A Reconciled Data Warehouse Layer based on CCNxA Reconciled Data Warehouse Layer based on CCNx...

A Reconciled Data Warehouse Layer based on CCNx

A Reconciled Data WarehouseLayer based on CCNx

M. Alexander

Vienna University of Technology

CCNxCon 2013PARC September 5, 2013


Table of Contents

Problem

Content Centric Networks

Architectural Proposition

Prototype Implementation

Demonstrator Environment

Extensions and Conclusion


Problem

Outline

I Very Large DB Data Warehousing PropertiesI Content Centric Networking Primitives Congruent

with Base Reconciled Data Warehouse LayerStorage Operations

I Basing a Distributed DWH on the CCNx ProtocolI Demonstrate Content Centric MySQL/MariaDB

Storage Engine Elements ImplementationI Motivate Broader Content Centric (Super) Distributed

Operations


Problem

Data Warehousing

I Central Data RepositoryI Variant: Distributed DWH

I Little NoSQL, Mostly RelationalI Data Retrieval: Typical SQLI 3 Layers

I Data Sources, Extract, Transform, Load (ETL)I Emphasis: Reconciliation LayerI DWH, Retrieval Layer

I Distributed Case: Physical Data Mart


Problem

Problem Statement

I Large Database Data WarehousingI Analytic WorkloadsI Required Data Locality

I Bringing the Code to the Data viable Alternative

I Geographically Dispersed Users and Data Provider

How do you Disseminate TB Daily Load Rates to ManyNon-Data Center Users Requiring Data Locality?



Content Centric NetworksA Networking Paradigm based on Named Data Entities

Addresses Observed Asymmetry of Internet Traffic FlowProducer-Consumer.

CCNs [13] and the CCNx stack [14] provide primitives thatcan be applied to the domain of data warehousing (DWH).I ForwardingI (IP) Transport Agnostic Communication InterfaceI Data StoreI Caching



Content Centric NetworkingProperties

I Forwarding via Data Namespace vs. NetworkAddress-based

I Publish/Subscribe on the Network LayersI Affinity to Delay Tolerant Networks, Content

Distribution Networks, Content Addressable Memoryand Dataflow

I Maps to Multicast, AnycastI Data Granularity: Chunks

I Streaming through Segmentation

I Caching Inherent in CCN Protocol Stack



Content Centric NetworkingProperties II

I Current Implementation Utilizes IP TransportI Interest Packets Trace Content Caching WeightsI Forwarding Tables: Added Pending Interest TableI Hierarchical (Logical Data) Topologies

I Source-Sink ConesI Additional Parameters Relative to TCP/IP

I Chunk Size, Cache Sizes, Request Rates, MissRatios, ...



Explorative Use Cases SurveyI Sequential Scan vs. Index-Based

I Indexes on all RelationsI Depends on Table Size, Width, Load Rate to Index

Build Cost, Results Size of Predicate Evaluation, ...I Query Plan Cost/Time Experiments Easy to Carry

Out using e.g. PostgreSQL EXPLAINI Intuitive Ranking of Regions

I Too Many Dimensions, Tablesize RangesI Distinct Regions with Indexing

I Local Index + Local DataI Local Index + Record ContentObjectsI Index Retrieved as ContentObject, Records as

I Distinct Regions without IndexingI Convergence on Very Large Analytics DB +

Dissemination Case



Use CaseEarth Observation Data Dissemination-Analytics

I Large Vector and Very Large Image Dataset SizesI Fifth Climate Model Intercomparison Project (CMIP)v 1PB

I ESA Earth Observation Data Sets

I High ETL Rates: 100s MB/dayI Suits Hierarchical Dissemination TopologyI Data Locality with Above Parameters Unsolved



Architectural PropositionTowards a Network-Provided DWH Infrastructure

I Towards DWH Operations on the NetworkI Objective of a (Super) Distributed DWHI Post Grid/Cloud Scientific Data

Dissemination/ProcessingI Shared-nothing with Explicit Lock Consistency ModelI Global Light-Weight State Holding/Mutex ServiceI Flat Node structure

I But of a Supernode for Query Planning(not in Demonstrator)

I Index Server (not in Demonstrator)I Analyic Segment Servers

I Sharding for Scale-Out



Architectural Proposition IITowards a Network-Provided DWH Infrastructure

I Proposed Architecture Supports both:I RelationalI Hierarchical

I Compare Google F1 RDBMS and Spanner Datastore[17]

I Clusters as Tree Branch Segments for Child Rows



TopologyTopology Options - What is the Optimal Topology for a Table-based CCDB?

