• Nenhum resultado encontrado

On the Cost of Database Clusters Reconfiguration

N/A
N/A
Protected

Academic year: 2021

Share "On the Cost of Database Clusters Reconfiguration"

Copied!
67
0
0

Texto

(1)

On the Cost of Database Clusters Reconfiguration

R. Vilaça, J. Pereira ,R. Oliveira Universidade do Minho

Portugal

J.E. Armendáriz-Iñigo, J.R. González de Mendívil Universidad Pública de Navarra

Spain

(2)

Online Database Recovery

Reconfiguration of a database cluster required to face load spikes or to restore the resilience of the system

Update new replicas to the most current database state online

Minimize time required to finish recovery

Minimize impact on resource usage and performance of

(3)

Motivation

Prominent practical problem

O'en addressed before (eg., [Kemme et al. , 2001], [Jiménez-Peris et al. , 2002], [Armendáriz-Íñigo et al. , 2007]

Applied to consistent database replication

Existing refined techniques to improve performance Existing work

Assumes simplified models without taking into

account system limitations such as I/O and CPU

(4)

Goals

Combine the proposed techniques in a single protocol Systematically benchmark different reconfiguration scenarios

Assess each technique’ s performance impact and overhead on the clustered database service

Determine fundamental limits to cluster reconfiguration

Discuss the relative merits of each approach

(5)

Based on the algorithm presented in [Kemme et al. , 2001]

Parallel Recovery [Jiménez-Peris et al. , 2002]

Convergence Phases [Armendáriz-Íñigo et al. , 2007]

Includes several obvious optimizations, eg:

purging of redundant data changes data compression

parallel and batch applier for received recovery data at the recovering replica.

Configurable: number of donors for Parallel Recovery

(6)

Recovery Protocol

Full Transfer

Missed Updates or Delta Transfer

Parallel Recovery

Convergence Phases

(7)

Full Transfer

(8)

Full Transfer

r1

r2

r3

t

(9)

Full Transfer

r1

r2

r3

t

(10)

Full Transfer

r1

r2

r3

t

V iew Change

(11)

Full Transfer

r1

r2

r3

t

V iew Change

(12)

Full Transfer

r1

r2

r3

t

V iew Change V iew Change

(13)

Full Transfer

r1

r2

r3

t

V iew Change V iew Change

(14)

Full Transfer

r1

r2

r3

t

Send Full Database

V iew Change V iew Change

(15)

Full Transfer

r1

r2

r3

t

Send Full Database

Received Full Database

V iew Change V iew Change

(16)

Full Transfer

r1

r2

r3

t

Send Full Database

Received Full Database

V iew Change V iew Change

(17)

Full Transfer

r1

r2

r3

t

Send Full Database

Received Full Database

Applied Full Database

V iew Change V iew Change

(18)

Full Transfer

r1

r2

r3

t

Send Full Database

Received Full Database

Applied Full Database

Catch-Up

V iew Change V iew Change

(19)

Full Transfer

r1

r2

r3

t

Send Full Database

Received Full Database

Applied Full Database

Catch-Up Normal Processing

V iew Change V iew Change

(20)

Missed Updates

(21)

Missed Updates

r1

r2

r3

t

V iew Change V iew Change

(22)

Missed Updates

r1

r2

r3

t

V iew Change V iew Change

(23)

Missed Updates

r1

r2

r3

t

Send Missed Updates

V iew Change V iew Change

(24)

Missed Updates

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

V iew Change V iew Change

(25)

Missed Updates

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

V iew Change V iew Change

(26)

Missed Updates

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

Applied Missed Updates

V iew Change V iew Change

(27)

Missed Updates

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

Applied Missed Updates

Catch-Up

V iew Change V iew Change

(28)

Missed Updates

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

Applied Missed Updates

Catch-Up Normal Processing

V iew Change V iew Change

(29)

Missed Updates

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

Applied Missed Updates

Catch-Up Normal Processing

V iew Change V iew Change

Mini mal set

of upd ate, ins ert

or d eletes

(30)

