01ClusterComputing

Cluster Computing

Outline

Computer Cluster
Big Data
Traditional Data Architectures
Meet Hadoop !
Big Data on Database Clusters
Big Data on Kubernetes

What is a computer cluster ?

A computer cluster is a collection of computers (also called as nodes) connected through high speed local are network that work together to simulate a single much more powerful computer system.

Each node in a computer cluster is controlled by its own operating system. All nodes in a computer cluster a controlled by a specialized scheduler/resource allocator that enforces even

distribution of work among the nodes in a computer cluster.

Each node in a computer cluster performs a different version of the same task.

A difference between computer cluster and computer grid is such that the nodes in a computer grid perform different tasks.

An architecture of computer cluster ranges from a simple two-node system connecting two personal computers to a supercomputer with a cluster architecture that consists

of 50,000 - 100,000 nodes (also called as a warehouse computer).

Computer Cluster

What is a computer cluster ?
A computer cluster is a collection of computers (also called as nodes) connected through high speed local are network that work together to simulate a single much more powerful computer system
Each node in a computer cluster is controlled by its own operating system
Each node in a computer cluster performs a different version of the same task
A difference between computer cluster and computer grid is such that the nodes in a computer grid perfrom different tasks
An architecture of computer cluster ranges from a simple two-node system connecting two personal computers to a supercomputer with a cluster architecture

Computer clusters are used to speed up computing through shared nothing (sharding) partitioning of data and parallelization of data processing on the nodes of a cluster.

Computer clusters provide high availability through automatic replacement of failed node with a replica node.

Advantages of computer clusters: faster processing speed, larger storage capacity, better data integrity, greater reliability and wider availability of resources.

A Linux cluster is a collection of connected array computers that can be viewed and managed as a single system.

A sample computer cluster: Specification of Nova cluster at Iowa State University:

54 regular compute nodes (with two 32-Core Intel 8358 processors, 1.6TB of local NVME storage and 512GB of memory each) and

5 GPU nodes with two 24-Core AMD EPYC 7413 processors, eight A100 GPU cards, 960GB of local storage and 512GB of memory each.

Computer Cluster

Computer clusters are used to speed up computing through shared nothing (sharding) partitioning of data and parallelization of data procesing on the nodes of a cluster
Computer clusters provide high availability through automatic replacement of a failed node with a replica node
Advantages of computer clusters: faster processing speed, larger storage capacity, better data integrity, greater reliability and wider availability of resources
A Linux cluster is a collection of connected computers that can be viewed and managed as a single system
A sample computer cluster: 54 regular compute nodes (with two 32-Core Intel 8358 processors, 1.6TB of local NVME storage and 512GB of memory each) and 5 GPU nodes with two 24-Core AMD EPYC 7413 processors, eight A100 GPU cards, 960GB of local storage and 512GB of memory each

What is cluster computing ?

Cluster computing is the process of sharing the computation tasks among multiple computers included in a computer cluster.

Advantages of cluster computing: cost efficiency, processing speed, expandability, high availability of resources.

At the moment cluster computing is an attractive paradigm for processing large scale science, engineering and commercial applications. Cluster computing

allows for parallelization of computations and data processing with relatively cheap hardware. It is probably the most cost-effective solution for parallelization

of data processing.

Cluster computing requires the specialized algorithms like load balancing, resource sharing and resource scheduling for optimization of data processing on a cluster

of computers.

Cluster computing is an attractive alternative to data processing on large parallel supercomputers. It is mainly due to availability of cheap hardware components

and very simple extensibility of the system.

The simplest configuration of nodes for cluster computing consists of a mater node and slave nodes.

Computer Cluster

What is a cluster computing ?
Cluster computing is the process of sharing the computation tasks among multiple computers included in a computer cluster
Advantages of cluster computing: cost efficiency, processing speed, expandability, high availability of resources
At the moment cluster computing is an attractive paradigm for processing large scale science, engineering and commercial applications
Cluster computing requires the specialized algorithms like load balancing, resource sharing and resource scheduling for optimization of data processing
Cluster computing is an attractive alternative to data processing on large parallel supercomputers
The simplest configuration of nodes for cluster computing consists of a mater node and slave nodes

Cluster Computing

Big Data

What does Big Data mean and how big is Big Data ?
Big Data is so big that it cannot be stored on the persistent storage devices attached to a single computer system
Big Data may also mean an infinite amount of data

Very frequently interpretation of a term "Big Data" depends on the sources of data. So, what are the sources of Big Data ?

