Server consolidation for internet applications in virtualized data centers

Bo Wang, Ying Song, Yuzhong Sun, Jun Liu

TL;DR

This paper introduces a novel approach to server consolidation in data centers specifically designed for "elastic" Internet applications. By modeling the problem as a three-tiered mapping (Application-to-VM-to-PM) and using a new heuristic called 3MAX, the researchers can minimize the number of physical servers required while guaranteeing application performance, even as workloads fluctuate.

What Problem Does This Paper Address?

In modern data centers, server consolidation is the go-to strategy for energy efficiency: you pack as many Virtual Machines (VMs) onto as few Physical Machines (PMs) as possible so you can power down the idle hardware.

However, most existing consolidation techniques treat the number of VMs as a fixed constant. This doesn't work well for "elastic" Internet applications (like e-commerce or web services) where the number of instances needs to scale up or down based on user demand. If you have too many VMs, you waste power; too few, and performance tanks. Furthermore, most models assume that a VM with a specific configuration performs identically on any physical server, ignoring the reality of hardware heterogeneity in the cloud.

Methods

The authors tackle this by formulating the consolidation as an Integer Linear Programming (ILP) problem. Unlike traditional "bin-packing" approaches that just look at VM-to-PM mapping, this model handles a three-tier mapping: Application $\rightarrow$ VM $\rightarrow$ PM.

Since ILP is NP-hard (meaning it takes too long to solve for large data centers), they propose the 3MAX heuristic algorithm:

  1. Step 1: It selects a PM and calculates the ratio between provided performance and the proportion of resources used for different VM types.
  2. Step 2: It selects the application with the highest "performance-to-resource" ratio and allocates a VM to it on that PM.
  3. Step 3: It repeats this process until the PM is full or all application requirements are met.

This greedy approach allows the system to make consolidation decisions in polynomial time, making it practical for real-world use.

Datasets

The researchers evaluated their approach using a mix of industry-standard benchmarks to simulate diverse workloads:

  • TPC-W: A transactional web e-commerce benchmark.
  • YCSB (Yahoo! Cloud Serving Benchmark): Used to simulate cloud-based data services.
  • Apache Benchmark (ab): Used to simulate web server requests for different file sizes (1KB and 1MB).
  • SysBench: Used to test CPU-intensive prime number calculations.
  • Workload Traces: They used real-world traffic data from the 1998 World Cup Web site to generate realistic, fluctuating request patterns over a 24-hour period.

Key Insights and Ideas

  • Joint Scaling: The paper emphasizes that horizontal scaling (adding/removing VMs) and vertical scaling (resizing VMs) must be considered simultaneously with placement on physical hardware to achieve true energy efficiency.
  • Hardware Matters: They explicitly account for the fact that the same VM might perform differently on an Intel Xeon vs. an AMD Opteron. Their model includes a performance variable ($\mu_{i,k,l}$) that is specific to the application, the PM type, and the VM type.
  • The "3MAX" Logic: The name comes from the algorithm's goal to maximize the ratio of performance provided to resources consumed. By picking the "best" application-VM-PM match first, it leaves more room for subsequent applications.

Key Results

  • Near-Optimal Efficiency: The 3MAX heuristic consumes less than 4.3% more resources than the theoretically optimal solution found by solving the ILP.
  • Better than Industry Standards: Compared to common packing algorithms like First Fit Decreasing (FFD) and Least Loaded (LL), 3MAX reduced the number of required physical machines by about 1.06%.
  • High Scalability: In simulations with 3,000 PMs and 500 applications, the algorithm made decisions in just 0.22 seconds, proving it is fast enough for dynamic cloud environments.
  • Performance Guarantees: The "Relative Difference" (RDi) between required and provided performance stayed very close to zero, meaning applications got exactly the resources they needed without significant waste.

Limitations and Future Work

  • Resource Scope: The study primarily focused on CPU, Memory, and Network Interface (NIC) resources. It did not consider disk I/O, as the authors noted it was not a bottleneck in their specific experimental setup.
  • Core Pinning: The results assume VCPUs are pinned to physical cores to avoid virtualization overhead, which might not be the default configuration in every public cloud.
  • Migration Costs: While the paper mentions that live migration has overhead, the 3MAX algorithm focuses on the target mapping. Future work could more deeply integrate the specific "path" or cost of moving from the current state to the 3MAX-optimized state.

Summary updated

Abstract

Server consolidation based on virtualization technology simplifies system administration and improves energy efficiency by improving resource utilizations and reducing the physical machine (PM) number in contemporary service-oriented data centers. The elasticity of Internet applications changes the consolidation technologies from addressing virtual machines (VMs) to PMs mapping schemes which must know the VMs statuses, i.e. the number of VMs and the profiling data of each VM, into providing the application-to-VM-to-PM mapping. In this paper, we study on the consolidation of multiple Internet applications, minimizing the number of PMs with required performance. We first model the consolidation providing the application-to-VM-to-PM mapping to minimize the number of PMs as an integer linear programming problem, and then present a heuristic algorithm to solve the problem in polynomial time. Extensive experimental results show that our heuristic algorithm consumes less than 4.3% more resources than the optimal amounts with few overheads. Existing consolidation technologies using the input of the VM statuses output by our heuristic algorithm consume 1.06% more PMs.
BibTeX
@inproceedings{wang2016server,
  author = {Bo Wang and Ying Song and Yuzhong Sun and Jun Liu},
  title = {{Server consolidation for internet applications in virtualized data centers}},
  booktitle = {SpringSim (HPS)},
  year = {2016},
  url = {https://openreview.net/forum?id=u7L9WI-rci5},
}

Loading related content...

Figure