PaperlayerView exact revision on arXiv

Source-linked AI summary

GreenDCN: a General Framework for Achieving Energy Efficiency in Data Center Networks

Lin Wang, Fa Zhang, Jordi Arjona Aroca, Athanasios V. Vasilakos, Kai Zheng, Chenying Hou, Dan Li, Zhiyong Liu

arXiv:1304.3519v2cs.NI

TL;DR

Data-center network energy is significant, while existing traffic-engineering approaches insufficiently exploit application and data-center structure. The paper proposes a coordinated VM-assignment and routing framework, proves the resulting problem NP-hard, and reports up to 30% savings in the supplied experiment.

  • Problem

    Network energy is a substantial data-center issue, while existing approaches primarily rely on traffic engineering and may be inaccurate when application traffic patterns vary.

  • Method

    The paper combines application and network information, assigns VMs to create favorable traffic patterns, and then uses energy-efficient routing to reduce active switches and balance flows.

  • Results

    30% savings: Greedy-EER achieves up to 30% energy savings compared with Greedy-SP in the reported Fat-Tree experiments.

  • Takeaways & Limitations

    The framework uses distinctive data-center features to coordinate VM placement and routing for network energy conservation.

Abstract

from arXiv · show

The popularization of cloud computing has raised concerns over the energy consumption that takes place in data centers. In addition to the energy consumed by servers, the energy consumed by large numbers of network devices emerges as a significant problem. Existing work on energy-efficient data center networking primarily focuses on traffic engineering, which is usually adapted from traditional networks. We propose a new framework to embrace the new opportunities brought by combining some special features of data centers with traffic engineering. Based on this framework, we characterize the problem of achieving energy efficiency with a time-aware model, and we prove its NP-hardness with a solution that has two steps. First, we solve the problem of assigning virtual machines (VM) to servers to reduce the amount of traffic and to generate favorable conditions for traffic engineering. The solution reached for this problem is based on three essential principles that we propose. Second, we reduce the number of active switches and balance traffic flows, depending on the relation between power consumption and routing, to achieve energy conservation. Experimental results confirm that, by using this framework, we can achieve up to 50 percent energy savings. We also provide a comprehensive discussion on the scalability and practicability of the framework.

I. INTRODUCTION

The paper proposes a data-center-specific framework that combines application, VM-placement, topology, and traffic information to improve network energy efficiency. It addresses limitations of traffic-engineering approaches by modeling VM assignment and routing as coordinated optimization stages.

  • Motivation: Network energy is a substantial data-center concern, reaching approximately 50% of total power when server utilization falls to 15%.At 100% server utilization, network power is approximately 20% of total power.
  • Limitations of Existing Work: Existing energy-efficient DCN studies mainly consolidate traffic onto fewer links and switch off network elements, but traffic prediction can be inaccurate because application patterns vary.The paper argues that application-dependent traffic makes traditional prediction-based approaches insufficiently precise.
  • Proposed Framework: The framework exploits regular DCN topologies and controllable VM endpoints while coherently using application and network information.These features support coordinated decisions about VM placement and network traffic.
  • Proposed Framework: VM assignment is designed to create favorable traffic conditions before traffic engineering is applied, followed by routing solutions tailored to traffic patterns and network features.This separates the optimization into a placement stage and a traffic-engineering stage.
  • Contributions: The paper models the problem, analyzes its complexity, develops algorithms for VM assignment and routing, and evaluates their efficiency experimentally.It also discusses scalability, practicability, and future challenges including profiling accuracy and topology- or traffic-specific engineering.

III. MODELING THE ENERGY-SAVING PROBLEM

The model represents DCN energy as load-dependent switch power over time, using application communication periods and traffic matrices to describe changing network demand. It focuses on varying communication-intensive periods and assumes a uniform switch cost function for tractability.

  • Data Center Network Model: Switch traffic load is computed from the loads on incident links, with a factor of 1/2 preventing double counting of flows.The model denotes link load by y_e and node load by x_v.
  • Switch Energy Model: The framework integrates speed scaling and power-down strategies through a switch energy curve f_v(x_v).An unloaded switch can be shut down, while an active switch pays a fixed cost and an increasing load-dependent cost.
  • Application Model: The model uses varying-height and varying-width communication peaks, allowing multiple intensive periods with different durations and traffic volumes.This is selected as the most general of the three communication-pattern categories discussed.
  • Application Model: Jobs are represented across timeslots with relatively stable traffic, and each job has a time-dependent traffic matrix for its associated VMs.Communication-intensive intervals are represented as transfers with start and end times and a traffic matrix.
  • Application Model: Outside defined transfers, the model treats traffic as background traffic approaching zero with only a small influence on the network.This simplifies the representation of communication activity between intensive periods.