The oldest sources include databases created and expanded since the first applications of the relational database technology.

The other source of data are generator of synthetic data commonly used for testing and performance evaluations. Yet another older source of data are data recordings from business processes.

More modern group of sources for BigData include sources like video documents , streams of messages coming from social websites and email servers and information from various

websites collected from Internet. Such group of sources is commonly called as enterprise and human generated data.

The largest group of sources of Big Data includes machine generated data. it includes the streams of data coming from sensor attached to industrial manufacturing

lines, satellite data obtained from record the signal coming from outer space, video data and images obtained from recording of real world events.

Machine generated data also include logs, audits, performance traces obtained from operational software systems.

Big Data

What are the sources of Big Data ?

Another way to characterize BigData is through so called nVs features. The original characterization of BigData used 3V features that included

Volume: e.g., billions of rows ? millions of columns, Variety: Complexity of data types and structures, Velocity: Speed of new data creation and growth.

Volume feature means that only very large volumes of data can be called as BigData. Variety feature means that Big Data must come in many different varieties

like text data, images, videos, semi-formatted data, etc. Velocity feature means that only the amounts of data growing and/or changing very fast can be

categorized as BigData.

Over time more V-like features have been added. The additional Vs include

Veracity: Ability to represent and process uncertain and imprecise data,

Value: Data is the driving force of the next-generate business and

Viability: Benefits we can potentially have from data analysis.

There are many, many other Vs, the largest number of Vs I found on Web was 42 ! (https://url.au.m.mimecastprotect.com/s/YCCKCXLWDDCzPonQUkswfW406_?domain=kdnuggets.com)

Vagueness: The meaning of found data is often very unclear, regardless of how much data is available

Validity: Rigor in analysis is essential for valid predictions where data is the driving force of the next-generate business

Vane: Data science can aid decision making by pointing in the correct direction

... and many, many others ... :)

Big Data

Big Data is characterized by so called 3V features:
- Volume: e.g., billions of rows ? millions of columns
- Variety: Complexity of data types and structures
- Velocity: Speed of new data creation and growth
Additional Vs:
- Veracity: Ability to represent and process uncertain and imprecise data
- Value: Data is the driving force of the next-generate business
- Viability: Benefits we can potentially have from data analysis
There are many, many other Vs, the largest number of Vs I found on Web was 42 !
- Vagueness: The meaning of found data is often very unclear, regardless of how much data is available
- Validity: Rigor in analysis is essential for valid predictions where data is the driving force of the next-generate business
- Vane: Data science can aid decision making by pointing in the correct direction
- ... and many, many others ... :)

This slide list the examples of BigData. It includes:

Examples of Big Data:

Clickstream data - data obtained from recording navigations of the customers over online electronic business web sites,

Call centre data - data obtained from recording calls performed by the employee of call centres,

E-mail and instant-messaging - data obtained from recording sequences of emails and messages from message passing systems.

Sensor data - data obtained from sensors recording the signals obtained from various sources,

Unstructured data - data that do not need to be consistent with a single pattern (schema), usually such data are obtained from unification of data objects from the slightly different domains within a single data container,

Geographic data - geographic data included many different types of maps and the descriptions of the real world objects marked on the maps, like roads, electricity networks, water pipelines, etc,

Satellite data - streams of data obtained from satellites,

Image data - all sorts of images, x-ray images, mri images, ultrasound images, etc

Temporal data - data obtained from "measurements" systematically repeated over a longer period of time, for example changing prices on stock exchnage,

and more ...

Big Data

Examples of Big Data:
- Clickstream data
- Call centre data
- E-mail and instant-messaging
- Sensor data
- Unstructured data
- Geographic data
- Satellite data
- Image data
- Temporal data
- and more ...

Cluster Computing

Outline

Computer Cluster
Big Data
Traditional Data Architectures
Meet Hadoop !
Big Data on Database Clusters
Big Data on Kubernetes

Traditional data architectures are based on data warehousing technologies. A data warehouse is a database that contains temporal data, i.e. data collected over

a longer period of time, for example orders submitted by the customers in the last twenty years.

