1 / 20

A Resource Allocation Mechanism of Data Center for Public Cloud Service

A Resource Allocation Mechanism of Data Center for Public Cloud Service . 指導教授:王國禎 學生:連懷恩 國立交通大學網路工程研究所 行動計算與寬頻網路實驗室. Outline. Introduction Scenario and settings A branch and bound approach Discussion and ongoing works. Introduction.

daria
Download Presentation

A Resource Allocation Mechanism of Data Center for Public Cloud Service

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A Resource Allocation Mechanism of Data Center for Public Cloud Service 指導教授:王國禎 學生:連懷恩 國立交通大學網路工程研究所 行動計算與寬頻網路實驗室

  2. Outline • Introduction • Scenario and settings • A branch and bound approach • Discussion and ongoing works

  3. Introduction Dealing with Number of active servers, number of VMs for each application, VM placement problem, and optimization over a series of time slots. Apply an existing load-prediction method to get the forecast of resource demand of each application over a series of time slots.

  4. Scenario and Settings All servers and VMs have fixed capacity. When the workload of an application changes, we readjust the number of VMs dedicated to that application.

  5. Scenario and Settings – The Power Consumption Model

  6. Scenario and Settings – The Cost Function Thus, for the operating cost, we define Pidleas the basic energy cost in a time slot when a server is idle, and α as the incremental energy cost in a time slot every time when we add in a new VM. For the switching cost, we define δs as the cost to switch on/off a server, and δv as the cost to switch on/off a VM.

  7. Scenario and Settings – The Cost Function The cost function will be the sum of operating cost and switching cost, considering all running servers and VMs. Suppose there are T time slots in our scenario, then the objective function will be the minimization of cost function over T time slots, subject to the number of hosted VMs must not exceed the server capacity.

  8. A Branch-and-Bound Approach – Problem Simplification • Claim 1: • The solution of optimum(# of servers, # of VMs) is a superset of the solution of optimum(# of servers, #number of VMs, VM placement). • Thus, we can apply a two-pass algorithm to reduce the complexity of the problem.

  9. A Branch-and-Bound Approach – Problem Simplification • Claim 2: • When doing pass1 of the algorithm, we only need to consider the concave upwards parts.

  10. A Branch-and-Bound Approach Use the branch and bound approach to find the optimal solution of (# of servers, # of VMs). Given the status of source and sink, the resource demand of each time slot, find a minimal weighted path from source to sink.

  11. A Branch-and-Bound Approach – Problem Complexity A simple relation to decide when to keep the unnecessary VM is ∆α < 2δv , and for the server, it is ∆Pidle < 2δs , ∆ is the time interval by that we need the resource again. The problem comes when that VM is the last remaining VM on that server, or, last two VMs from two apps, and so on. What is more, the result of deciding # of VMs will affect the way deciding # of servers. Its hard to find a simple way to find the optimal solution.

  12. A Branch-and-Bound Approach – Main Idea The key idea of the BB algorithm is: if the lower bound for some tree node A is greater than the upper bound for some other node B, then A may be safely discarded from the search. A global upper bound: the minimum cost among all accomplished paths so far. The lower bound of each node: current accumulated cost + minimum remaining operating cost + switching cost to the status of sink node

  13. A Branch-and-Bound Approach – A Paradigm Set α= 1.0, Pidle = 0.5, δv = 0.6, δs = 0.7

  14. A Branch-and-Bound Approach – Tree Structure Suppose there are n applications, define the resource provision tuple as a (n+1)-tuple. The 1st element represents the # of active servers, the 2nd element represents the # of VMs for the first app, and i-th element represents the # of VMs for the (i-1)-th app. A search tree with depth T. Every node consists of a resource provision tuple, the current accumulated cost field and lower bound field. Using doubly linked list. A path can be represented by its tail node and backtracked to its preceding nodes.

  15. A Branch-and-Bound Approach – Branch Table When it happens a resource demand degradation, it may be worthy of keeping the VM if ∆α < 2δv . The false-positive is allowed. Construct a branch table composed of T resource provision tuple. Each element in the tuple indicates how many possible branches of # of VMs for that app.

  16. A Branch-and-Bound Approach – Branch Table Similarly, for the server field in the tuple, consider ∆Pidle < 2δs . Note that the total server capacity must accommodate the maximal possible number of VMs in that time slot. From a certain node, the possible number of branches in the next time slot is the product of all elements of that tuple.

  17. A Branch-and-Bound Approach – Shortcut Table A shortcut is a discovered path from a certain time slot to the sink node with minimal cost. In the shortcut table, every time slot records adiscovered best path and its cost. Note that the shortcut of (t = a) is not necessarily contained in shortcut of (t = b), for any a < b .

  18. A Branch-and-Bound Approach – ShortcutTable • Updating shortcut table: every time when we finish a path to the sink node. The cost of a shortcut is reversely accumulated. • Applying shortcut: apply a shortcut when (current accumulated cost + shortcut cost + switching cost transferring to shortcut) < upper bound. • The reason of using shortcut : • To reduce the complexity of the algorithm, we must discard nodes as early as possible. It is possible to lower the upper bound only when we finish another path to the sink.

  19. A Branch-and-Bound Approach – The Whole Procedure The baseline path: the minimal operating path. Using a priority depth-first search.

  20. Discussion and Ongoing Works The worst case complexity is still extremely high. We don’t know the effectiveness of the those acceleration method. Need to prove or analyze those two claims. Need an approach to deal with the VM placement problem.

More Related