C. Problem Description

The problem description decomposes energy optimization into VM assignment and traffic engineering. Given VM placement, routing minimizes the total switch cost for time-varying demands subject to flow and capacity constraints.

  • VM Assignment: VM assignment minimizes total energy by placing VMs on servers while allowing suitable routing paths between VM pairs.The objective is defined over switch loads across the modeled timeslots.
  • Problem Decomposition: The optimization procedure has two related stages: VM assignment followed by energy-efficient traffic engineering.An assignment determines the demands that the routing stage must handle.
  • VM Assignment: The assignment formulation uses binary variables indicating whether each VM is assigned to a server, resource requirements, server capacities, and one-server-per-VM constraints.Traffic demands are obtained from the active transfers after an assignment is selected.
  • Traffic Engineering: Energy-efficient routing minimizes the sum of node costs over all switches while routing every demand in the time-slot demand set.The node cost depends on each switch's total traffic load.
  • Traffic Engineering: Flow conservation and switch-capacity constraints ensure feasible routing, while link loads aggregate routed demand and node loads aggregate incident-link traffic.The model uses binary routing indicators and limits each switch load to capacity C.

D. Complexity Analysis

The paper shows that optimal energy-saving in data center networks is NP-hard and develops VM-assignment principles to exploit network structure and switch power behavior.

  • The complexity proof reduces the problem from the strongly NP-hard Quadratic Assignment Problem, for which constant-factor approximation is also NP-hard.
  • Optimal energy-saving in data center networks is NP-hard.
  • The paper seeks VM assignments that exploit structured data-center topologies and application characteristics to improve energy efficiency.
  • Power rate is defined as the power consumed per unit of switch load, f(x)/x for x > 0.
  • Network power is minimized when the number of active switches is optimal and their loads are evenly balanced near R∗.
  • Practical switch profiles indicate that idle power consumption can range from 76 to 150 watts and may increase by approximately 40 watts or more at full speed.

1) Minimizing energy at the rack level:

The rack-level strategy assigns VMs according to traffic and resource structure, applying three principles to reduce network energy: compact high-traffic VMs, distribute VMs across racks when beneficial, and keep each job’s VMs in one pod when feasible.

  • Principle 1: The optimal rack assignment compacts VMs into racks as tightly as possible to minimize ToR-switch power consumption.
  • Principle 2: Distributing a job’s VMs across multiple racks can reduce power because traffic moves to the richly connected upper-layer network.For α = 2, distributing VMs across k = 16 or more racks reduces energy relative to compacting them into one rack.
  • Principle 3: An optimal assignment keeps VMs from the same job in one pod whenever feasible.The framework also clusters jobs with different traffic patterns to improve network-resource utilization and reduce interference.
  • Algorithm: The optEEA algorithm applies the three principles through super-VM formation, traffic-aware pod clustering, and minimum k-cut rack assignment.Super-VMs are formed by compacting highly communicating VMs, jobs are clustered into pods, and each job’s super-VMs are partitioned across racks.
  • VM transformation: Compacting VMs with high communication traffic reduces network power consumption.

V. ENERGY-EFFICIENT ROUTING

The routing stage relates energy conservation to traffic engineering and designs a two-phase energy-efficient routing algorithm for data center networks.

  • The routing stage first studies the relation between energy consumption and routing, then designs a two-phase energy-efficient routing algorithm.

A. Exploring Energy-Saving Properties

Energy-efficient routing must determine how many switches to activate and how to distribute traffic across them. The framework minimizes active switches and balances their loads to reduce energy consumption.

  • The routing problem asks how many switches are sufficient and how traffic flows should be distributed.
  • With the optimal number of switches fixed, energy savings are achieved by balancing traffic across all used switches.The paper attributes this to the convex relationship between power consumption and traffic load and identifies ECMP and VLB as applicable multipath protocols.
  • The routing procedure estimates active aggregation and core switches, distributes flows evenly over the selected network, and puts unused switches into sleep mode.
  • The optimal energy-efficient routing algorithm uses as few aggregation switches as possible.
  • In an optimal energy-saving solution, the number of active switches is minimized.

B. Two-Phase Energy-Efficient Routing

The energy-efficient routing algorithm selects enough switches to carry all flows, then uses multipath routing to balance traffic. Experiments compare this approach with greedy, shortest-path, and alternative VM-assignment combinations across two Fat-Tree sizes.