Data warehouse is fed with data coming from the operational databases. An operational database keeps a snapshot of the present information. At the end of a given period of time, typically at the end of a day operational data are appended to data warehouse database. Relational database systems like DB/2, Oracle, MySQL etc

are typical used fro implementation of operational databases. It is possible that operational database contains data in Excel spreadsheet file or even in flat files

controlled by an operating system.

A specialized piece of software is used for data analysis.The outcomes of data analysis are visualized in a form of pie charts, block charts, or functional charts.

Traditional Data Architectures

Data warehousing technologies

Data warehousing technology has been developed over many years and at the moment it is a mature technology that has important strengths.

The strength of traditional data architectures include:

- Centralised governance of data repositories,

- Light-fast inquires performed regularly in daily business,

- Optimisation for OLTP (OnLine Transaction processing) and OLAP (On-Line Analytical Processing),

- Security and access control,

- Fault-Tolerance and backup.

Progresses in management of BigData impose new challenges on data warehousing technology. The challenges for traditional data architectures include:

- New types of data such as unstructured data and semi-structured data,

- Increasingly large amounts of data flowing into organisations,

- New computational paradigms use non-traditional NoSQL databases to rapidly mine and analyse very large data sets,

- Increasing cost of storing and analysing the large amounts of data,

- Increasing use of data analytics, which requires significant storage and processing capabilities.

Traditional Data Architectures

The strength of traditional data architectures:
- Centralised governance of data repositories
- Light-fast inquires performed regularly in daily business
- Optimisation for OLTP and OLAP
- Security and access control
- Fault-Tolerance and backup
The challenges for traditional data architectures:
- New types of data such as unstructured data and semi-structured data
- Increasingly large amounts of data flowing into organisations
- New computational paradigms use non-traditional NoSQL databases to rapidly mine and analyse very large data sets
- Increasing cost of storing and analysing the large amounts of data
- Increasing use of data analytics, which requires significant storage and processing capabilities

Traditional Data Architectures

A sample Data Lake architecture

Hardware for Big Data has two scalability dimensions. It means that it can be scaled-up in two different ways.

Assume that a computer used for processing of all our data provides us wth satisfactory performance. Such computer is represented by an icon located

close to a centre of the system of coordinates. Now, assume that we significantly increase the amounts of data processed and we also increased the total number of

applications processed our data. Now, due to more data and more applications a computer does not provide us with a satisfactory performance.

We have two options. The first option is to purchase a more powerful computer that will be able to deal with the increased workload. Such solution solves

the performance problems. However, it has two important disadvantages. First, a new and more powerful computer costs a lot of money. Second, we have a problem

what to do with the old computer ?

The second option is to purchase more computers like the computer we have now, use fast networks to link the systems and to create cluster of computers.

Such solution better scales up. When the current cluster of computer is not sufficient to process a given workload then we can easily add more systems to a cluster

increasing its processing power. There is no need to remove the earlier purchased systems.

Traditional Data Architectures

Hardware for Big Data has two scalability dimensions

Cluster Computing

Outline

Computer Cluster
Big Data
Traditional Data Architectures
Meet Hadoop !
Big Data on Database Clusters
Big Data on Kubernetes

Hadoop, in terms of its developers, is a project that develops open-source software for reliable, scalable, distributed computing. Hadoop runs on a cluster

computers and supports parallel processing of data.

Features of Hadoop

- Capability to handle large data sets, e.g. simple scalability and coordination,

- File size range from gigabytes to terabytes,

Can store millions of those files,

- High fault tolerance,

- Supports data replication,

- Supports streaming access to data,

- Supports batch processing,

- Supports interactive, iterative and stream processing,

- Implements a data consistency model of write-once-read-many access model,

- Runs on commodity hardware, not high-performance computers,

- Inexpensive,

- It can be deployed on premises or in the cloud,

Meet Hadoop !

Hadoop, in terms of its developers, is a project that develops open-source software for reliable, scalable, distributed computing
Features of Hadoop
- Capability to handle large data sets, e.g. simple scalability and coordination
- File size range from gigabytes to terabytes
- Can store millions of those files
- High fault tolerance
- Supports data replication
- Supports streaming access to data
- Supports batch processing
- Supports interactive, iterative and stream processing
- Implements a data consistency model of write-once-read-many access model
- Runs on commodity hardware, not high-performance computers
- Inexpensive
- It can be deployed on premises or in the cloud

Meet Hadoop !

