80 likes | 185 Views
Research Idea. Xi Chen. Idea 1. Hybrid Control Scheme: Bandwidth utilization Control + Delay Jitter Control Motivation:
E N D
Research Idea Xi Chen
Idea 1 • Hybrid Control Scheme: Bandwidth utilization Control + Delay Jitter Control • Motivation: • When bandwidth utilization is low (the traffic does not reach the peak bandwidth), there is no queue or jitter. However, it is not what we desired. bandwidth utilization control is necessary to avoid waste of resources. • When the bandwidth is over 1, the bandwidth control will regulate it to 1, which suffers from the control saturation, then it can not enforce the desired bandwidth utilization with fast convergent speed. At this point, since the traffic exceed the peak bandwidth setpoint, excess frames are queued for later transmission, which will enlarge the delay to the frame. We want to control the delay jitter to ensure satisfactory QoS of system. In this way, we should enable the queue delay jitter control.
Idea 1 (continue) • Method: • For each control, we establish a control loop. Since the two loops have mutual exclusive control saturation, they can compensate each other’s fault.
Idea 1 (continue) • Control Method • Delay Jitter Control Loop: measured input is delay jitter (can be reflected to the queue length), control input is source rate . Method: RLS model estimator + Minimum Variance Controller • Bandwidth Utilization Control : measured input is the bandwidth utilization, control input is the source rate. The same control method. • The switching point: is a bandwidth utilization control value near 1.
Idea 2 • Synchronous Adjustment to Arrival and Service Rate for QoS Control • Traditional way for queuing delay control is by Feedforward(prediction) + Feedback (PI or Adaptive Control for compensate residual error) • Measured input: queuing delay – desired delay • Measured output: variation to arrival rate by AC or variation to service rate by PR
Idea 2 (continue) • How to combine the two? • Necessity: • only AC may result in the low QoS since sometimes high drop probability will result most incoming requests are denied • Only PR may result in overload or extremely high demand (a large pool of processes should be created to serve the queued requests) on the server.
Idea 2 (continue) • My idea: • Take ae= arrival rate(ar) – service rate(sr) as a whole, a combined factor as control input. • use Feedforward + Feedback(adaptive) to provide the optimal variation to “ae” for desired delay. • Then map it to the desired Arrival rate (the drop probability) and the service rate (the processes ratio) simultaneously.
Idea 2 (continue) • If ar= mean(armin, armax) && sr= mean(srmin, srmax) • Then (dealta ae(k)+ ae(k-1))/ae(k-1) = x*(pa*ar(k-1)-ps*sr(k-1)). • pa_new = x*pa ps_new = x*ps; • Elseif ar> mean(armin, armax) && sr<=mean(srmin, srmax) • (dealta ae(k)+ ae(k-1))/ae(k-1) = pa*ar(k-1)- x *ps*sr(k-1) • pa_new = pa; ps_new = (pa*ar(k-1)- (dealta ae(k)+ ae(k-1))/ae(k-1))/x; • Elseif ar<= mean(armin, armax) && sr>=mean(srmin, srmax) • (dealta ae(k)+ ae(k-1))/ae(k-1) = x *pa*ar(k-1)- ps*sr(k-1) • ps_new = ps; pa_new = ( (dealta ae(k)+ ae(k-1))/ae(k-1)+ ps*sr(k-1))/ ar(k-1); End if