Openpower & IBM Power Systems

OpenPOWER & IBM Power Systems The best platform for PostgreSQL workload

René Akeret Systems Architect IBM Switzerland akeret@ch.ibm.com

Version 1.6

PGDay 2017 | © IBM Corporation PostgresPURE and Power Systems

Higher Performance Lower TCO PostgreSQL tps Price per Transaction

365’974 tps EUR 0.32

190’000 tps EUR 0.11

IBM Power S822L x86 Intel Xeon IBM Power S822L x86 Intel Xeon • 20c / 160t • 60c/ 120t • 20c / 160t • 60c/ 120t

For detailed benchmark results, see Benchmark Report “Splendid Data PostgresPURE on IBM Power System S822L” (http://www.splendiddata.com/benchmark-results-postgrespure-ibm-powerlinux-8/).

2 PGDay 2017 | © IBM Corporation Agenda

POWER8 Processor Technology

OpenPOWER & IBM Power Systems

PostgreSQL and POWER - The winning team

Summary Quarter Century of POWER Legacy of Leadership Innovation Driving Client Value

14nm

POWER10 2020+ POWER9

Big Data & Analytics Era

Modern UNIX Era

1990 1995 2000 2005 2010 2015 2017 …

4 PGDay 2017 | © IBM Corporation POWER8 Four Level Cache Design

Up to 2 TB per socket Up to 230 GB/s sustained

C C C C C C Mem L4 L4 Mem L1 L1 L1 L1 L1 L1 L2 L2 L2 L2 L2 L2

Mem L4 L4 Mem

Shared eDRAM L3

Mem L4 L4 Mem

L2 L2 L2 L2 L2 L2 L1 L1 L1 L1 L1 L1 Mem L4 L4 Mem C C C C C C

L1: 96 KB per Core L3: 96 MB shared per socket L2: 512 KB per Core L4: 16MB per memory card

5 PGDay 2017 | © IBM Corporation Cache is Critical to Good Performance

Memory is slow 1 Core relative to clock cycle cache 2-100 Cache clock cycles

400-800 Memory clock cycles

6 PGDay 2017 | © IBM Corporation PostgreSQL database, server point of view

Server processes Core performance • Single thread • Multithread • Throughput

SMP interconnect • Bandwidth Locking • Latency

Memory performance Shared memory • Bandwidth • Latency

I/O performance • Bandwidth • Latency Data files

7 PGDay 2017 | © IBM Corporation Exploits POWER8 Hardware - Speed

Up to Up to 230 GB/s Up to 4.35 GHz memory bandwidth 4 NVIDIA P100 GPU processors per socket workload accelerators

Up to 230 MB Up to 96 GB/s cache per peak I/O bandwidth socket per socket

8 PGDay 2017 | © IBM Corporation POWER8 developed for Big Data & Analytics

4X 4X 5X threads per core vs Intel memory bandwidth vs more cache vs Intel (up to 1536 threads per Intel (up to 230MB cache per system) (up to 32TB of memory) socket)

Processors Memory Cache flexible, fast execution of large, fast workspace to ensure continuous data analytics algorithms maximize business insight load for fast responses

Continously Dataflow Extreme I/O Parallel Processing Flash for very good Large Memory Blocks Bandwidth performance

First especially for Big Data developed Processor

9 PGDay 2017 | © IBM Corporation OpenPOWER and IBM Power Systems

10 PGDay 2017 | © IBM Corporation IBM Power Systems portfolio

S822 S814 S824 • Up to 20 Cores • Up to 8 Cores • Up to 24 Cores • Up to 1TB Memory • Up to 1TB Memory • Up to 2TB Memory

S812L S822L • Up to 12 Cores • Up to 24 Cores • Up to 512GB Memory • Up to 1TB Memory

S824L • Up to 24 Cores • Up to 2TB Memory

E850C E870C E880C • Up to 48 Cores • Up to 80 Cores • Up to 192 Cores • Up to 4TB Memory • Up to 16TB Memory • Up to 32TB Memory

11 PGDay 2017 | © IBM Corporation Power Enterprise Pools & CoD Lower Costs

A B C A B C

Before After

• Planned maintenance • Failover clusters • Rebalance capacity • POWER8 migration

12 PGDay 2017 | © IBM Corporation Power Reliability and Availability Strategy

Avoid Make soft Let Focus on errors by errors hardware planned and design harmless take care of unplanned hardware outages

13 PGDay 2017 | © IBM Corporation Virtualization Options for Power Systems

Hypervisor

PowerVM is Power Virtualization that will continue to be enhanced to support AIX, IBM i Workloads as well as Linux Workloads. PowerVM is supported on non-LC Systems (Exxx, Sxxx, SxxxL)

