WAM-The Weighted Average Method or Predicting The Performance of Systems With Bursts of Customer Sessions

WAMThe Weighted Average Method or Predicting the Performance of Systems with Bursts of Customer Sessions
By C.M.Srinidhi
Predictive performance models are important tools that support system sizing, capacity planning, and systems management exercises. Weighted Average Method (WAM) improves the accuracy of predictive performance models for systems with bursts of concurrent customers. WAM considers the customer population distribution at a system to reflect the impact of bursts. We demonstrate the effectiveness of WAM using a case study involving a multitier TPC-W benchmark system. To demonstrate the utility of WAM with multiple performance modeling approaches, we developed both Queuing Network Models and Layered Queuing Models for the system. Results indicate that WAM improves prediction accuracy for bursty workloads for QNMs and LQMs by 10 and 12 percent, respectively.
Introduction
The purpose of this work is to improve the accuracy of predictive performance modeling techniques so that they may be more reliably applied in system sizing, capacity planning, and systems management exercises. The systems we consider are session-based systems, e.g. e-commerce and enterprise application systems. A session is defined as a sequence of related customer requests that accomplish some business purpose. For example, in an ecommerce system, requests may be for a Home page or a Buy transaction. We argue that for systems with bursts in the number of concurrent sessions, knowing the mean or maximum number of concurrent customer sessions is not generally sufficient to enable accurate performance predictions. It is necessary to consider the distribution of concurrent customer sessions. We refer to this distribution as the session population distribution session length.
Several studies have indicated that multitier sessionbased systems experience bursty arrivals of requests from customers which can adversely affect performance. Several Performance models such as QNM and LQM have been proposed to predict the impact of such burstiness. However, they are not efficient enough.
Techniques are not adequate for capturing the impact of complex underlying workload parameter distributions that can contribute to burstiness in population distribution. Hence we use Weighted Average Method (WAM) to estimate population distribution.
WAM
WAM applies the well-known concept of hybrid modeling to session-based systems. Hybrid modeling typically combines simulation with analytic modeling techniques to study complex systems. It exploits a fast Monte Carlo simulation to estimate the distribution for any given workload. The estimated distribution is combined with a performance model (e.g., a QNM) to predict a systems performance under that workload. The primary advantage of the newly proposed method is that it is more robust.
A trace file S can be based on a historical session log from a real system, or it can be synthetically generated using a tool such as SWAT. Each session in the trace has an arrival time. Each request has a session identifier, a start time and end time such that (end time start time) is the response time of the request, and a flag that indicates whether a request is the last request for its session. For all but the last request in a session, we define a requests think time as the time between its end time and the start time of the next request in the session. The first request of a session has a start time that is equal to its sessions arrival time. The sequence R of response time estimates is obtained from a performance model for the system, e.g., a QNM or an LQM. The sequence X of throughput estimates are obtained from the trace S using the method described shortly. The population distribution estimator operates as follows: When using a historical trace file S, per request response times are known so the sequenceRof response time estimates is not needed to compute the population distribution.
observations
The maximum errors observed for are nearly as high as 43 and 32 percent for the QNM and the LQM, respectively. Furthermore, for bursty workloads, using the QNM and LQM models in combination with a In contrast, the WAM approach significantly improves the accuracy of mean response time predictions. For bursty workloads, average prediction errors dropped by 12 and 10 percent for LQMs and QNMs, respectively, as compared to the Markov birth-death approach. Moreover, the LQM-based WAM approach had much lower average error and range of errors than the QNM-based WAM approach. WAM permits well-known modeling methodologies such as QNMs and LQMs to be applied for studying systems characterized by bursts in concurrent sessions.
Thank you

WAM-The Weighted Average Method or Predicting The Performance of Systems With Bursts of Customer Sessions

Uploaded by

Copyright:

Available Formats

WAM-The Weighted Average Method or Predicting The Performance of Systems With Bursts of Customer Sessions

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

WAM-The Weighted Average Method or Predicting The Performance of Systems With Bursts of Customer Sessions

Uploaded by

Copyright:

Available Formats

WAMThe Weighted Average Method or Predicting the Performance of Systems with Bursts of Customer Sessions

You might also like