Upload
angel-conde-manjon
View
281
Download
0
Embed Size (px)
Citation preview
© COPYRIGHT IKERLAN 2016 www.ikerlan.es
Solving the Industry 4.0. challenges on the logistics domain using Apache Mesos
Technological Centre
Angel Conde, Cristobal Arellano, Santi Charramendieta (IK4-Ikerlan)
Oscar Ocariz (ULMA Handling Systems)
© COPYRIGHT IKERLAN 2016 www.ikerlan.es
I. Introduction
II. Components
III. Architecture
IV. Conclusions & Future Work
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 3
ULMA Handling Systems
Provides all-round logistics systems (e.g., automatic warehouses)
Custom turnkey solutions
Design, development, assembly & maintenance
World Wide presence
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 4
An ULMA Warehouse
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 5
COMPONENTS
ConveyorStacker Crane
Elevator
A system that is physically distributed composed by independent devices that carry out automatic tasks
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016
COMMON FAILURES
• Physical failures &Deterioration
• Logical failuresSW errorsSW updatesLogical/Physical mismatch
“Automatic elements fail sooner or later”
6
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 7
ULMA & the Industry 4.0.
Reduce down time
Integration with 3rd systems
Lower maintenance costs
Predictive maintenance
Smart Warehouse
Industry 4.0
BIG DATA
CIBERSECURITY
IIOT
CLOUD
MODELING & SIMULATION
M2M
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 8
Supervisor
Message Bus
DDSAdapter
DDSAdapter
DDSAdapter
ULMA - Supervisor
Message Bus based on DDS
Operational Data
“The Supervisor system gathers information about operational data and malfunctions”
Distributed: monitoring system natureScalable: different size of warehouseExtendible: useful with different devicesReliable: without errors
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 9
The Supervisor & Industry 4.0.Useful for local maintenanceLocal data valuable for later analytics
However….
No remote monitoringManual data collectionOperational data can be lost
Industry 4.0
BIG DATA
CIBERSECURITY
IIOT
CLOUD
MODELING & SIMULATION
M2M
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 10
The Cloud Supervisor
Live Mirror of each Real Supervisor
Responsible of storing the operational data
Can define its own rules
Aggregates must be supported
Supervisor1
Supervisor2
Real Supervisors
Cloud Superviso
r 1
Cloud Superviso
r 1 & 2
Cloud
Operational Data
© COPYRIGHT IKERLAN 2016 www.ikerlan.es
I. Introduction
II. Components
III. Architecture
IV. Conclusions & Future Work
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 12
PLATFORM GOALS
Public Cloud, Private Cloud, Hybrid
Avoid Vendor Locking
Resiliency
Elasticity
Efficient resource usage
?
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 13
CLOUD AGNOSTIC PLATFORMWhy ?
From Static Resource Partitioning to Elastic Sharing
REST APICassandraSpark
Static Partitioning
Elastic Sharing
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 14
MESOS OVERVIEW
HTTPJVM
Python
Ruby
distributed resources: CPU, GPU, RAM, I/O, FS, rack locality, etc.
distributed file system
Cluster
DFS
Frameworks
KernelC++
Apps & Services
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 15
MESOS FEATURES
Battle tested on Twitter
Up to 10,000 nodes
Launch any task using or cgroups
Resource isolation
Commercial support via
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 16
INFRASTRUCTURE
Launch platform on any cloud provider
Provisioning machines
Automate updates/upgrades
Configuration
“Infrastructure as a code”
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 17
BUILDING BLOCKS
Data Ingestion & Message Bus (Data Sources)
RT/Batch Data Analytics (Machine learning, failure prediction)
Data Storage (Horizontal scaling)
Advanced Visualization and User Experience
Clou
d In
frast
ruct
ure
17
Supervisor
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 18
REVERSE PROXY& LB
Multiple backends (Mesos, Consul, Docker…)
Dynamic Watchers for backends
Open Source
HTTPS SNI
Deployed on the “edge node”
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 19
MESSAGE BUSDecouple the components
