CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS...

16
CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building [email protected]

Transcript of CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS...

Page 1: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

CSE 691: Energy-Efficient ComputingLecture 6

SHARING: distributed vs. localAnshul Gandhi

1307, CS [email protected]

Page 2: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

energy_routing paper

Page 3: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

# servers

Page 4: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

workload

Page 5: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

softscale paper

Page 6: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

6

Goals of a data center

• Low response times • Goal: T95 ≤ 500 ms

Performance• 70% is wasted• Goal: Minimize waste

Power

Load

Time

BUSY: 200 W IDLE: 140 W OFF: 0 W

Intel Xeon server

Page 7: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

7

Scalable data centers

Performance Power

BUSY: 200 W IDLE: 140 W OFF: 0 W

Intel Xeon server

Reactive: [Leite’10;Horvath’08;Wang’08]Predictive: [Krioukov’10;Chen’08;Bobroff’07]

Setup cost300 s

200 W(+more)

Only if load changes slowly

Load

Time

Page 8: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

8

Problem: Load spikesLo

ad

Time

x

2x

Page 9: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

9

Prior work

Dealing with load spikes• Spare servers [Shen’11;Chandra’03]

Over provisioning can be expensive

• Forecasting [Krioukov’10;Padala’09;Lasettre03]

Spikes are often unpredictable

• Compromise on performance [Urgaonkar’08;Adya’04;Cherkasova’02]

Admission control, request prioritization

x

Load

Time

2x

Page 10: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

10

Our approach: SOFTScale

• No spare servers• No forecasting • Does not compromise on

performance (in most cases)

Can be used in conjunction with prior approaches

x

Load

Time

2x

Page 11: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

Closer look at data centers

Always on

Use caching tier to “pick up the slack”

11

Scalable

Page 12: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

High-level idea

OFF

OFF

OFFSETUP

SETUP

SETUP

Load

Time

x

2x

Dual purpose

Leverage spare capacity

12

ON

ON

ON

Page 13: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

Experimental setup

PHP(CPU-bound)

Memcached(memory-bound)

Response time: Time for entry to exitAverage response time: 200ms (with 20X variability)Goal: T95 ≤ 500ms

Apache

13

Page 14: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

Experimental setup

14

8-core CPU4 GB memory 4-core CPU

48 GB memory

PHP(CPU-bound)

Memcached(memory-bound)

Apache

Page 15: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

Results: Instantaneous load jumps

15

Load

Time

50%

61%

10% 29%

T 95 (m

s)av

erag

ed o

ver 5

min

s

baseline = provisioned for initial load

Page 16: CSE 691: Energy-Efficient Computing Lecture 6 SHARING: distributed vs. local Anshul Gandhi 1307, CS building anshul@cs.stonybrook.edu.

Conclusion

16

• Problem: How to deal with load spikes?

• Prior work: Over provision, predict, compromise on performance

• Our (orthogonal) approach: SOFTScale

Leverages spare capacity in “always on” data tiers

Look at the whole system Can handle a range of load spikes