KVM provides an open source choice for Power Virtualization for Linux workloads. Best for clients that aren’t familiar with Power and Linux centric admins. PowerKVM is supported on L and LC Systems

14 PGDay 2017 | © IBM Corporation IBM Power Systems run at Higher Utilization

For a single workload, we require a machine capacity When we consolidate 4 workloads we 6.0x the average demand only require 3.5x average demand

When we consolidate 144 workloads we only require 1.42x average demand

• Fewer servers • Improved data center management • Reduced software licensing • Reduced HW & SW maintenance • Opportunity for significant growth within a single physical System

☛ Lower TCO

15 PGDay 2017 | © IBM Corporation Performance Utilization Guarantee

65% 70% 80%

POWER8 POWER8 POWER8 Scale-Out E850 E870 or E880

Real TCA and TCO Savings with

To qualify clients must implement each of the criteria set forth in the “POWER8 Performance Guarantee Requirements”. See IBM for full details and remediation steps IBM will take in accordance with the Guarantee.

16 PGDay 2017 | © IBM Corporation OpenPOWER Foundation - 300+ Members

This is What a Revolution Looks Like

17 PGDay 2017 | © IBM Corporation OpenPOWER Linux Cluster (LC) Systems

S822LC for S822LC for S812LC S822LC - GCA S821LC Big Data HPC

1 socket, 2U, Linux 2 socket, 2U, Linux Up to 20 Cores Up to 20 Cores Up to 20 Cores 8 or 10 cores 16 or 20 cores Up to 512 GB POWER8 with Up to 512 GB Up to 1 TB memory Up to 1 TB memory memory NVLink memory Up to 112 TB 2 Disks 12 HDD/SSD / Up to 1 TB memory 4 HDD/SDD / Storage 5 PCI slots NVMe 2 HDD/SDD NVMe 12+2 Disk Bays 5 PCI Slots 3 PCI Slots. 2 CAPI 4 PCI Slots 4 PCI Slots (4 CAPI, 2 K80 Up to 4 P100 GPU (3 CAPI, 1 K80 GPU) GPU)

KVM / Bare Metal KVM / Bare Metal KVM / Bare Metal Bare Metal KVM / Bare Metal

Hadoop & Spark Commercial Big Data HPC Compute

Get the maximum number of VMs running pgbench while average response-time is < 1ms

IBM Power System Intel x86 Xeon System

Start “pgbench –S –j2 –c2” on 16 new VMs

Number of VMs Wait 15 seconds

else If average response-time is < 1ms then

After 5min of stabilized run, the server who runs the more VMs wins !

For detailed benchmark info’s, see https://www.youtube.com/watch?v=5uKuQ8nzJns&t=120s 21 PGDay 2017 | © IBM Corporation Higher Performance with Linux on POWER

IBM Power System S822L x86 Intel Xeon

• 20x POWER8 cores, 3.42GHz • 4x Intel Xeon 2.80GHz (60 cores) • 256 GB memory • 256 GB memory • PowerKVM, RHEL 7.1 LE, • RHEL 7.1 / PostgreSQL 9.4.4 PostgreSQL 9.4.4

365’974 transactions / second 190’000 transactions / second

For detailed benchmark results (pgbench), see Benchmark Report “Splendid Data PostgresPURE on IBM Power System S822L” http://www.splendiddata.com/benchmark-results-postgrespure-ibm-powerlinux-8

IBM Power System S822L x86 Intel Xeon

• 20x POWER8 cores, 3.42GHz • 4x Intel Xeon 2.80GHz (60 cores) • 256 GB memory • 256 GB memory • PowerKVM, RHEL 7.1 LE, • RHEL 7.1 / PostgreSQL 9.4.4 PostgreSQL 9.4.4

0.11 EUR / transaction 0.32 EUR / transactions

For detailed benchmark results (pgbench), see Benchmark Report “Splendid Data PostgresPURE on IBM Power System S822L” http://www.splendiddata.com/benchmark-results-postgrespure-ibm-powerlinux-8

For detailed benchmark info’s, see https://www-03.ibm.com/systems/power/solutions/data-platform/

• 600 developers in over 100 open source projects to improve Linux • 500+ IBM Linux SW products • Work with industry organizations to expand the use of Linux

Germany

Israel China United States India

Brazil Australia

Highest throughput per Superior virtualization and core and core/memory management features to bandwidth to deliver faster afford flexibility and business results, up to 2x maximum utilization. Intel-based alternatives.

Truly open ecosystem and Lower total cost of ownership collaboration for extreme compared to virtualization on accelerator innovation with x86, including reduced industry-leading database licensing costs. technology partners

Thank you!

IBM (Switzerland) René Akeret - Systems Architect