Integrated PlayBack Sensing and Networked Control Vincenzo Liberatore
Integrated Play-Back, Sensing, and Networked Control Vincenzo Liberatore Division of Computer Science Research supported in part by NSF CCR-0329910, Department of Commerce V. Liberatore Control Playback TOP 39 -60 -04003, NASA NNC 04 AA 12 A, and an Ohio. ICE training grant.
Networked Control • Computing in the physical world • Components – Sensors, actuators – Controllers – Networks V. Liberatore Control Playback 2
Networked Control • Enables – – – – Industrial automation [BL 04] Distributed instrumentation [ACRKNL 03] Unmanned vehicles [LNB 03] Home robotics [NNL 02] Distributed virtual environments [LCCK 05] Power distribution [P 05] Building structure control [SLT 05] • Merge cyber- and physical- worlds – Networked control and tele-epistemology [G 01] • Sensor networks – Not necessarily wireless or energy constrained – One component of sense-actuator networks V. Liberatore Control Playback 3
Information Flow • Flow – Sensor data – Remote controller – Control packets • Timely delivery – Stability – Safety – Performance V. Liberatore Control Playback 4
Playback Buffers [Infocom 2006] • Play-back buffers – Main objective – Smooths out network non-determinism • Multimedia buffers – Important source of inspiration – Physics versus multimedia quality – Playback delay computed in advance • Affects control signal computation – Round-Trip Times • TCP RTO – Another source of inspiration – Large time-out cost V. Liberatore Control Playback 6
Algorithm V. Liberatore Control Playback 7
Main Ideas • Predictable application time – If control applied early, plant is not in the state for which the control was meant – If control applied for too long, plant no longer in desired state • Keep plant simple – Low space requirements • Integrate Playback, Sampling, and Control V. Liberatore Control Playback 8
Algorithm • Send regular control – Playback time • Late playback okay – Expiration • Piggyback contingency control V. Liberatore Control Playback 9
Deadwood packets • Old – Received after the expiration time • Out-of-order – Later control more appropriate for current plant state • Would get us into a deadlock – New packet resets the playback timer – Keep resetting until no signal applied – “Quashed” packet • Discard! controller plant Playback delay V. Liberatore X Control Playback X 10
Countermand control • Scenario – Packet i+1 overtakes packet I – i+1 << i – Likely caused by delay spike • New signal countermands previous one controller plant Playback delay V. Liberatore i+1 Control Playback i 11
Playback delays • Modular component • Compute playback delay and sampling period T • Use short term peak-hopper [EL 04] – Original peak-hopper for TCP RTO • Too conservative for networked control – Aggressively attempt to decrease • Aggressively attempt to decrease T • Add upper bound on playback delay – Avoid dropping deadlock packets – Bound ≤ T+RTT • Caps and T • Must estimate lower-bound on RTT – Use symmetric of peak-hopper – Add negative variability estimate to compensate for short-term memory V. Liberatore Control Playback 12
Playback Delays (I) Calculate current RTT variability Positive variability coefficient if Negative variability coefficient then Update min RTT estimate Age min RTT estimate Calculate V. Liberatore Control Playback 13
Playback Delays (II) if then Attempt to avoid quashed packets else Increase sampling period V. Liberatore Control Playback 14
Control Pipes • Bandwidth and delays – is playback delay – T is sampling period • 1/T proportional to bandwidth • Control pipe – T « – Multiple in-flight packets • Pipe depth – Bound by constraint ≤ T+RTT – Keep pipe predictable V. Liberatore Control Playback 15
Observer • Estimate future plant state – Plant sample current state, including local variables – Keep log of outstanding control packets • Assumption on packet delivery – Future packet delivery is uncertain • Purge from log – Old packets – Packet that should be overtaken by new control • Countermands signals generated when delay spike is transient – Out-of-order packets V. Liberatore Control Playback 16
Evaluation V. Liberatore Control Playback 17
Network Model • Simulated network • Losses: Gilbert model • Delays – – Shifted Gamma distribution Heavy tail Low probability of out-of-order delivery Correlate delays to introduce delay spikes • Wide-area implementation • Use RT scheduling whenever possible • Use otherwise unloaded machines – RT made little difference • Host worldwide, heterogeneous conditions V. Liberatore Control Playback 18
Plant • Scalar linear plant – – Plant state x(t) Input u(t) (control) Output y(t) Disturbances v(t), w(t) • Akin to white noise • Deadbeat controller – Aggressive V. Liberatore Control Playback 19
Metrics • Metrics – Root-mean square output – Output: 99 -percentile • Comparison – Open-loop plant u(t)=0 – Proportional controller (no buffer) – Proportional controller with constant delays V. Liberatore Control Playback 20
Plant output Open Loop V. Liberatore Play-back Control Playback 21
Packet losses Figure 8 V. Liberatore Control Playback 22
Sampling period Root-mean-square error V. Liberatore Control Playback Imperfection of the control pipe 23
Conclusions (I) • Sense-and-Respond – Merge cyber-world and physical world – Critically depends on physical time • Playback buffers integrated with – Sampling (adaptive T) – Control (expiration times, performance metrics) • Packet losses – Reverts to open loop plant (contingency control) V. Liberatore Control Playback 24
Conclusions (II) • Playback delay – Adapts to network conditions • Sampling period T – Avoids imperfection of control pipe • Simulations and emulations – Low variability around set point – Robust V. Liberatore Control Playback 25
- Slides: 24