in-memory computing essentials for architects and engineers

© 2017 GridGain Systems, Inc.

In-Memory Performance

Durability of Disk


In-Memory Computing Essentials

for Java Developers

Denis MagdaIgnite PMC Chair

GridGain Director of Product Management


• Apache Ignite Overview

• Clustering and Deployment

• Distributed Storage

• Distributed SQL

• Distributed Computations

• Machine Learning

• Memory Architecture & Persistence

Agenda


Apache Ignite In-Memory Computing Platform

Memory-Centric Storage

Ignite Native Persistence(Flash, SSD, Intel 3D XPoint)

Third-Party Persistence(RDBMS, HDFS, NoSQL)

SQL Transactions Compute Services MLStreamingKey/Value

IoTFinancialServices

Pharma &Healthcare

E-CommerceTravel & Logistics

Telco


Clustering and Deployment


Clustering

• Server Nodes

• Act as containers for data and computations

• Generally started as standalone processes

• Client Nodes

• Provide a cluster entry point to run operations

• Embedded in applications code


Deployment

• Nodes are logical entities

• Runs in a JVM process

• Many nodes in a single JVM process

• On-Premise and Cloud

• Physical server or VM

• AWS, Azure, Google Compute Engine

• Kubernetes, Mesos, YARN


Distributed Storage


Distributed Storage

JCache Transactions Compute SQL

RDBMS

NoSQL

HDFS

Server Node

Distributed Key-Value Store

Dynamic

Scaling

Distributed

partitioned

hash map

ACID TransactionJCache & SQL

Server Node Server Node

3rd party storage caching

DURABLE MEMORY DURABLE MEMORY DURABLE MEMORY


Where Entry Goes?

Ignite Node 1 Ignite Node 2

put (key, value)

? ?


Key to Node Mapping

Key Partition

Server Node

ON-DISK


Caches and Partitions

K1, V1

K2, V2

K3, V3

K4, V4

Partition 1

K5, V5

K6, V6

K7,V7

K8, V8 K9, V9

Partition 2

Cache


Partitions Distribution


0 2 4 6 8

10 12 14

1 3 5 7 9

11 13 15


Where Entry Goes?


put (key, value)

0 2 4 1 3 5

? ?


Where Entry Goes?


put (key, value)

0 2 4 1 3 5


Backup Copies

Ignite Node Ignite Node


0 1

2 3


Backup Copies



0 1

2 3

0

1

2

3


Distributed SQL


Distributed SQL

JDBC ODBC SQL API

Java .NET C++ BI

SELECT, UPDATE,

INSERT, MERGE,

DELETE, CREATE

and ALTER

DDL, DML Support

Cross-platform

Compatibility

Indexes in

RAM or Disk

Dynamic

Scaling

Server Node Server NodeServer Node

Apache Ignite Cluster


Tools


Connectivity

• JDBC

• ODBC

• REST

• Java, .NET and C++ APIs

// Register JDBC driver.

Class.forName("org.apache.ignite.IgniteJdbcThinDriver");

// Open the JDBC connection.

Connection conn = DriverManager.getConnection("jdbc:ignite:thin://192.168.0.50");

./sqlline.sh --color=true --verbose=true -u jdbc:ignite:thin://127.0.0.1/


Data Definition Language

• CREATE/DROP TABLE

• CREATE/DROP INDEX

• ALTER TABLE

• Changes Durability

• Ignite Native Persistence

CREATE TABLE `city` (

`ID` INT(11),

`Name` CHAR(35),

`CountryCode` CHAR(3),

`District` CHAR(20),

`Population` INT(11),

PRIMARY KEY (`ID`, `CountryCode`)

) WITH "template=partitioned, backups=1, affinityKey=CountryCode";


Data Manipulation Language

• ANSI-99 specification

• Fault-tolerant and consistent

• INSERT, UPDATE, DELETE

• SELECT

• JOINs

• Subqueries

SELECT country.name, city.name, MAX(city.population) as max_pop

FROM country JOIN city ON city.countrycode = country.code

WHERE country.code IN ('USA','RUS','CHN')

