Cost Aware Fault Recovery in Clouds (IM 2013)

8/22/2019 Cost Aware Fault Recovery in Clouds (IM 2013)

1/29

COST AWARE FAULT RECOVERYIN CLOUDSAssaf Israel, Danny RazTechnion - Israel Institute of Technology


2/29

FAULTS IN DATACENTERS

Weve come a long way in terms of server resilience

Enterprise gra

Component A

Compute

(CPU, RAM, Fans, Net)

~

Storage ~


3/29

FAULTS IN DATACENTERS Typical first year of a new 1800 servers cluster @ Google:

- thousands of hard drive failures

~1000 individual machine failures

~3 router failures (have to immediately pull traffic for an hour)

~5 racks go wonky (40-80 machines see 50% packet loss)

~20 rack failures (40-80 machines instantly disappear, 1-6 hours to get ba

~1 network rewiring (~5% of machines down over 2-day span)

~1 PDU failure (~500-1000 machines suddenly disappear, ~6 hours to com

~0.5 overheating (power down most machines in


4/29

FAULTS IN DATACENTERS

Other factors also contribute to lack of resilience

Distribution of service disruption evenThe Datacenter as a Computer (200


5/29

RECOVERY

Most of the time we would like to recover as quickly as p

Backup


6/29

RECOVERY

Most of the time we would like to recovery as quickly as Single host recovery may take advantage of vacant re

Backup


7/29

RECOVERY

Most of the time we would like to recovery as quickly as Single host recovery may take advantage of vacant re

Backup


8/29

RECOVERY

Larger failures (Racks, Network segments, Power regionsMay require powering more machines

Backup


9/29

RECOVERY

Larger failures (Racks, Network segments, Power regionsMay require powering more machines

Backup


10/29

RECOVERY COST

ServiceDegradation

BackupInfrastructure

RecoveryCost


11/29

RECOVERY COST

ServiceDegradation


RecoveryCost

,

,,

, - Service deg. cost of when recovered at - Infrastructure cost of

, , - 0/1 Decision vectors

Can be formally expressed as:


12/29

RECOVERY COST

Service degradation depends on: Task setup/initialization

Host setup/initialization

Network configuration (if recovered to a different network segme

Storage mapping

Storage migration (if recovered to a different SAN)

Software patches

Integrity checks

Manual host configuration

Recovery target location (latency/bandwidth)

ServiceDegradation



13/29

RECOVERY COST

Pre-planning can help reduce recovery cost

Activating additional backup infrastructure: Can help lowering some of Service Degradation costs

At the expense of additional maintenance costs

ServiceDegradation



14/29

OBSERVATION

Not all tasks are equal Interactive & vital monitoring

High-priority non-interactive

Non-interactive user-facing

Batch

Housekeeping tasks

Some are more susceptible to long downtimes than oth

Web-scW. Cirne

Tight SLA

Relaxed SLA


15/29

GOAL

We would like to recover expensive tasks faster Balance service degradation and infrastructure costs

Backup


16/29

GOAL

We would like to recover expensive tasks first Balance service degradation and infrastructure costs

Backup


17/29

GOAL

Formal: Minimize the total recovery cost

Infrastructurecosts

Service degradationcosts

Under somepacking constraints


18/29

APPROXIMATION - OVERVIEW

Integer Program

LP Relaxation

Linear

Transformations ||Light Graphs

CycleBreaking

Activation

RoundingApproximation bounds

Cost 1 Load

6


19/29

IF WE HAD MORE INFO

Backup


20/29

IF WE HAD MORE INFO

If we knew which of backup hosts are active we could approximate the Service degradation costs

Backup


21/29

MINIMUM GENERAL ASSIGNMENT PROB

Bins, Items Each item have a size, depends on the target bin

Each item have a cost, depends on the target bin

Goal:Packall items into bins at minimum cost, under packing c


22/29

MIN-GAP

Has been studied extensively Known results:

LP-Based 2-Approx. (Shmoys and Tardos, 1993)

LP-Based

-Approx. (Fleischer, Goemans, Mirrokni and Svir Local Ratio-Based 2 -Approx. (Cohen, Katzir and Raz, 2006)


23/29

LOCAL SEARCH

Iteratively find the next backup machine to activate Stop when theres no improvement in recovery costs

Backup

Active host

Inactive hostBase cost - All backups are

inactive Next AcFind theactivate

recover

is minim(Using it

Stop conditionIf( < ):

return last RPElse:

Activate return +


24/29

SIMULATIONS

Based on data from IBM Research Compute Cloud (RC

Several hundreds hosts, with a few thousands VMs

4 host configurations, 3 VM configurations

EC2-like SLA policies(higher availability guaranties, at higher rates)


25/29

RECOVERY COST BY RACK SIZE

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1 2 3 6 10 17 34

Cost[%]

Rack size (#hosts/rack)

Normalized Recovery Cost by Rack size


26/29

RECOVERY COST BY VM SLA DISTRIBUTIO

0

50000

100000

150000

200000

250000

0

0.0

2

0.0

4

0.0

6

0.0

80.1

0.1

2

0.1

4

0.1

6

0.1

80.2

0.2

2

0.2

4

0.2

6

0.2

80.3

0.3

2

0.3

4

0.3

6

0.3

80.4

0.4

2

0.4

4

0.4

6

0.4

80.5

Cost

SLA Distribution

2 host racks - Total & Service costs

20% - Cheap to recover

80% - Expensive to recover


27/29

RECOVERY COST BY VM SLA DISTRIBUTIO

0

50000

100000

150000

200000

250000

0

0.0

2

0.0

4

0.0

6

0.0

80.1

0.1

2

0.1

4

0.1

6

0.1

80.2

0.2

2

0.2

4

0.2

6

0.2

80.3

0.3

2

0.3

4

0.3

6

0.3

80.4

0.4

2

0.4

4

0.4

6

0.4

80.5

Cost

SLA Distribution

2 host racks - Total & Service costs

Active

Servic

Inactiv

ServicLocal

Servic


28/29

CONCLUSION

Large scale infrastructure mandates fault tolerance tec

Pre-planning can help reduce recovery cost

Classifying tasks by SLAs can improve overall recovery c

LP-Based Load/Cost Approximation with guaranteed pe Local Search heuristic with good practical performance


29/29

THANK YOU !

Questions ?

Cost Aware Fault Recovery in Clouds (IM 2013)

Documents