Advantages of running OpenShift on Power/OpenPOWER systems

Pack-em or stack-em: Advantages of running
OpenShift on Power
—
Mithun HR
Software Engineer - Cognitive Systems
Krishna Harsha Voora
Software Engineer - Cognitive Systems

Agenda
IBM ISV Team/ June 26, 2020 / © 2020 IBM Corporation 2
IBM and Red Hat Partnership
Overview of Red Hat OpenShift
Use Cases for OpenShift
Meet The POWER9 Family
Objective
The Workload
Demo

3
Together, IBM and Red Hat will support our clients with forward-thinking,
flexible, cloud-ready infrastructure solutions – optimized for Power Systems.
IBM and Red Hat – Partners for 20 Years
For over 20 years, IBM and Red Hat have collaborated with the Open Source community to drive innovation
and power businesses around the world.
IBM ISV Team/ June 26, 2020 / © 2020 IBM Corporation

4
Overview of Red Hat
OpenShift

5

6
Meet The POWER9 family
Big Data Workloads Enterprise AI Workloads
LC922 AC922
Mission Critical Workloads
S922/S914/S924
H922/H924/L922
Power Enterprise SystemsPower Scale-Out Systems
E950 E980

7
ppc64le for Data intensive workloads

8
Objective
• Develop Proof Point for Price/Perf per core Advantage
• Identify Workload
• Choose CPU Intensive Workload
• Geospatial Workload from MongoDB
• Based on MEAN Stack
• Define SLAs
• Identify Container Density Testing
• Pack as many containers per host as possible
• Work through Price/Perf per core

9
The Workload
(idea from MongoDB Manual è https://docs.mongodb.com/manual/tutorial/geospatial-tutorial/)
Collection #1 - Neighborhoods
• 195 rows
• 4 Megabytes raw
• Each row has an array of geospatial points that outline the neighborhood
Collection #2 - Restaurants
• 23,359 rows (restaurant listings)
• Each row has the location- single geospatial coordinate (Lat/Long) and the establishment’s name
• 3.2 Megabytes raw
Each MicroService (Node.Js Application, Express, Mongoose) call runs 1 Query (Jmeter Script v4)
• Get 10 neighborhoods (MongoDB picks 10 and returns them)
• Get My Neighborhood (fixed location)
• GetNeighborhoods (pass in location coordinates)
• GetRestaurants located in a Donut shape around me (max & min range)
• GetRestaurants in range
Each “Jmeter user” calls all 5 micro services per loop. 4 users, 10 loops = 200 transactions
Short runs are about 8 mins long – 675,000 transactions
Long runs are about 30 mins long – 2,700,000 transactions

10
Components POWER9 x86_64
Hypervisor PowerVM KVM
OpenShift Container Platform v3.11.104 v3.11.104
OS RHEL 7.6 (Maipo) RHEL 7.6 (Maipo)
Kernel 3.10.0-957.12.1.el7.ppc64le 3.10.0-957.12.2.el7.x86_64
Container OS CentOS Linux 7(Core) CentOS Linux 7(Core)
Container OS Kernel 3.10.0-957.12.1.el7.ppc64le 3.10.0-957.12.2.el7.x86_64
MongDB Enterprise 4.0.2-1.e17.ppc64le 4.0.2-1.e17.x86_64
Node.js v8.14.1-linux-ppc64le v8.14.1-linux-x64
The Software Stack

11

12

13
MongoDB workload on IBM Power
3.2X greater containers per core than tested Intel Xeon SP Gold 6150 servers (Skylake)
IBM
Power L922
(20-core, 256GB, 2 LPARs)
174 containers
Intel Xeon SP based
2-socket server
(36-core, 256GB, 2 VMs)
98 containers
Server price 2,3,4
-3-year warranty
$28,821 $28,805
Geospatial workload
Total Transactions per Second
- With 2 VM’s
2542 tps 2290 tps
Containers/core 8.7 2.7
3.2 X
Greater
containers/core
1. Based on IBM internal testing running MongoDB’s Geospatial queries at 700 users, each running 1000 transactions using jmeter v4. Each container uses MongoDB 4.0.2 & Node.js v8.14.1 (REST APIs) with socket bound containers. Testing added
containers to each server until servers reached response time limit of 99% of transactions completing in under 1 second. Results valid as of 7/17/19. Conducted under laboratory condition with speculative execution controls to mitigate user-to-kernel and user-
to-user side-channel attacks on both systems, Individual result can vary based on workload size, use of storage subsystems & other conditions. Details about MongoDB workload: https://docs.mongodb.com/manual/tutorial/geospatialtutorial/
2. 3.2X greater containers/core is based on 174 containers/20 cores for Power L922 and 98 containers/36 cores for Intel Xeon. – (2,531/20)/(2,290/36) = 3.2
3. IBM Power L922 (2x10-core/typical 2.9 GHz/256 GB memory) 2x 388 GB SSD, 2x 10 Gb two-port network, RHEL 7.6 with PowerVM (2 partitions@10-cores each),
4. Competitive stack: 2-socket Intel Xeon Skylake Gold 6150 (2x18-core/ 2.7 GHz/256 GB memory), 2 x 480 GB SSD, 3 x 10 Gb two-port network, RHEL 7.6, KVM (2 VMs@18-cores each)
5. Pricing is based on Power L922 https://www.ibm.com/it-infrastructure/power/scale-out, and publicly available x86 pricing https://ark.intel.com/content/www/us/en/ark/products/120490/intel-xeon-gold-6150-processor-24-75m-cache-2-70-ghz.html

14
A container is the smallest compute unit
CONTAINER

15
image name
labels
cpu
memory
Storage
replicas
POD
CONTAINER
POD
CONTAINER
POD
CONTAINER
DEPLOYMENT
pods configuration is defined
in a deployment

16
TRAIN MORE. BUILD MORE. KNOW
MORE.
Industry leading innovations
• Large Model Support
• Distributed Deep Learning
• Coherence
• NVLink 2.0
3.7x
2.3x
3.8x

17

18
Summary
Get more work done -- Perf/Core is high
Fastest memory lives on cores -- P9 runs more containers
More data than ever lives -- Higher throughput means more containers
Faster innovation and value -- $/Container pp64le vs Intel.
IBM PowerVM:
https://www.ibm.com/in-en/marketplace/ibm-powervm
Red Hat OpenShift on PowerVM:
http://www-01.ibm.com/common/ssi/ShowDoc.wss?docURL=/common/ssi/rep_sm/p/877/ENUS5639-
OCP/index.html&lang=en&request_locale=en

Demo
19IBM ISV Team/ June 26, 2020 / © 2020 IBM Corporation
https://github.com/krishvoor/cloud_interop
https://developer.ibm.com/tutorials/mongodb-nodejs-on-openshift/

Advantages of running OpenShift on Power/OpenPOWER systems

More Related Content

What's hot

Similar to Advantages of running OpenShift on Power/OpenPOWER systems

More from Ganesan Narayanasamy

Recently uploaded

Advantages of running OpenShift on Power/OpenPOWER systems