  • Routing algorithm: EER estimates active switches from total traffic and switch capacity, then applies first-fit decreasing to ensure selected switches can carry all flows.The bin-packing approximation treats flows as objects and switch transmission capacity as the bin size.
  • Routing algorithm: MPTCP routes flows over multiple subflows, helping move traffic away from hotspots created by uneven randomized path selection.The second routing phase uses congestion control across subflows to improve traffic distribution.
  • Energy savings: Up to 30% energy savings arise from EER alone relative to Greedy-SP, while combining OptEEA with EER reduces network energy consumption by up to 50%.EER reduces active switches and balances their load; VM assignment and routing reduce power through different mechanisms.
  • Experimental setup: The running-time evaluation uses 320- and 720-switch networks and compares the proposed algorithms with greedy assignment and routing alternatives.The corresponding running-time results are presented in Figure 5.

D. Running Time

The proposed algorithms remain computationally practical in the tested settings and are reported to scale to much larger data centers. However, online deployment and traffic-pattern effects remain practical boundaries discussed as future work.

  • Measured running time: Its running time is only 50% longer than the greedy algorithm in the reported comparison.The comparison is based on the running-time results presented in Figure 5.
  • Scalability: On topologies with tens of thousands of servers, the proposed algorithm usually finishes within tens of seconds.The authors characterize this runtime as acceptable for production data centers.
  • Practical boundaries: The model and method are designed for offline cases, while a deliberate adaptation to dynamic job arrivals and departures remains future work.The paper suggests sequential assignment and pod placement as a possible online adaptation.
  • Practical boundaries: Energy-saving efficiency depends on application traffic patterns, with uneven within-job traffic offering more savings and comprehensive impact analysis left for future study.The discussion identifies MapReduce as advantageous and recommends co-locating jobs with different traffic patterns.

APPENDIX

The appendix establishes NP-hardness through a reduction from QAP and analyzes VM placement and traffic routing principles for reducing network energy consumption. It shows that compacting communicating VMs and minimizing active ToR switches can improve energy efficiency.

  • NP-hardness: A polynomial-time optimal solver for the DCN energy-saving problem would solve QAP, establishing NP-hardness.The reduction maps QAP locations to servers, facilities to VMs, and distance-dependent switch power functions to the DCN instance.
  • NP-hardness: The reduction preserves optimal VM assignments because the DCN and QAP objective values differ only by the constant n^2σ.
  • Traffic engineering: Convex load-dependent switch costs favor routing choices that minimize the power rate fv(xv)/xv on every active switch.The stated objective balances switch loads because convexity makes concentrated load more expensive.
  • VM assignment: For α > 1, compacting VMs into racks as tightly as possible and minimizing the number of ToR switches improves network energy efficiency.The comparison considers moving VMs between two ToR-connected server groups and shutting down an emptied ToR switch.

D. Proof of Theorem 3.

The proof compares evenly distributing one job’s VMs across k racks with assigning them to a single rack under roughly uniform intra- and inter-rack traffic. Under the stated tractability assumptions, concentration can reduce power consumption.

  • D. Proof of Theorem 3.: The comparison assumes one networking-intensive job, even VM distribution across k racks, and approximately uniform loads u and w.Startup cost σ is excluded because no switch is switched off in this comparison.
  • D. Proof of Theorem 3.: For MapReduce-like traffic with roughly equal intra-rack and inter-rack loads, assigning all VMs to one rack can reduce total switch power.The analysis defines ΔP = P2 − P1 and uses convexity of traffic-dependent power consumption.

E. Proof of Theorem 4.

The proof shows that moving an entire rack of VMs from one Fat-Tree pod to another does not reduce switch traffic and adds core traffic. Therefore, keeping the job within one pod is more energy-efficient under the stated topology properties.

  • E. Proof of Theorem 4.: Moving VMs from a rack in pod A to an empty rack in pod B never reduces ToR or aggregation traffic and adds traffic at core switches.Fat-Tree’s outer fan-out is no larger than its inner fan-out, so cross-pod placement cannot lower the relevant lower-level traffic.

F. Proof of Proposition 2.

This proof establishes that concentrating highly communicating VMs reduces network traffic, then analyzes how many aggregation switches should remain active. Under the stated conditions, using fewer aggregation switches yields better energy efficiency.

  • F. Proof of Proposition 2.: Placing VMs for the same job on the same servers reduces network traffic because same-server communication bypasses physical NICs and the network.The result specifically motivates compacting VMs with high communication traffic.
  • F. Proof of Proposition 2.: For n ≥ N_agg, the proof compares optimal power consumption with n and n + 1 aggregation switches.The argument orders the loads and analyzes the difference ΔP between the two optimal solutions.
  • F. Proof of Proposition 2.: When ΔP > 0, using fewer aggregation switches results in better energy efficiency.The proof derives this conclusion under α > 1 and the stated restriction on σ.
Loading 1304.3519v2…