Core components of Hadoop

Hadoop Ecosystem

Hadoop ecosystem

Commercial Hadoop Landscape

Commercial Hadoop landscape

Meet Hadoop !

Master-slave architecture of Hadoop clusters

Meet Hadoop !

Hadoop clusters can support up to 10,000 server and receives near-to-linear scalability in computing power
A typical Hadoop cluster consists of:
- A set of master nodes (servers) where the daemons supporting key Hadoop frame-works run
- A set of worker nodes that host the storage (HDFS) and computing (YARN) work
- One or more edge servers, which are used for accessing the Hadoop cluster to launch applications
- One or more relational databases such as MySQL for storing the metadata repositories
- Dedicated servers for special frameworks such as Kafka

Meet Hadoop !

Hadoop also supports the pseudo-distributed mode
- All HDFS and YARN daemons running on a single node.
- Highly simulates the full cluster
- Easy for beginner's practice
- Easy for testing and debug
Our lab setting is the pseudo-distributed mode
- The single node is Ubuntu 22.4 Virtual Machine (VM)

Cluster Computing

Outline

Computer Cluster
Big Data
Traditional Data Architectures
Meet Hadoop !
Big Data on Database Clusters
Big Data on Kubernetes

There are two other ways how BigData can be implemented on computer cluster: database clusters and Kubernetes.

A database cluster is a collection of databases that is managed by a single instance of a running database server. A very large database in a database cluster is partitioned over a number of smaller databases each located on a separate node of a computer cluster.

Database clustering requires replication and sharding. Database clustering improve performance, availability, and scalability

The classes of database system that allow for database clustering:

NoSQL systems: MongoDB, RavenDB, Cassandra, Amazon Aurora, ...

NewSQL systems: ClustrixDB, NuoDB, CockroachDB, Pivotal GemFire XD, Altibase, MemSQL, VoltDB, ...

Improved OldSQL systems: Oracle RAC, SQL Server (Windows server Failover Cluster), DB2 Cluster, PostgreSQL, MySQL Cluster, ...

Big Data on Database Clusters

A database cluster is a collection of databases that is managed by a single instance of a running database server
A very large database in a database cluster is partitioned over a number of smaller databases each located on a separate node of a computer cluster
Database clustering requires replication and sharding
Database clustering improve performance, availability, and scalability
The classes of database system that allow for database clustering:
- NoSQL systems: MongoDB, RavenDB, Cassandra, Amazon Aurora, ...
- NewSQL systems: ClustrixDB, NuoDB, CockroachDB, Pivotal GemFire XD, Altibase, MemSQL, VoltDB, ...
- Improved OldSQL systems: Oracle RAC, SQL Server (Windows server Failover Cluster), DB2 Cluster, PostgreSQL, MySQL Cluster, ...

Cluster Computing

Outline

Computer Cluster
Big Data
Traditional Data Architectures
Meet Hadoop !
Big Data on Database Clusters
Big Data on Kubernetes

Big Data on Kubernetes

Kubernetes (K8) is a container or microservice platform that orchestrates computing, networking, and storage infrastructure workloads
in a plain language Kubernetes is an orchestration platform to manage any containerized application
A Kubernetes cluster consists of a single master node and potentially multiple corresponding worker nodes
The benefits of Kubernetes:
- horizontal scaling,
- automated rollouts and rollbacks,
- service discovery and load balancing,
- storage orchestration,
- self healing,
- batch execution,
- automatic binpacking

References

White T., Hadoop The Definitive Guide: Storage and analysis at Internet scale, O'Reilly, 2015 (Available through UOW library)
Vohra D., Practical Hadoop ecosystem: a definitive guide to Hadoop-related frameworks and tools, Apress, 2016 (Available through UOW library)
Aven J., Hadoop in 24 Hours, SAMS Teach Yourself, SAMS 2017
Alapati S. R., Expert Hadoop Adiministration: Managing, Tuning, and Securing Spark, YARN and HDFS, Addison-Wesley 2017