Missed Updates

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

Applied Missed Updates

Catch-Up Normal Processing

V iew Change V iew Change

Mini mal set

of upd ate, ins ert or d eletes

No rma l del ega te trans acti on

pro cess ing

(31)

Parallel Recovery

(32)

Parallel Recovery

r1

r2

r3

t

V iew Change V iew Change

(33)

Parallel Recovery

r1

r2

r3

t

V iew Change V iew Change

(34)

Parallel Recovery

r1

r2

r3

t

Send Missed Updates

V iew Change V iew Change

Send Missed

Updates

(35)

Parallel Recovery

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

V iew Change V iew Change

Send Missed

Updates

(36)

Parallel Recovery

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

V iew Change V iew Change

Send Missed

Updates

(37)

Parallel Recovery

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

Applied Missed Updates

V iew Change V iew Change

Send Missed

Updates

(38)

Parallel Recovery

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

Applied Missed Updates

Catch-Up

V iew Change V iew Change

Send Missed

Updates

(39)

Parallel Recovery

r1

r2

r3

t

Send Missed Updates

Received Missed Updates

Applied Missed Updates

Catch-Up Normal Processing

V iew Change V iew Change

Send Missed

Updates

(40)

Convergence Phases

(41)

Convergence Phases

r1

r2

r3 t

V iew Change V iew Change

(42)

Convergence Phases

r1

r2

r3 t

V iew Change V iew Change

(43)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

V iew Change V iew Change

(44)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

V iew Change V iew Change

(45)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

V iew Change V iew Change a se 1

(46)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

V iew Change V iew Change a se 1

(47)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

V iew Change V iew Change a se 1

(48)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

V iew Change V iew Change a se 1 a se 2

(49)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

V iew Change V iew Change a se 1 a se 2

(50)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

V iew Change V iew Change a se 1 a se 2

(51)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

V iew Change V iew Change a se 1 a se 2 a se n

(52)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

tc h - U p

V iew Change V iew Change a se 1 a se 2 a se n

(53)

Convergence Phases

r1

r2

r3 t

Send Missed Updates

tc h - U p

Normal Processing

V iew Change V iew Change a se 1 a se 2 a se n

(54)

4 commodity servers on par with the systems used in recent related work

Intel Core 2 Duo at 2.13GHz, 1GB RAM and dedicated SATA HD.

Replicas ran an instance of PostgreSQL 8.1 and a Java Virtual Machine (1.5.0) for the Replication service.

No failures during the recovery process

At most one replica was recovering during evaluation Flow Control on the incoming rate of update

transactions during recovery

(55)
(56)

Two standard benchmarks

TPC-C: write intensive workload, IO stressing (15 warehouses, 150 clients, 2.2GB database)

TPC-W , read intensive workload, CPU stressing (Shopping Mix,

400 clients, 10000 items, 2.4GB database)

(57)

Two standard benchmarks

TPC-C: write intensive workload, IO stressing (15 warehouses, 150 clients, 2.2GB database)

TPC-W , read intensive workload, CPU stressing (Shopping Mix, 400 clients, 10000 items, 2.4GB database)

Evaluation of recovery with all replicas close to their

nominal capacity, maximum load that does not saturate machines.

TPC-C TPC-W

(58)

Recovery Time: Full Transfer

TPC-C 150 clients 2.2GB database throughput 110 tpm

TPC-W 400 clients 2.4GB database

throughput 3000 tpm

0 5 10 15 20 25

Recovery T ime (m)

0 20 40 60 80 100

Flow Control (%)

2.5 5 7.5 10

Recovery T ime (m)

Specific database

dump and restore tools

(59)

TPC-C 150 clients 2.2GB database throughput 110 tpm

TPC-W 400 clients 2.4GB database

throughput 3000 tpm

Recovery Time: Convergence Phases

1 phase 2 phases 5 phases

0 25 50 75 100

Recovery T ime (m)

0 20 40 60 80 100

Outdatedness (MB)

0.25 0.5 0.75

1

Recovery T ime (m)

