Resource allocation for autonomic data centers using analytic performance models.

Transcription

1 Bennani, Mohamed N., and Daniel A. Menasce. "Resource allocation for autonomic data centers using analytic performance models." Autonomic Computing, ICAC Proceedings. Second International Conference on. IEEE, Summarized by: Cristopher Flagg

2 Abstract Large data centers host several application environments (AEs) that are subject to workloads whose intensity varies widely and unpredictably. Therefore, the servers of the data center may need to be dynamically redeployed among the various AEs in order to optimize some global utility function. Previous approaches to solving this problem suffer from scalability limitations and cannot easily address the fact that there may be multiple classes of workloads executing on the same AE. This paper presents a solution that addresses these limitations. This solution is based on the use of analytic queuing network models combined with combinatorial search techniques. The paper demonstrates the effectiveness of the approach through simulation experiments. Both online and batch workloads are considered.

3 Previous Work - Limitations Previous work uses a table-driven approach that stores response time values obtained from experiments for different values of the workload intensity and different number of servers. Interpolation is used to obtain values not recorded in the table. Some limitations: i) it is not scalable with respect to the number of transaction classes in an application environment, ii) it is not scalable with respect to the number of AEs, and iii) it does not scale well with the number of resources and resource types. Building a table from experimental data is time consuming and has to be repeated if resources are replaced/updated within the data center

4 Proposed Solution This paper proposes an alternative solution to the data center resource allocation problem along the lines of our previous work. We replace the table-driven approach with predictive multiclass queuing network models. We show that the proposed solution does not suffer from the limitations mentioned above, primarily scalability and need for offline pre-processing.

5 Problem Definition M application environments (AEs). Each AE has: N servers (variably assigned). Si classes of transactions. Reports stats to global utility. Online (poisson) workloads OR Batch (concurrent jobs) workload. Servers dynamically moved among AEs Primarily a problem of moving servers between AEs to meet SLA.

6 Problem Definition - Utility Global Utility Ug: Based on response time per class per AE. Online Workloads: Utility drops sharply as it approaches SLA. Batch Workloads: More throughput provides better results. Other workloads possible, not explored. Online Workload Batch Workload

7 Problem Definition - Local Controller The workload monitor collects transaction arrival rates per transaction class. Workload database. Workload forecaster uses database to make future workload intensity predictions. Predictive model solver makes performance predictions: For current workload level. For a forecast workload level. Utility function evaluator computes the utility function for the AE.

8 Problem Definition - Global Controller Updates at set interval or in response to global utility function changes. Possible combinations of servers to SA determined using Beam Search. Ug for 'neighbors' calculated. Highest neighbors kept. Neighbors of kept neighbors calculated. Process repeats set number of times. Batch neighbors determined based on number of servers to meet concurrency requirements. Online Servers determine neighbors based on not overflowing resource usage.

9 Problem Definition - Online Transaction Performance Model E-commerce, Database transactions Service demand at a device i for class s. transactions can be measured using the Service Demand Law: The service demand at device i is the ratio between the utilization of device i due to class s and the throughput of class s. Response time Ri,r is a function of the number of servers ni allocated to AEi. Percentage of Arrivals: Class 1-30% Class 2-25% Class 3 - remaining 45%.

10 Problem Definition - Batch Transaction Performance Model Long report generation for decision support systems or data mining applications over large databases. Throughput is often of higher concern than response time. Workload intensity is measured by the concurrency level, i.e., the number of concurrent jobs in execution in each class

11 Experimental Setting Data center with three AEs: 2 Online AEs - Each with separate workload generator. 1 Batch AE - Threads allocated based on concurrency level. 25 Servers divided between AEs. Data center simulated on one machine, controller on another. Assume switching cost is Zero - all applications are installed in all servers, no server spin up time for specific applications. Staggered demand between online and batch to show controller can move machines to maximize global utility.

12 Results - Utility v. Workload

13 Results - AE1

14 Results - AE2

15 Conclusion We showed how analytic performance models can be used in an efficient manner to design controllers that dynamically switch servers from one application environment to another as needed. The approach scales very well with the number of AEs, resources within an AE, and transaction classes in comparison with simulation models. We are investigating the impact on the global utility function if switching costs are non-negligible. The utility function needs to reflect the unavailability of a server while it is moving from one application environment to another We are also investigating the use of adaptive controllers and workload forecasting techniques as we did in previous work.

16 Conclusion - Observations The batch AE was included primarily as something from which the online AEs could steal resources to improve performance. In one experiment, batch AE was given additional resources to show how online AEs adapted. Published in 2004, cited by 467 papers.

17 Conclusion Questions?