Spin-down Disk Model

1
Spin-down Disk Model
Spinning Seek
Spinning Access
Spinningup
Request
Triggerrequest or predict
Predictive
NotSpinning
Spinning Ready
Spinningdown
Inactivity Timeout threshold
2
Spin-down Disk Model
Etransition Ptransition Ttransition
1- 3s delay
Spinning Seek
Spinning Access
Spinningup
Request
Triggerrequest or predict
Predictive
Tidle
Etransition Ptransition Ttransition
NotSpinning
Spinning Ready
Spinningdown
Pdown
Pspin
ToutInactivity Timeout threshold
Tdown
3
Reducing Energy Consumption

• Energy S Poweri x Timei

• Energy S Poweri x Timei
• To reduce energy used for task
• Reduce power cost of power state I through better
  technology.
• Reduce time spent in the higher cost power
  states.
• Amortize transition states (spinning up or down)
  if significant.
• PdownTdown 2Etransition Pspin Tout lt
  PspinTidle
• Tdown Tidle - (Ttransition Tout)

i e powerstates
4
Power Specs

• IBM Microdrive (1inch)
• writing 300mA (3.3V)1W
• standby 65mA (3.3V).2W

• IBM TravelStar (2.5inch)
• read/write 2W
• spinning 1.8W
• low power idle .65W
• standby .25W
• sleep .1W
• startup 4.7 W
• seek 2.3W

5
Spin-down Disk Model
2.3W
4.7W
2W
Spinning Seek
Spinning Access
Spinningup
Request
Triggerrequest or predict
Predictive
NotSpinning
Spinning Ready
Spinningdown
.2W
.65-1.8W
6
Spin-Down Policies

• Fixed Thresholds
• Tout spin-down cost s.t. 2Etransition
  PspinTout
• Adaptive Thresholds Tout f (recent accesses)
• Exploit burstiness in Tidle
• Minimizing Bumps (user annoyance/latency)
• Predictive spin-ups
• Changing access patterns (making burstiness)
• Caching
• Prefetching

7
Dynamic SpindownHelmbold, Long, Sherrod
(MOBICOM96)

• Dynamically choose a timeout value as function of
  recent disk activity
• Based on machine learning techniques (for all you
  AI students!)
• Exploits bursty nature of disk activity
• Compares to (related previous work)
• best fixed timeout with knowledge of entire
  sequence of accesses
• optimal - per access best decision of what to do
• competitive algorithms - fixed timeout based on
  disk characteristics
• commonly used fixed timeouts

8
Metrics

• Spindown cost (s) Tidle (sec.) s.t.
  2Etransition PspinTidle
• Seconds of Energy 1 unit Pspin 1sec. -
  Pdown 1 sec. Joules
• Energy used by timeout
• idle time if idle time lt timeout
• timeout spindown cost if idle gt timeout

PspinTidle____________ Pspin - Pdown
9

• Energy used by optimalidle time if idle lt
  spindown costspindown cost otherwise (immediate
  spindown)
• Excess energyEnergy used by timeout - Energy
  used by opt
• LossExcess energy / idle time.

10
Share Algorithm

• Learning algorithm
• Each trial, set of experts make prediction
• Weighted average of experts predictions
• After each trial update weights of experts
• reduce weights of misleading experts
• share the slashed weights among good experts

11
Specifics

• For Disk Spindown timeout calculation
• Each expert is a fixed timeout value
  (100)weights wi initially 1/n
• Learning rate used to slash misleading experts h
• 4.0
• Share parameter to recover expert a
• .08

12
For each trial

• Timeout S wi xi / S wi
• Slashes weights for next timewi wie-h
  Loss(xi)
• Shares remaining weightspool S wi(1- (1- a)
  Loss(xi))wi (1- a) Loss(xi) wi 1/n pool
• Claim little sensitivity to parameters
• Behavior idleness makes timeout slowly get
  smaller busy makes timeout jump to longer

13
Experiments and Results

• Used traces of HP C2474s disks (not many details
  about environment - multiuser?)
• Most idle times in 0-1 sec range, lt1sec next most
  frequent
• Results this algorithm beats everything except
  optimal
• Avoids inappropriate spindowns