ElasticTree: Saving Energy in Data Center Networks 許倫愷2013/5/28

About the paper

Brandon Heller, Srini Seetharaman, Priya Mahadevan, Yiannis Yiakoumis, Puneet Sharma, Sujata Banerjee, Nick McKeown

NSDI’10 (USENIX conference on Networked systems design and implementation)

Citation: 174

16 pages

Outline

The big picture

Introduction

ElasticTree system

Analysis

Conclusion

Outline

The big picture

Introduction

ElasticTree system

Analysis

Conclusion

The motivation

The motivation

Very inefficient!!

Desired

Why wasting power

Provisioning for peak

Time varying traffic demands

Low efficiency at low loads

The goal of ElasticTree

The approach…

Turn off unneeded links and switches

The challenge

Performance

Fault tolerance

Scalability

Outline

The big picture

Introduction

ElasticTree system

Analysis

Conclusion

Introduction

What is ElasticTree:

ElasticTree is a system for dynamically adapting the energy consumption of a data center network

• What does it do:

Finding minimum-power network subsets across a range of traffic patterns

Trade-off:

energy efficiency, performance and robustness

Introduction

Data center network

(Traditional) 2N Tree:

One failure can cut the effective bisection bandwidth in half; two failures can disconnect servers

Data center network

Fat tree: SIGCOMM 2008, A Scalable, Commodity Data Center Network Architecture

Data center network

provision for peak workload

Traffic varies daily, weekly, monthly, and yearly.

Energy Proportionality

The strategy: turn off the links and switches that we don’t need

Outline

The big picture

Introduction

ElasticTree system

Analysis

Conclusion

ElasticTree

ElasticTree is a system for dynamically adapting the energy consumption of a data center network

ElasticTree

If 0.2 Gbps of traffic per host ,1 Gbps link…

ElasticTree

13/20 switches and 28/48 links stay active

ElasticTree reduces network power by 38%

0.2

0.4

0.8

ElasticTree

The optimizer: find the minimum- power network subset which satisfies current traffic conditions

Optimizer

As traffic conditions change, the optimizer continuously re-computes the optimal network subset

3 approaches:

Formal Model , Greedy Bin-Packing , Topology-aware Heuristic

Optimizer comparison

Formal model

Finding the optimal flow assignment alone is an NP-complete problem for integer flows.

Derived from standard multi-commodity flow (MCF) problem

The model outputs a subset of the original topology, plus the routes taken by each flow to satisfy the traffic matrix

O(n^3.5+)

Greedy Bin-Packing

Strategy: choose the leftmost one with sufficient capacity

O(n^2+)

1G link

Greedy Bin-Packing

1G link

Topo-aware Heuristic

1. does not compute the set of flow routes

2. assumes perfectly divisible flows

=> pack every link to full utilization and reduce TCP bandwidth

=> starter subset

Decoupling power optimization from routing :

=> can be applied alongside any fat tree routing algorithm

Topo-aware Heuristic

An edge switch doesn’t care which aggregation switches are active, but instead, how many are active

Topo-aware Heuristic

Decoupling power optimization from routing

Optimizer comparison

Outline

The big picture

Introduction

ElasticTree system

Analysis

Conclusion

How to test

K = 6, fat tree

OpenFlow

Analysis

Traffic pattern:

Near: servers communicate only with other servers through their edge switch

Far: servers communicate only with servers in other pods

Analysis

Random demand:

Individual aggregation/core switches turning on/off

Analysis70% to outside, 30% inside DCN

Different traffic load

Analysis: redundancy

If only the MST is on

=> no redundancy => no fault tolerance

Analysis: redundancy

+MST: additive cost, multiplicative benefit

Analysis: latency

0.250.33 0.5

Need safety margin!!

Ethernet overheads (preamble, inter-frame spacing, and the CRC) cause the egress buffer to fill up Packets either get dropped or significantly delayed

Analysis: latency

Safety margin is the amount of capacity reserved at every link by the optimizer

Traffic overload is the amount each host sends and receives beyond the original traffic matrix

Trade-off between Energy and Performance

Outline

The big picture

Introduction

ElasticTree system

Analysis

Conclusion

Summary

Reference

The paper

The slide (by the author)

A youtube video (by the author, too)

http://www.youtube.com/watch?v=G2_D-CH4tQk

Questions

Thank you!