GreenSlot : Scheduling Energy Consumption in Green Datacenters

GreenSlot: Scheduling Energy Consumption in Green Datacenters

Íñigo Goiri, Kien Le, Md. E. Haque,Ryan Beauchea, Thu D. Nguyen,

Jordi Guitart, Jordi Torres, and Ricardo Bianchini

Motivation• Datacenters consume large amounts of energy• Energy cost is not the only problem– Brown sources: coal, natural gas…

• Lots of small and medium datacenters• Connect datacenters to green sources– Solar panels, wind turbines…– Green datacenter

Green datacenter• Energy sources– Solar/wind: variable availability over time– Electrical grid: backup

• Other (problematic) approaches– Batteries: losses, cost, environmental– Bank energy on the grid: losses, cost, unavailability

Wind Pow

er

Time

Sola

r Pow

er

Scheduling scientific workloads

• Batch jobs• User specifies: #nodes, estimated runtime, deadline• Challenge– Match workloads with green energy availability

Power

Time

Load

GreenSlot

• Predict green energy availability– Weather forecast

• Schedule jobs– Maximize green energy use– If green not available, consume cheap brown

• May delay jobs but must meet deadlines• Turn off idle servers to save energy

Dealing with energy costs

• Schedule jobs: evaluate energy cost– Green energy is “free” (amortization): $0.00/kWh– Cheap (off peak, 11pm to 9am): $0.08/kWh– Expensive (on peak, 9am to 11pm): $0.13/kWh

• Optimization goal– Minimize energy cost while meeting deadlines

J3

Conventional vs GreenSlot

J1

J2J3

J2J3

Nod

esPow

erN

odes

Power

Time

J3 J1 J2Now

J1

GreenSlot: scheduling round

Time

Power

1. Divide “scheduling window” into slots (15 minutes)2. Predict green energy availability3. Consider jobs by earliest start deadline

– Calculate cost starting at every slot– Schedule job at the cheapest slot

4. Dispatch actions– Calculate and start required servers– Start jobs to be executed now– Deactivate unneeded servers (ACPI S3 state)

1. Divide “scheduling window” into slots (15 minutes)2. Predict green energy availability3. Consider jobs by earliest start deadline

– Calculate cost starting at every slot– Schedule job at the cheapest slot

4. Dispatch actions– Calculate and start required servers– Start jobs to be executed now– Deactivate unneeded servers (ACPI S3 state)

10 5 0 0 0 5 10 15 X X

GreenSlot: scheduling round

Time

Power

J1

GreenSlot behavior

J2

Time

J1

J2

Now

Nod

esPow

er

J1J2

Schedule:

Brown electricity priceJob deadlineScheduling window

J1, J2

J1J3

J4

GreenSlot behavior

J2

Time

J1

J2

J4

J3

Nod

esPow

er

J3J4

Schedule:

Brown electricity priceJob deadlineScheduling window

Now

J3, J4

J1

J4

J3

GreenSlot behavior

J2

Time

J2

J1J3

Nod

esPow

er

J4

J4

Schedule: J4 Weather prediction was wrong

Brown electricity priceJob deadlineScheduling window

Now

J1

J4

J5J3

GreenSlot behavior

J2

Time

J2

J1J3 J5

Nod

esPow

er

J4

J5

Schedule:

Brown electricity priceJob deadlineScheduling window

Now

J5

Evaluation methodology

• Cluster with 16 nodes– Modified version of SLURM– GreenSlot implemented on top

• Energy profile– NJ electricity pricing (on/off peak)– Solar farm energy availability (10 panels)– Four weeks (most, best, average, and worst)

• Schedulers– Conventional: EASY backfilling– GreenSlot: Green energy, Brown electricity price

Evaluation methodology

• Workload– Real workload from BSC– Workflows for sequencing yeast genome– 5 days (Monday to Friday)– Deadlines: 9am, 1pm, and 4pm

Monday Tuesday Wednesday Thursday Friday

Energy prediction vs actual

6:00 AM

7:00 AM

8:00 AM

9:00 AM

10:00 AM

11:00 AM

12:00 PM

1:00 PM

2:00 PM

3:00 PM

4:00 PM

5:00 PM

6:00 PM

7:00 PM0

0.51

1.52

PredictionActual

Ener

gy (k

Wh)

0 6 12 18 24 30 36 42 480

5

10

15

20

Hours ahead

Erro

r (%

)

GreenSlot for BSC workloadCo

nven

tiona

lGr

eenS

lot

26 kWh75 kWh

$8.00

38 kWh63 kWh

$6.06 -24%

24% cost savings

GreenSlot for BSC workload

Green energy increase Cost savings0

20

40

60

80

100

120MostBestAverageWorst

%

Other results

• Impact of weather miss-predictions– Less than 1% cost savings

• Workloads variations: Staggered and Multi-node– Consistent green energy increases and cost savings

• Workload intensity (datacenter utilization)– Works well with low/medium utilization– High switches to conventional

• Inaccurate user run time estimations– Maximum cost increase of 2%

Staggered workloadCo

nven

tiona

lGr

eenS

lot

32 kWh69 kWh

$8.58

38 kWh63 kWh

$6.00 -30%

30% cost savings

Conclusions

• Parallel job scheduler for green datacenters• Predicts green energy availability

• Increases the use of green energy• Reduces energy related costs• Solar array amortized in 11 years (18 years originally)

• We are building a solar-powered μDatacenter

GreenSlot: Scheduling Energy Consumption in Green Datacenters

Íñigo Goiri, Kien Le, Md. E. Haque,Ryan Beauchea, Thu D. Nguyen,

Jordi Guitart, Jordi Torres, and Ricardo Bianchini