Online Feedback-based Estimation of Dynamic Page Service Time Ashwini Kumar Kaushik Veeraraghavan Benjamin Wester Kang Shin

Embed Size (px)

Citation preview

  • Slide 1

Online Feedback-based Estimation of Dynamic Page Service Time Ashwini Kumar Kaushik Veeraraghavan Benjamin Wester Kang Shin Slide 2 2 Motivation Dynamic pages Increasingly prevalent Pages difficult to cache Service time applications Differentiated QoS (prioritizing requests) Load balancing Virtual host services Estimates: credit card payment times Slide 3 3 Idea History Maintain records of past service times Predict based on existing records Assume a correlation between past and future What is the invariant property? What system state can we store to track it? We explore three estimators and their design Slide 4 4 Outline Motivation Design & Implementation Evaluation Conclusion and Future Work Slide 5 5 History 1.Maintain per-URL fixed-size history table 2.Data depends on the estimator used 3.Tag incoming requests 4.Update history when request processing finishes Slide 6 6 SEDA: Haboob Slide 7 7 Sirocco Slide 8 8 Estimators SchemeStore in historyEstimation Average STQ STQT Slide 9 9 Outline Motivation Design & Implementation Evaluation Conclusion and Future Work Slide 10 10 Evaluation How close are our estimators? Load patterns Steady load: httperf 50ms / request Changing load: SURGE Slide 11 11 Average Estimator (steady) Request takes ~35ms to complete Estimator should not be too responsive Slide 12 12 Average Estimator (changing) Each point averages 20 requests Vertical bars indicate the addition of a handling thread Slide 13 13 Average Estimator How to read the graph: Enters system:t = 5.4s Estimated service time:1.6s Real service time:4.9s Slide 14 14 Average Estimator History becomes out-of-date Base estimation on the current state of the system Slide 15 15 STQ Estimator Its better than the Average Estimator Inaccurate around a thread change Slide 16 16 STQ Estimator Estimations will be inaccurate unless the estimator knows about the coming change in the number of threads After a change, recently-completed requests show a mixed view of system state and will pollute the history Slide 17 17 STQT Estimator Overshoots target because of a polluted history Fix: Integrate request times across the thread change Slide 18 18 Outline Motivation Design & Implementation Evaluation Conclusion and Future Work Slide 19 19 Concluding Remarks Takeaways Dont adapt to transient delays Use current state along with the history Be aware of underlying changes Future Directions Extrapolate requests service times URL service time as a distribution Slide 20 Questions?