Presentation is loading. Please wait.

Presentation is loading. Please wait.

1 The Challenges of Reflexive Control Systems Lui Sha

Similar presentations


Presentation on theme: "1 The Challenges of Reflexive Control Systems Lui Sha"— Presentation transcript:

1 1 The Challenges of Reflexive Control Systems Lui Sha lrs@uiuc.edu

2 2 Acknowledgement Many have contributed to the collaboration. In particular, I want to thank Xue Liu, Jin Jeo at UIUC, Tarek Abdelzaher at UVA and Joe Hellerstein at IBM Research

3 3 Web Service Performance Control Web Servers Application ServersEnd Users KeepAlive TImeout Number of Threads MaxClients DB Connections Fast response cache MaxRequestsPerChild ThreadsPerChild Max simultan. requests ListenBackLog URL Cache EJB threads JVM heap size Servlet reload int Courtesy of Joe Hellerstein, IBM Research Network based server systems, e.g., Web servers, have now become an integral part of our society.

4 4 Overview In a queueing network, service performance control can be viewed as a hierarchical control Setting the performance goal at each node Rout the traffic to the nodes Performance regulation on each node There are many interesting problems. Among them “Noise” is data Event driven control Reflexivity in control

5 5 Some Interesting Facts Suppose we want to correct a biased coin with Prob(head)= 0.4. We begin by soldering a small weight to the side of head, and then do some experiments and adjust the weight. From a control perspective, the transfer function between the change of weight and the change of probability of head manifests itself clearly only when the sample size becomes large. Fast control actions do not lead to fast convergence Low pass filter, except moving average, won’t work because there is no noise. Event (sample size) based control action yields better results. Fixed sample size control typically has random time intervals

6 6 Reflexive Control Normal control system: the plant model is invariant to control actions Reflexive control system: the control alters the plant Genetics -> (physique, intelligence) -> mating -> genetics Laws -> control social behaviors -> changes laws Queueing system control is reflexive

7 7 Reflexivity and Uncertainty Uncertainty principle in measurement. In quantum physics, the act of measurement distorts what we try to observe. “Uncertainty principle” in reflexive control System. The act of control alters the model used to design the controller

8 8 The Bright Spot One of the worst that can happen in a network of servers is performance failure due to congestions. This turns out to be an easier task Heavy traffic creates long queue Long queue is persistent Low of large number kicks in Allows for fluid approximation Linear model with PI control works well High performance regulation is reflexive. Why?

9 9 Delay vs Control for M/M/1

10 10 Keeping the model relevant  D ref “Matching” the measured arrival rate with computed service rate e.g., Let µ= + 0.5 will “lock” the system at the point of linearization Slope = d  /dD ref = - 4

11 11 Queuing Model Based Control Measured Delay d dd Server Queue Control Request Queueing Model qq   Ref Delay D ref Control of Average Delay

12 12 Experimental Results Sha, L., Li, X., Lu, Y., and; Abdelzaher, T. “Queueing model based network server performance control”, the proceedings of IEEE Real-Time Systems Symposium, 2002

13 13 Long Shadow of Reflexivity Alas, there are still considerable variances The queueing model is a function variances in arrival time and service time. Good admission control smooth the input variances Good execution time control smooth the execution time variances Control invalidates queueing model! Feed forward over-allocates when delay turns lower Feed forward under-allocates when delay turns higher

14 14 Jumps in Loads In addition to over/under allocation in feed forward errors, the jumps in arrival process creates another challenge The queue length from previous traffic load takes time to settle to the new equilibrium queue distribution. Leads to large transients Whatever the sources of errors, large variance in service time correlates with large variances in queue length Adding a control term on (Actual_queue_length – expected_queue_length) Reducing both transients induced by workload changes and variances during steady state

15 15 Comparison of two delay regulators using web trace Queueing Model Based Feedback Control Queue Length Model Based Feedback Control Steady state feed forward + delay feedback + model based queue control Actuator parameters: K: delay control;  : multiplicative noise control

16 16 Summary – theoretical problems Heavy traffic not only cures “reflexivity”, it also allows for fixed time interval control works as well as sample size based control upside: simple fixed rate linear model works well downside: the fun is gone Under moderate traffic control, we need to deal with event (sample size) driven control reflexivity queueing model based feed forward helps but the feed forward both over allocate & under allocate Optimized supervisory control for networked server set point adjustments

17 17 Summary: Engineering Problems Poor observability in actual systems: Many queues and I/O states are not visible and not controllable. Poor actuators: Rejecting individual requests  TCP window adjustment (non-linear, coarse grain actuator in admission control) Hard to do fine grain control in CPU cycle allocation Nearly no control of I/O bandwidth Control OS


Download ppt "1 The Challenges of Reflexive Control Systems Lui Sha"

Similar presentations


Ads by Google