De Facto Standard in Big Data Architectures
Publish / Subscriber model
HA enabled
High Performance
Deployed using the Mesos Framework
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 20
STORAGE
HDFS deployed in HA (Mesos Master Nodes)
Files stored in Columnar format (Parquet)
Not deployed as Mesos framework
Horizontal scalability
“Small Files problem”
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 21
ANALYTICS
Big Data Analytics in Memory
Standard SQL Support
Deployed using the Mesos Spark Scheduler
Machine Learning
Real Time Processing
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 22
SERVICE DISCOVERY & MONITORING
Service Discovery using DNS records
Port discovery via DNS SVR records
Service Health-monitoring
Circuit Breakers
Node/Service Monitoring
Task registration via Marathon/Consul
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 23
PLATFORM OVERVIEW
Edge Node
MESOS Agent Nodes
Consul Client
Consul Client
Agent
Proxy
MESOS Master NodeNameNode
Consul Server
Master
External traffic
(HTTPS)
Secured Cloud Infrastructure
DataNode
Consul Client
Agent
DataNode
HTTP
ZooKeeper
Mesos coordinatorsHDFS NamenodesConsul for Service DiscoveryTraffic Ingestion
HTTPS SNI
Mesos Agent Consul ClientHDFS DatanodesTasks (Supervisors)
© COPYRIGHT IKERLAN 2016 www.ikerlan.es
I. Introduction
II. Components
III. Architecture
IV. Conclusions & Future Work
© COPYRIGHT IKERLAN 2016 www.ikerlan.es 25
Message BusIngestion/ Analytics
ULMA Cloud
Storage
Storage
Storage
Cloud Supervisor
Cloud Supervisor
Cloud Supervisor
Ingestion/ Analytics
Ingestion/ Analytics
1,000 FT OVERVIEW
Supervisor1
Supervisor2
Supervisor3
Oper
atio
nal D
ata
© COPYRIGHT IKERLAN 2016 www.ikerlan.es
JVM web based app
Dynamic port binding via Marathon
Cgroups used as containerizer
Entry points assigned via labels
Look for:- Random number generation exhaustion - JVM DNS Caching
26
PLATFORM COMPONENTSThe Cloud Supervisor
© COPYRIGHT IKERLAN 2016 www.ikerlan.es
Kafka producer module
Sends data to Kafka instead of the local system
Local queue used in case of connection problems
Each supervisor sends data to its own topic (aggregates)
27
PLATFORM COMPONENTSThe Save Interface for the (cloud) Supervisor
© COPYRIGHT IKERLAN 2016 www.ikerlan.es
Deployed using Spark Mesos Scheduler
Lambda architecture (RT / Batch)
Data Ingestion & Storage from Kafka
AVRO schemas simulating IEC61850.
Ingestion tested up to 100,000 events/sec
28
The Spark ServerPLATFORM COMPONENTS
© COPYRIGHT IKERLAN 2016 www.ikerlan.es 29
The Spark Server (II)PLATFORM COMPONENTS
Cloud Superviso
r1
Cloud Superviso
r2.
Operational DataStreaming
Retains recent data on memoryPersists operational data to HDFSLambda (RT/Batch) queries via SQL endpointREST API
SQL
Decouples the systemEach supervisor has its own topic
© COPYRIGHT IKERLAN 2016 www.ikerlan.es 30
The Spark Server (III)PLATFORM COMPONENTS
Operational Data
StreamingSQL
• Real-Time Data“Recent Data” in memory tableRT algorithmsData is saved to the “staging” HDFS directoryUncompacted Partitioned Parquet (date/supervisor)
• Batch DataExposes “batch” data via “Old Data” Table Reload compacted data from HDFS when the
compaction is done
© COPYRIGHT IKERLAN 2016 www.ikerlan.es
HDFS must: Avoid small files!!!!How to solve it?
Scheduled task executed by Chronos
Kite project is used for the task
“Staging” “Compacted”
31
PLATFORM COMPONENTSThe Compactor
© COPYRIGHT IKERLAN 2016 www.ikerlan.es
I. Introduction
II. Components
III. Architecture
IV. Conclusions & Future Work
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 33
CONCLUSIONS
Remote Monitoring already valuable for clients
Global data collection available 24x7 Platform easily deployable on any provider
Efficient resource usage
Analytics on recent and batch data using standard SQL
Industry 4.0
BIG DATA
CIBERSECURITY
IIOT
CLOUD
MODELING & SIMULATION
M2M
© COPYRIGHT IKERLAN 2016 www.ikerlan.es© COPYRIGHT IKERLAN 2016 34
FUTURE WORKMove to
Real-time analytics
Evaluate Cassandra as backend storage
Structured streaming, dynamic allocation on Spark 2.0.+
Run more ULMA software on the platform
Overlay networks, aka “IP per container”