Dynamic scaling for multi-tiered distributed systems using payoff optimization of application classes

• US 9,386,086 B2
• Filed: 09/11/2013
• Issued: 07/05/2016
• Est. Priority Date: 09/11/2013
• Status: Active Grant

• Current Assignee: Cisco Technology, Inc. (Cisco Systems, Inc.)
• Original Assignee: Cisco Technology, Inc. (Cisco Systems, Inc.)
• Inventors: Fenoglio, Enzo, Andrieux, Fabien, Surcouf, Andre
• Primary Examiner(s): Wai, Eric C
• Application Number:
  US14/023,686
• Publication Number:
  US 20150074679A1
• Time in Patent Office: 1,028 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G06F 2209/5019 Workload prediction
• CPC Class Codes: G06F 9/505 considering the load
• CPC Class Codes: H04L 67/1008 based on parameters of servers, e.g. available memory or workload monitoring of computer activity G06F11/30

Abstract

In one embodiment, a method is described. The method includes: monitoring workloads of a plurality of application classes, each of the application classes describing services provided by one or more applications in a multi-tiered system and comprising a plurality of instantiated execution resources; estimating, for each of the application classes, a number of execution resources able to handle the monitored workloads, to simultaneously maintain a multi-tiered system response time below a determined value and minimize a cost per execution resource; and dynamically adjusting the plurality of instantiated execution resources for each of the application classes based on the estimated number of execution resources.

Claims

1. A method comprising:

monitoring workloads of a plurality of application classes, each of said application classes describing services provided by one or more applications in a multi-tiered system, and each of said application classes comprising a plurality of instantiated execution resources;

estimating, for each of said application classes, a number of execution resources able to handle said monitored workloads, to simultaneously maintain a multi-tiered system response time below a determined value and minimize a cost per execution resource, wherein said multi-tiered system response time is maintained below the determined value and said cost per execution resource is minimized by optimizing a payoff function defined for each of said application classes, wherein said payoff function defined for each of said application classes depends on;

a workload arriving at an application class;

a local response time of said application class; and

a unitary cost for instantiating an execution resource for said application class; and

for each one application class of said plurality of application classes;

dynamically adjusting said plurality of instantiated execution resources comprised in said one application class based on said estimated number of execution resources, said dynamically adjusting comprises increasing or decreasing the number of execution resources for an application class for said monitored workload for said application class.

2. The method of claim 1, wherein said estimating comprises a minimal over-provisioning to account for unpredictable burst workload for each of said application classes.

3. The method of claim 1, wherein said local response time of said application class is estimated as a ratio of a number of requests waiting to be served by a plurality of execution resources instantiated for said application class to said workload of said application class.

4. The method of claim 1, wherein said multi-tiered system response time is computed as a cumulative sum of local response times of all application classes belonging to said multi-tiered system.

5. The method of claim 1, wherein said determined value for a multi-tiered system response time is specified by an operator in a service level agreement contract.

6. The method of claim 1, wherein said estimating further comprises:

providing first estimates for each of said application classes of a number of execution resources able to handle said monitored workload; and

providing second estimates for each of said application classes by applying weighting factors to said first estimates.

7. The method of claim 6, wherein said second estimates are provided using a double exponentially-weighted moving average filter.

8. The method of claim 1, wherein said estimating further comprises:

associating a set of parameters to one or more application classes of said multi-tiered system; and

correcting, for one or more application classes, said estimated number of execution resources able to handle said monitored workloads according to said set of parameters.

9. The method of claim 8, wherein said set of parameters characterizes an application class and comprises at least one of the following:

a maximum number of requests that said application class is able to handle per second;

a unitary cost for instantiating an execution resource;

a default minimum number of execution resources to be instantiated; and

at least one time range associated with a minimum number of execution resources to be instantiated.

10. The method of claim 9, wherein said correcting comprises using said default minimum number of execution resources to be instantiated if said estimating is lower than said default minimum number of execution resources to be instantiated.

11. The method of claim 9, wherein said correcting comprises using said minimum number of execution resources to be instantiated if said estimating is lower than said minimum number of execution resources to be instantiated for said at least one time range.

12. The method of claim 1, wherein said dynamically adjusting is performed, for each of said application classes, only when a change is detected between two consecutive estimated number of execution resources.

13. An auto-scaler device comprising a processor, said processor being operating to:

monitor workloads of a plurality of application classes, each of said application classes describing services provided by one or more applications in a multi-tiered system and each of said application classes comprising a plurality of instantiated execution resources instantiated;

estimate, for each of said application classes, a number of execution resources able to handle said monitored workloads, to simultaneously maintain a multi-tiered system response time below a determined value and minimize a cost per execution resource, wherein said multi-tiered system response time is maintained below the determined value and said cost per execution resource is minimized by optimizing a payoff function defined for each of said application classes, wherein said payoff function defined for each of said application classes depends on;

a workload arriving at an application class;

a local response time of said application class; and

a unitary cost for instantiating an execution resource for said application class; and

for each one application class of said plurality of application classes, to dynamically adjust said plurality of instantiated execution resources comprised in said one application class based on said estimated number of execution resources by increasing or decreasing the number of instantiated execution resources for said one application class.

14. One or more computer readable non-transitory tangible storage media encoded with software comprising computer executable instructions being operative, when the software is executed by a processor, to:

monitor workloads of a plurality of application classes, each of said application classes describing services provided by one or more applications in a multi-tiered system, and each of said application classes comprising a plurality of instantiated execution resources;

estimate, for each of said application classes, a number of execution resources able to handle said monitored workloads, to simultaneously maintain a multi-tiered system response time below a determined value and minimize a cost per execution resource, wherein said multi-tiered system response time is maintained below the determined value and said cost per execution resource is minimized by optimizing a payoff function defined for each of said application classes, wherein said payoff function defined for each of said application classes depends on;

a workload arriving at an application class;

a local response time of said application class; and

a unitary cost for instantiating an execution resource for said application class; and

for each one application class of said plurality of application classes;

dynamically adjust said plurality of instantiated execution resources comprised in said one application based on said estimated number of execution resources, said dynamically adjusting comprising adjusting said plurality of instantiated execution resources by increasing or decreasing the number of instantiated execution resources for said one application class.