Managing Energy and Server Resources in Hosting Centers

1
Managing Energy and Server Resources in Hosting
Centers
Jeff Chase, Darrell Anderson, Ron Doyle, Prachi
Thakar, Amin Vahdat Duke University
Review by Paskorn c
2
Managing Energy and Server Resources

• Key idea a hosting center OS maintains the
  balance of requests and responses, energy inputs,
  and thermal outputs.

• Adaptively provision server resources to match
  request load.
• Provision server resources for energy efficiency.
• Degrade service on power/cooling failures.

energy
responses
requests
waste heat
3
Proposed Idea

• Architecture/prototype for adaptive provisioning
  of server resources in Internet server clusters
  (Muse)
• Software feedback
• Reconfigurable request redirection
• Foundation for energy management in hosting
  centers
• 25 - 75 energy savings
• Simple economic resource allocation
• Continuous utility functions customers pay for
  performance.
• Balance service quality and resource usage.

4
Before Muse

• Dedicate fixed resources per customer
• Reprovision manually as needed
• Overprovision for surges
• High variable cost of capacity

How to automate resource provisioning for managed
hosting?
5
Load Is Dynamic

• ibm.com external site
• February 2001
• Daily fluctuations (3x)
• Workday cycle
• Weekends off

• World Cup soccer site
• May-June 1998
• Seasonal fluctuations
• Event surges (11x)
• ita.ee.lbl.gov

6
Adaptive Provisioning
- Efficient resource usage - Load multiplexing -
Surge protection - Online capacity planning -
Dynamic resource recruitment
- Balance service quality with cost - Service
Level Agreements (SLAs)
7
(No Transcript)
8
Muse Architecture
Executive
performance measures
configuration commands
Control
offered request load
storage tier
reconfigurable switches
server pool

• Executive controls mapping of service traffic to
  server resources by means of
• reconfigurable switches
• scheduler controls (shares)

9
Utilization Targets
?i allocated server resource for service i
?i utilization of ?i at is current load ?i
( ?I request throughput (hit per minute)
?target configurable target level for ?i Leave
headroom for load spikes.
?i gt?target service i is underprovisioned
?i lt?target service i is overprovisioned
10
Energy vs. Service Quality
A
A
B
B
C
Active set A,B,C,D
Active set A,B
D

• ?i lt?target
• Low latency

• ?i ?target
• Meets quality goals
• Saves energy

11
Resource Economy

• Input the value of performance for each
  customer i.
• Common unit of value money.
• Derives from the economic value of the service.
• Per-customer utility function Ui bid penalty.
• Bid for traffic volume (throughput ?i).
• Bid for better service quality, or subtract
  penalty for poor quality.
• Allocate resources to maximize expected global
  utility (revenue or reward).
• Predict performance effects.
• Sell ? to the highest bidder.
• Never sell resources below cost.

12
Bid and Penalties

• Bid (/hpm)
• Example 5 cents for 1k to 10k hpm
• Delivered Throughput
• ? i(t,?i)
• Hence, revenue
• Bid x Deliveraed Throughput
• bidi x ? i(t,?i)

13
Bid and Penalties

• Penalties
• Since, ? resource (Supply)
• , r demand from user
• Then, penalties happens when
• ?i lt ri and ?i gt?target ( Customer is
  underprovisioned)
• Amount of penalty is function of
• ri / ?i

14
Objective Function
Maximize ? bidi(?i(t, ?i)) Subject to ??i ?
?max
15

• To make model tractable
• Price is updated by

16

• To change amount of resource for each service
• There are two functions
• Shrink(i, target) reclaims resource from service
  i and increase price to target
• grow(i, target) assigns more resource to service
  i and decrease price to target

17
(No Transcript)
18
To change p and ?i(t, ?i))

• EWMA-based filter alone is not sufficient.
• Average At for each interval t At ?At-1
  (1-?)Ot
• The gain ? may be variable or flip-flop.
• Load estimate Et Et-1 if ?Et-1 - At? lt
  tolerance
• else Et At
• Stable
• Responsive

19
Muse Prototype and Testbed
Executive
SURGE or trace load generators
power meter
Extreme GigE switch
LinkSys 100 Mb/s switch
redirectors (PowerEdge 1550)
server pool
client cluster
Request IBM 43M in Feb 5-11 01
FreeBSD-based redirectors resource containers APM
and Wake-on-LAN
20
IBM Trace Run (Before)
Power draw (watts) Latency (ms50)
Throughput (requests/s)
1 ms
21
IBM Trace Run (After)
1 ms
22
Action Items

• Read more paper
• Thinking of simulaiton to show
• Use Muse
• Configuration
• Heterogeneos
• Homogeneous diff power consumption
• Green grids
• Siligon Valley Conf. AUTOOMOUS.

23
Server Power Draw
866 MHz P-III SuperMicro 370-DER (FreeBSD) Brand
Electronics 21-1850 digital power meter
boot 136w
CPU max 120w
CPU idle 93w
watts
Idling consumes 60 to 70 of peak power demand.
disk spin 6-10w
off/hiber2-3w
work