GROUP BY country.name, city.name ORDER BY max_pop DESC LIMIT 3;


Affinity Collocation

Country

Languag

eCity

Server Node

ON-DISK

Server Node

ON-DISK

key (country = 5) 10

Partition

key (cityId = 10, countryId = 5)10

Partition

key (cityId = 11, countryId = 9) 12

Partition


Collocated Joins

1. Initial Query

2. Query execution over local data

3. Reduce multiple results in one

Ignite Node

Canada

Toronto

Ottawa

Montreal

Calgary

Ignite Node

IndiaMumbai

New Delhi

1 SELECT ct.name, c.name

FROM Country as ct

JOIN City as c ON ct.id = c.countryId

WHERE ct.name = “Canada”;

2

23


Non-Collocated Joins

1. Initial Query

2. Query execution (local + remote data)

3. Potential data movement

4. Reduce multiple results in one

Ignite Node

Canad

a

Toronto

Calgary

1 SELECT ct.name, c.name

FROM Country as ct

JOIN City as c ON ct.id = c.countryId

WHERE ct.name = “Canada”;

2

24 Ignite Node

India

Montreal

Ottawa

3Montreal

Ottawa

Mumbai

New Delhi


Distributed Computations


Compute Grid

DURABLE MEMORY

DURABLE MEMORY

Ignite Cluster

C1

R1

C2

R2

C = C1 + C2

R = R1 + R2

C = Compute

R = Result

in T/2 time

Automatic Failover

Load Balancing

Zero Deployment


1. Initial Request

2. Fetch data from remote

nodes

3. Process entire data-set

3

1

Data 12

2 Data 2

Client-Server Processing Co-located Processing

Server Node

ON-DISK

Server Node

ON-DISK

1. Initial Request

2. Co-located processing with

data

3. Reduce multiple results in

one

2

2

1Client Node

Server Node

ON-DISK

Server Node

ON-DISK

Client Node

3


Machine Learning


Genetic Algorithm Grid

DURABLE MEMORY

DURABLE MEMORY

Ignite Cluster

F2, C2, M2

F = F1 + F2

C = C1 + C2

Collocated

Computation

Biological Evolution

SimulationChromosome and Genes Cluster

M = M1 + M2

F1, C1, M1

F = Fitness Calculation

C = Crossover

M = Mutation


Machine Learning Grid

K-Means Regressions Decision Trees

R C++ Python Java

Server Node Server NodeServer Node

Distributed Core Algebra


Scala REST

Random ForestDistributed Algorithms

Dense and Sparse

Algebra

Large Scale

Parallelization

Multi-Language

Support

Dense and Sparse

Algebra

No ETL


Memory Architecture & Persistence


Durable Memory

Off-heap Removes

noticeable GC

pauses

Automatic

Defragmentation

Stores

Superset of

Data

Predictable memory

consumption

Fully Transactional

(Write-Ahead Log)


Server Node Server Node Server Node

Ignite Cluster

Instantaneous

Restarts


Regions and Segments

• Memory split into regions

• Regions split into segments

• Segments include pages


B+Tree

• Self-balancing tree

• Memory & Disk

• Sorted Index

• Secondary Indexes

• Hash Index

• Primary Keys

• Hash code based sorting


Free Lists

• Tracks pages of ~ equal free space

• 25% free

• 75% free

• Essential for updates

• Gives page with min size needed

• Reduces fragmentation

• Lowers pages compaction activity


Ignite Native Persistence

1. Update

RA

M

2. Persist

Write-Ahead Log

Partition File 1

3. Ack

4. Checkpointing

Partition File N

Server Node


Any Questions?

Thank you for joining us. Follow the conversation.

http://ignite.apache.org

#apacheignite

#denismagda

in-memory computing essentials for architects and engineers

Engineering