Join the Conversation #OpenPOWERSummit

Jez Wain, Escala Competency Centre Atos, Grenoble

Making Elephants Fly

Not so long ago

Power

x86

image: www.thegreenhead.com

RISC vs. CISC

https://media-cdn.tripadvisor.com/media/photo-s/04/12/37/ed/gatlinburg-photography.jpg

Data lock-in

image: http://pinterest.com/pin/121878733635791126

Power

Power x86

Common data format

Linux-el

Power x86

Common application data format

Linux-el

Common application data format

Power Database Files x86

Cold Standby

Power Database Files

Cold Standby

Power Database Files

Cold Standby

Power Database Files x86

Cold Standby - Fail-back

Database Files x86Power

Cold Standby

Database FilesPower

Improving the recovery time and recovery point

PostgreSQL Operation

Database Files

WALWALWAL

Ubuntu/x86Ubuntu/Power

Hot/Warm Standby

WALWALWAL Archived WAL

WALWALWAL

Transaction

Stream

Log

Shipping

WAL Directory

x86

Power

Web Server

Read + Write

Read Only

Log ship & TX stream

192.160.160.170

10.197.160.58

INSERT INTO rtdata (ts, ip, value)

VALUES(now(), inet_server_addr(), 0)

ID Timestamp DB IP Value

1631 2018-05-14 19:37:22 192.160.160.170 0

1630 2018-05-14 19:37:16 192.160.160.170 0

1629 2018-05-14 19:37:11 192.160.160.170 0

1628 2018-05-14 19:37:01 192.160.160.170 0

ID Timestamp DB IP Value

1631 2018-05-14 19:37:22 192.160.160.170 0

1630 2018-05-14 19:37:16 192.160.160.170 0

1629 2018-05-14 19:37:11 192.160.160.170 0

1628 2018-05-14 19:37:01 192.160.160.170 0

Both servers running

- The web-server is writing on main database

- The main database is replicated on the hot standby server.

- The web-server can read from both databases.

ID Timestamp DB IP Value

1713 2018-05-14 19:44:12 192.160.160.170 0

1712 2018-05-14 19:44:07 192.160.160.170 0

1711 2018-05-14 19:44:02 192.160.160.170 0

1710 2018-05-14 19:43:57 192.160.160.170 0

Primary DB halted

- Data still available on secondary (read-only) - Server cannot write until secondary promoted to primary

2018-05-14 19:44:17.390369+02 Cannot execute query INSERT INTO rtdata (ts, ip, value) VALUES(now(), inet_server_addr(), 0)

ID Timestamp DB IP Value

1723 2018-05-14 19:46:19 10.197.160.58 0

1722 2018-05-14 19:46:14 10.197.160.58 0

1721 2018-05-14 19:46:09 10.197.160.58 0

1720 2018-05-14 19:46:04 10.197.160.58 0

Standby promoted to primary

- After a few seconds the hot-standby is promoted as main db server. - Data are now written through the hot standby server (10.197.160.58)

Postgres-XL: Sharded PosgreSQL

- Scale-out, shared-nothing cluster

X86

Postgres-XL

X86 X86 X86

Coordinator

Global transaction manager

Data nodes

Postgres-XL: Sharded PosgreSQL

- Scale-out, shared-nothing cluster

X86

Postgres-XL

X86 X86 X86 Power Power

Postgres-XL: Sharded PosgreSQL

- Scale-out, shared-nothing cluster

Postgres-XL

Power Power

PostgreSQL on Power9 Performance Benchmark

- TPC-B style transactional workload

- Five queries: SELECT, UPDATE, INSERTRead-only (SELECT) option

- Simulates multiple clients

pgbench

pgbench scaling factor

Table # rows

pgbench_branches 1

pgbench_tellers 10

pgbench_accounts 100,000

pgbench_history 0

Scale = 1

Scale Rows in accounts DB size

100 10,000,000 1.5 GB

300 30,000,000 4.5 GB

1,000 100,000,000 15.0 GB

3,000 300,000,000 45.0 GB

Power9 14 cores, 64GBRHEL 7.5 4.14

PG Bench Injector

PG Bench Injector

PG Bench Injector

EMC Unity 300 (Flash)

NVMe

16GB/s FC

Postgres 10.5 & 11 beta4

pg_bench: scale 1000 = ~15 GB Select (read-only)

~700MB/s

8Gb FC Saturated

Tran

sact

ions

per

sec

ond

120K

220K

320K

420K

520K

620K

# Clients16 32 48 64 80 96 112 128 144

SCALE 1000 SCALE 1500 SCALE 2000 SCALE 2500

Power9: 14 cores 40 GB RAM

https://www.enterprisedb.com/blog/enterprisedb-postgres-plus-and-lenovo-3850-x6-equals-performance-and-scalability

Splendid Data – PostgresPURE on IBM Power System S822L benchmark report

16

7 Results

7.1 PostgreSQL 9.4.4 results

PostgreSQL 9.4.4 Scale 100 Scale 300 Scale 1,000 Scale 3,000

1:1 2:1 1:1 2:1 1:1 2:1 1:1 2:1

incl.

