Can we define semantic relatedness in any meaningful terms

Download Report

Transcript Can we define semantic relatedness in any meaningful terms

“ElasticTree: Saving energy in data center
networks“
by Brandon Heller, Seetharaman, Mahadevan,
Yiakoumis, Sharma, Banerjee, McKeown
presented by Nicoara Talpes, Kenneth Wade
About the paper
•Published in April 2010 at Networked
Systems Design & Implementation (NSDI)
Motivation 1
•Efforts spend so far on servers and cooling. Our focus is
on the network (10-20% total power)
•Environmental Protection Agency: estimate that in 2011
networks in data centers will consume 12 B kWh
•This is 6.542.640 tons CO21
Motivation 2
•Goal: energy proportionality
Motivation 3
•Cannot get to green line using the hardware
•Common network goal is to balance traffic evenly
among all links: power is constant regardless of load
•‘Data centers provisioned to run at peak workload,
below capacity most of the time’
•Today’s network elements not energy proportional:
switches, transceivers waste power at low loads
•Switches consume 70% of full power when idle
Existing networks
•2N: fault tolerant
Wasted power
•Servers draw constant power independent of traffic
•time varying demands, provisioned for peak
ElasticTree
• Goal: build network that has energy proportionality even
if switches don’t
• By using traffic management and control of switches:
turning on switch consumes most of the power; 8%:
going from zero to full traffic; turning off switch saves
most power
• Careful: minimizing effects on performance and fault
tolerance
• Has to work at scale to make an impact
• With ET, we do opposite than balanced networks: only
use a few links, lower power at low loads (ex: middle
night)
Existing networks: scale-out
•Ex: Fat-tree; incremental degradation
Existing networks: scale-out
Implementation 1
• Optimizer: find minimum power network subset
which satisfies current traffic. Inputs: topology, traffic
matrix, switch’s power models, fault tolerance
constraints. Outputs new topology
• Continually re-computes subset as traffic changes
• Power control: toggles power states of ports,
linecards, entire switches
• Routing: chooses paths for all flows, pushes routes
into network
Implementation
Optimizer methods: formal model
•
•
•
•
outputs subset & flow assignments
Evaluates solution quality of other optimizers
optimal
(con) scales to number of hosts ^ 3.5
Optimizer methods: formal model
• Doesn’t scale
Optimizer methods: Greedy-Bin packing
Optimizer methods: Greedy-Bin packing
Optimizer methods: Greedy-Bin packing
Optimizer methods: Greedy-Bin packing
Power savings: data centers
• 30 % traffic inside dc, greedy-bin packet optimizer,
scaled, reductions of 25-60%: energy elastic!
Need for redundancy
• Nice propriety: cost drops with increase of network
size since MST is smaller fraction
Optimizer methods: Greedy-Bin packing
• Scales better, optimal solution not guaranteed, not
all flows can be assigned
• Understand power savings for larger topologies
Optimizer methods: topology aware heuristic
• Quickly find subsets in networks with regular
structure (fat tree)
• Requires less information: only need the cross-layer
totals, not the full traffic matrix
• Routing independent: does not compute set of flow
routes, (con) assumes divisible flows; can be applied
with any fat tree routing algorithm (Portland); any
full-bisection-bandwidth topologies with any nr
layers (ex 1gb at edge, 10 gb core)
• Simple additions to this lead to quality solutions in a
fraction of time
Optimizer methods: topology aware heuristic
Optimizer comparison
Optimizer comparison
• formal model intractable for large topologies greedy
• Un-optimized single-core python implementation: 20s
Control software
• ET requires traffic data and control over flow paths.
we use Open Flow: generate traffic? and push
application level flow routes to switches
Implementation
Implementation 2
• Openflow: measure traffic matrix, control routing
flows
• Open flow: vendor neutral so no need to change
code when use HP/ECR switches
• Experiments show savings 25-40% feasible: 1 bill
KWhr annual savings; then we have proportional
reduction in cooling costs
Experiments
• Topologies: two 3-layer k=4 fat tree; one 3-layer k=6
fat tree
• Measurements: NetFPGA traffic generator: each
emulates four servers
• Latency monitor
Experiments
3 Power savings results
• Formal method. Savings depend on network utilization
• traffic all inside. near traffic at low utilization: 60%
reduction
Power savings: sine-wave demand
• Reduction up to 64%
Robustness: safety margins
• MSTs disadvantages: renounces path redundancy and fault
tolerance
• Added cost of fault tolerance insignificant for large networks
Performance
• Uniform traffic shows spikes, large delays for packets
Safety margins
• safety margins defer points of loss, degrade latency
• Margins are adjustable
Topology aware optimizer
• Better robustness by tweaks: setting the link rate
utilization in equations to absorb overloads and
reduce delay
• setting the switch degree to add redundancy for
improving fault tolerance
• Solves constraints: Response times dominated by
switch boot time (30sec - 3 min)
• Fault tolerance: move topology-aware optimizer to
separate host to prevent crashes to affect routing.
• Traffic prediction experiments encouraging; can use
greedy algorithm
7 Discussion
• During low to mid-utilization, it respects the
constraints while lowering the costs
References
Some images borrowed from the author’s presentation
available at his website or online at
http://www.usenix.org/events/nsdi10/tech/
1-http://www.nef.org.uk/greencompany/co2calculator.htm