Yi Cloud10 Pres

8/8/2019 Yi Cloud10 Pres

1/29

Reducing Costs of Spot Instances via

Checkpointing in the Amazon Elastic

Compute Cloud

Sangho Yi, Derrick Kondo, and Artur Andrzejak

- at CLOUD 2010 -

Sangho Yi

INRIA Grenoble, France([email protected])


2/29

Outline

Motivation and background

Spot Instances on Amazon EC2

Possible checkpointing strategies Performance evaluation

Conclusions and future work


3/29

Motivation:

Market-based Pricing in Cloud Amazon released Spot instances in Dec.

2009.

Current price depends on users bids, andavailable computing resources.

(based on the supply and demand relationship)

30~50% price of normal instances.

In-bid users get instances.

Out-of-bid users lose their instances.


4/29

Pricing in Amazon EC2

(normal instances on demand)Virtual MachineInstance Type

Linux

Cost/hour (USD)

MS Windows

Cost/hour (USD)

Standard Small 0.095 0.12

Standard Large 0.38 0.48

Standard Extra Large 0.76 0.96High CPU - Medium 0.19 0.29

High CPU - Extra Large 0.76 1.16Data Transfer

Type

Cost/GB-Month

(USD)

Inbound Transfer 0.10

First 10 TB 0.17

Next 10-50 TB 0.13

Next 50-150 TB 0.11

Over 150 TB 0.10

** Pricing on eu-west (Ireland), 2010


5/29

Pricing in Amazon EC2

(spot-instances bidding system)

Source: http:/cloudexchange.org


6/29

An Example of Failures as a

Function of Users Bid

Users bid = 0.081

Availability Availability Availability

Failure Failure

Current Price


7/29

PDF of Failure arrivals


8/29

Goal

Understand impact of checkpointing methods onspot instances in terms ofmonetary cost and

execution time

0

Price

Time (hour)1 2 3 4 5 6 7 8 9 10

0.1

0.20.3

0.4Bid

availability (5) availability (3)failure (2)


9/29

Possible Checkpointing

Strategies In terms of taking points

The optimal case

Without checkpointing

Hourly checkpointing

Rising edge-driven checkpointing

Checkpointing with adaptive decision whetherto skip or take a checkpoint

Combinations of aboves


10/29

Characteristics of Strategies

Hourly

Pay as you go (with checkpointing per hour)

Rising edge-driven Takes when price goes up (except for out-

of-bid cases)

Adaptive taking point decision

Has overhead for decision, also thedecision may not work well in some cases.


11/29

Hourly vs. Edge-driven

tc

tc

tc

Time (minute)0 60 120 180

Failure

(without

payment)

Recovery

Pay per hour Pay per hour Pay per hour

Task execution Task execution Task execution

Task

execution

Time

Price

fora

s

potinstance

Users bid

Available duration Available duration

Failure

Recovery

: checkpoint

: rising edge


12/29

Adaptive Taking Point Decision

Based on time series analysis,

it compares the expected costs at a certain point

when Taking a checkpoint

Skipping a checkpoint

Then, selects the lower-cost option.


13/29

System Model


14/29

To take or skip, that is the question

tc

tc

Time0

Failure

Recovery

(short rollback)


tc

Time0

Failure

Recovery

(long rollback)

Task execution Task execution

(a) When taking a checkpoint at the current time

Current time

Current time

(b) When skipping to take a checkpoint at the current time


15/29

Cost analysis on the choice


16/29

Delayed termination a way of

exploiting Amazons policy


Failure(out-of-bid)



Task

execution Without payment


End of execution

(when user stops)



Pay last partial-hour

(a) When a users spot instance is out-of-bid

(b) When a user stops using spot instance


Users bid

Available durationPrice for spot instance


17/29

Performance evaluation

Trace-based simulation

http://spotckpt.sourceforge.net

42 spot instances (now Amazon has 64)

24 (12x2) different checkpointing strategies


18/29

Checkpointing strategies

OPT: The optimal case

NONE: Without checkpointing

H: Hourly

E: Rising edge-driven

AH: H with adaptive decision

AE: E with adaptive decision

H+E, H+AE, AH+E, AH+AE

AF(10): adaptive decision every 10 mins

AF(30): adaptive decision every 30 mins


19/29

Results #1: Monetary cost

Total cost according to Users bid on two spotinstances


20/29

Results #2: Execution time

Total execution time according to Users bid onthe spot instances.


21/29

Results #3: Normalized one

Results #4: Benefit of using DT


22/29

Results #5: Best ones on the

mean price bidding

Hourly checkpointing shows good performance.

Fine-grained checkpointing shows good performance

sometimes.

Adaptive decision does not show significant impact onthe results.


23/29

Summary

Checkpointing strategies have ~50% higheroverhead (execution time * monetary cost)

than the optimal case.

Thus, better checkpointing strategies arerequired to reduce both cost and time.


24/29

Current and Future work

1) Finding relationship between past andfuture price of spot instances

2) Developing efficient checkpointingmethod to minimize monetary cost and

execution time (to appear @ EKC 2010)

3) Providing decision model to determineusers bid, given reliability and performance

constraints (to appear @ MASCOTS 2010)


25/29

Thanks! / Questions?

Lets walk (or work) on the Cloud!


26/29

** On-going work on 2)

Sangho Yi and Derrick Kondo How CheckpointingCan Reduce Cost of Using Clouds?, EKC 2010, July,

2010. (accepted)

We used the known characteristics of AmazonsSpot Instances, and better checkpoint decision

mechanism.

The proposed scheme shows much less

overhead for most cases. (~20% overhead)


27/29


28/29

** On-going work on 2)


29/29

Lets Make Our Cloud Better!

On a Cloud Vendors point of view,

TODO

(efficiently) Adapt management strategies to fullyutilize the limited (very-large) resources (with less

cost)

Better

Utilization,Less Cost

More

Cloud Users

MoreMoney

More Investment

on R&D

Yi Cloud10 Pres

Documents

Transcript of Yi Cloud10 Pres