conn

Clients/Threads 26/26 48/24 28/28 52/26 30/30 52/26 28/28 48/24

TPS /Latency 423,148/0.123

379,866/0.126

406,589/0.138

377,233/0.138

379,414/0.158

368,224/0.141

365,974/0.153

355,882/0.135

excl.

conn

Clients /Threads 124/124 260/130 90/90 292/146 30/30 244/122 28/28 48/24

TPS /Latency 726,022/0.749

709,271/0.819

435,311/0.589

566,129/1.017

380,335/0.158

446,477/0.859

366,383/0.153

356,229/0.135

0"

50000"

100000"

150000"

200000"

250000"

300000"

350000"

400000"

450000"

0" 20" 40" 60" 80" 100" 120" 140"

TPS$includ

ing$conn

ec.o

n$establishing$

Clients$

TPS$5$PostgreSQL$9.4.4$

scale"100"

scale"300"

scale"1000"

scale"3000"

0"

0,2"

0,4"

0,6"

0,8"

1"

1,2"

0" 20" 40" 60" 80" 100" 120" 140"

Latency(in(m

illise

cond

s(

Clients(

Latency(in(ms(0(PostgreSQL(9.4.4(

scale"100"

scale"300"

scale"1000"

scale"3000"

Sébastien Chabrolles IBM Montpellier + Splendid Data Power8: 20 cores 256 GB RAM

0

12500

25000

37500

50000

INTEL E7-4890 2.8GHz (60c)

PostgreSQL 9.6

POWER8 3.42GHz (20c)

PostgreSQL 9.5

POWER9 3.3GHz (14c)

PostgreSQL 11beta4

X4

TPS/per core

Tran

sact

ions

per

sec

ond

26K

30K

34K

38K

42K

46K

50K

54K

# Clients16 32 48 64 80 96 112 128 144 160

EXT4 nobarrier pgbench -j8

Read-Write - scale 1,00014 cores P9@3.3GHz SMT8, 40GB RAM, shared_buffers=24GB

E4-900 : 14 cores P9@3.3GHz 40GB RAM PostgreSQL11beta4 pgbench -S with shared_buffers=8GB

Tran

sact

ions

per

sec

ond

200K

300K

400K

500K

600K

700K

# Clients

16 32 48 64 80 96 112 128

SMT4SMT8

SMT4 vs SMT8 on POWER9

Leading performance and cost optimized for Big Data and Analytics workloads

Standardize on LinuxEscala supports industry standard Linux distributions from RedHat, Suse and Ubuntu along with applications optimized for emerging business challenges such as Big Data & analytics, mobile and Java platforms. The support for little endian Linux ensures easy portability for thousands of Open Source applications. Leading ISV’s such as SAP also support the Linux on Power platform.

OpenStack Cloud enabling / automationThe PowerVC management solution turns Escala VMs, storage and virtual networks into a OpenStack based IAAS cloud solution. PowerVC groups multiple physical servers into managed pools, greatly simplifying the creation and deployment of AIX and Linux VMs. Features like Snapshots and cloning reduce the average deployment time of applications from several hours to just a few minutes. Automation can be integrated or triggered by popular Open Source tools such as Chef or Puppet.

The constant economic pressure to reduce IT infrastructure costs while increasing flexibility through dynamic, on-demand provisioning has driven many organizations towards standardization onto x86 hardware and the Linux operating system.As price/performance has become the key measure for infrastructure purchases, TCO calculations are mainly conducted in isolated domains such as hardware, OS/virtualization and support costs. However, bigger picture aspects such as the cost impact on ISV licenses, platform reliability, OS vulnerability and VM isolation are often not factored in.Escala delivers key advantages which can have a significant overall TCO impact such as a leading performance per core, the best reliability in the industry, superior security isolation and virtualization efficiency.

Escala Mission Critical Linux serversEscala Mission Crit ical Linux servers are designed for commercial workloads requiring performance, reliability and scalability. The Escala L1-800 and L3-800 servers use the latest Power8 processor technology delivering unprecedented performance, scalability, reliability, and manageability, for demanding commercial workloads. They are optimized to help deliver new solutions and services faster, and with higher quality. They are also optimized for big data and analytics, and provide the ideal foundation for scale-out data and cloud environments in a compact 2U package. They deliver superior throughput compared to x86-based offerings for similar workloads.

Escala OpenPower-based Linux servers Escala OpenPower-based Linux servers are optimized for Big Data and cloud workloads. The Escala L1-OP80 and L3-OP80 Power8 servers are built for most OpenSource workloads, as well as OpenSource databases, due to their cost efficiency and expandability. They are designed to run either Hadoop environments thanks to larger internal storage capacity and powerful SMT8 cores, or in-memory analytics such as Spark, as well as innovative solutions from the OpenPower partner ecosystem such as CAPI. They are also ideal for hyperscale ISP environements where high-efficiency and cost-effectiveness are the most important features.

Atos and Linux advantages:• Market leading memory bandwidth

• Slim packaging with high internal disk expandability

• Ideal for Big Data as a Hadoop node thanks to internal disk expandability and powerful SMT8 cores

• Ideal for Spark, OpenSource databases (PostgresSQL, MariaDB..) and Business applications (SAP)

discover the open alternative: Escala linux solutions

escala mission critical servers

https://disneymovieyear.files.wordpress.com/2013/01/photo20.jpg

Photo by Austin Chan on Unsplash