(60)

TPC-C 150 clients 2.2GB database throughput 110 tpm

TPC-W 400 clients 2.4GB database

throughput 3000 tpm

Recovery Time: State Donors

1 donor 2 donors 3 donors

0 5 10 15 20

Recovery T ime (m)

0 20 40 60 80 100

Outdatedness (MB)

0 0.25

0.5 0.75

1

Recovery T ime (m)

2 Phases

2 Phases

(61)

Recovery Time: Flow Control

0% FlowControl 5% FlowControl 10% FlowControl

20% FlowControl 50% FlowControl 100% FlowControl

0 5 10 15 20

Recovery T ime (m)

0 20 40 60 80 100

Outdatedness (MB)

0.2 0.4 0.6 0.8

Recovery T ime (m)

2 Phases, 3 Donors

2 Phases, 3 Donors

TPC-C 150 clients 2.2GB database throughput 110 tpm

TPC-W 400 clients 2.4GB database

throughput 3000 tpm

(62)

Throughput

2 -> 1 state donor, 5 convergence phases and no flow control 3 -> 2 state donor, 2 convergence phases and no flow control 4 -> 3 state donor, 2 convergence phases and no flow control 5 -> 3 state donor, 2 convergence phases and 5% flow control 6 -> 3 state donor, 2 convergence phases and 10% flow control 1 -> 1 state donor, 2 convergence phases and no flow control 0

25 50 75 100 125

Throughput (TPM)

1 2 3 4 5 6

1000 2000 3000 4000

Throughput (TPM)

TPC-C, 150 clients, 2.2GB database TPC -W , 400 clients, 2.4GB database

Before and A'er recovery

(63)

Recovery is about 7x longer than donation

For a TPC-C workload, a 86MB recovery log and a single donor:

1145

Donor replica 168

Recovering replica

s

(64)

Conclusion

Database online recovery protocol combining several

previously proposed optimization techniques

(65)

Conclusion

Database online recovery protocol combining several previously proposed optimization techniques

The results of our tests do not reveal any relevant effect of the optimizations on the recovery time or on the

overall cluster performance either.

(66)

Conclusion

Database online recovery protocol combining several previously proposed optimization techniques

The results of our tests do not reveal any relevant effect of the optimizations on the recovery time or on the

overall cluster performance either.

The capacity of the recovering replica to apply the

received state turns out to be the salient limiting factor

(67)

Conclusion

Database online recovery protocol combining several previously proposed optimization techniques

The results of our tests do not reveal any relevant effect of the optimizations on the recovery time or on the

overall cluster performance either.

The capacity of the recovering replica to apply the

received state turns out to be the salient limiting factor Most research has been targeted at optimizing the

operations that are not, by a large margin, limiting

factors in overall performance

Referências

Documentos relacionados

É o caso de Gioconda Belli, em El pais de las mujeres (2010), mais de um século depois da obra de Emília Freitas. Mesmo separadas pelo tempo e espaço, é possível observar que

In this sense, the following practices of LDC played a key role: (1) the participatory attitude of the counselor in the client’s experience, providing process markers to guide

Innovation costs too high, Uncertain demand for innovative goods or services, Difficulties in finding cooperation partners for innovation, Lack of qualified personnel, Lack

Desta forma e inserindo os dados obtidos no modelo explicativo da evolução da ocupação do vale superior do rio Coura, entre o Bronze Final e a Romanização, e dada a

The main goal of the proposed approach is to allow the discovery of concise, and accurate, high-level rules (from a biological database named dbSNP - Database

La segunda contribución de las autoras es el desarrollo de un pensamiento de carácter feno- menológico-existencial – en el caso de Arendt, en diálogo crítico con

Da Silva complementa afirmando que “tornando-se um ser que busca a integridade, terá condições de elaborar recursos adaptativos que transformem o seu tempo, que é único,

Assim, partimos da hipótese de que a afetividade tem in- fluência direta no processo de ensino e aprendizagem em sala de aula, visto que, como afirma Leite (2006), ela pode