2023-05-12 Optimum t1 : How to find it. Design a bisection search to find the optimum t1. At each iteration a simple check for validity is performed This is possible because of their proven properties below. • 1: Plot ({tn} vs. n) does not cross each other for different n. • 2: All {tn} below those violate Condition1 are invalid. All {tn} above those violate Condition2 are invalid. 2023-05-12 • 3: Optimum {tn} always lies between two sequences that violates different conditions • 4: The algorithm achieves a result arbitrarily close to the optimum, with a convergence rate of 2-n Aperiodic Sampling: Results Penalty Vs Mean TTE for Optimum periodic policy and Optimum aperiodic policy 2023-05-12 Aperiodic Sampling: Results Penalty reduction attained by optimum aperiodic policy over optimum periodic policy Vs. System parameters 2023-05-12 Thank You.