I Intuitive Hypothesis as it to be Tree Form (Rooted,Directed Acyclic Graph Shaped)

I Covering Domain Vector and Image Datasets

I Hard to Test Proposition

I Simulation using Existing Toolchain Enables ParameterEstimation

I Less Discovery of Optimal Topologies

I Uses Constructed Skeleton Topology



SimulationI Regular Throughput, Latency, Hit Ratio Simulation, Tables

Repos Upstream/Downstream at Leafs Performed

I Throughput Shaping as Proxy for Node Utilization

I Possible to Craft Multi-RepositoryScenarios based on a-Priori Topology Choice

I Topology Choice Requires Alternate Simulation Approach



Alternate Optimal Topology SimulationProposalScenario Trees to Track Node Cache Changes over Time

I Proposed Ensemble Method following [9] and [8]

I Tree as Sample in Time of Stochastic ProgrammingProblem Vector

I Minimize Cost Along Path

I Procedure

I Construction of a sample of scenario trees: Monte Carloand CCN Simulation

I Solving Sample Trees using Stochastic Programming



Alternate Optimal Topology Simulation (2)

I Procedure contd.

I Fitting Policy Function using SVM-Machine Learning

I Ranking and Repeating

(ξ1, x1)

(ξ12 , x1

2 )

(ξ13 , x1

3 )

(ξ1t , x1

t )

(ξ23 , x2

3 )

(ξk4 , xk

4 )

(ξ32 , x3

2 )

(ξ33 , x3

3 )



PrototypeMariaDB/MySQL

I Storage Engines Extension Options

I MyISAM with Partitioning

I CSV: 3 Files per Table, no Indexes

I (M)Aria with Partitioning

I Archive: File per Table, Insert-Only, no Indexes

I InnoDB with Separate Table/Index Filesinnodb file per table



PrototypeCross-Layer DB Approaches

II Prior DB Storage Cross-Layer Approaches

I MySQL Falcon Engine: MySQL over ZFSI Management vs. Distributed DB Focus

I DB over DFS

I MySQL over Distributed {ZFS, GPFS, GluserFS, Lustre,PVFS2 etc.}

I Locking Issues, Use for Replication,TLOG etc.

I DB over Distributed Block Device

I MySQL over DRBD et al.



PrototypeCross-Layer DB Approaches

I Application with DB Protocol over CCN

I Hadoop with CCNx [16]

I NoSQL over CCN for Large Datasets

I Distributed Join Cavaet

I Possibly Implicitly Partitioned, Denormalized

I or Local Joins



PrototypeArchitecture

I MariaDB/MySQL Storage Engine

I Based on CSV-Engine (file-per-table)

I Engine Modification - New Native Commands

I Global Lock Service

I Modifications for Opening Tables

I Standard MySQL Behavior: Open/Cache TablespacesUpon First Operation on Open Database

I Extra Layer of CCDB Global Table Locks

I Explicit Write-back to CCNx from Local FS Store



PrototypeIndex-Based Access

I Applicable Regions:

I Low Change Velocity Tables

I SELECT Predication Results Sets est. << 10% RowNumbers

I Initially Examined: Index-Affine Content Object Retrievalbased on Local Index

I Depreciated in Favor of Very Large DB Use Case

I Where Indexing is Not Efficient vor Very Large Table Sizes,ETL RatesI Main Constraint: Index (Re-)Building Time



Prototype ImplementationDesign

I MariaDB Database 10.0.4 Fork

I Apache Zookeeper (formerly part of Hadoop)for Global State

I C/C++ CCDB Extensions, Mutex Library



Prototype ImplementationI Ongoing work on Adaption of MySQL/MariaDB CSV

Storage Engine

I Addition of a native MySQL Commands

I Simple Table-Based Granularity

I Fine Granularity with Multiple Supernodes: Partitioningand Sharding - Towards (Super) Distributed DBs

I Tablespaces

I Mix of - DB2 Terms - System and Database ManagedSpace (SMS/DMS) with Filesystem and CCNx



PrototypeArchitecture::Hierarchical Namespace

I Hierarchical Demonstrator

I Topology Agnostic but Distribution Tier-Aware

I Maps to CCNx URI Schema

I CCDB Schema

I ccnx:/ccvldb.org/cryosat2/{altimeter, doppler}/track[n]



PrototypeNew MariaDB/MySQL Commands - Partly Implemented

I CCDB PORT uri

I CCDB OPENTABLE uri

I SHOW CCTABLES

I SHOW CCDB PUBLISH STATUS

I SHOW CCDB STATUS uri

I CCDB PUBLISH uri

I CCDB LOCK CCTABLE uri

I CCDB UNLOCK CCTABLE uri



Prototype ImplementationFunctions Excerpt

char* ccdb_mutex_read(std::string uri);char* ccdb_mutex_check(std::string uri);std::string ccdb_mutex_lock(std::string uri,

std::string node, std::string user);char* ccdb_mutex_release(std::string uri);char* ccdb_publish_status(std::string uri);char* ccdb_show_status(std::string uri);char* ccdb_mutex_createlocknode(std::string uri);



Prototype ImplementationIn-Memory vector struct Mirroring Zookeper State

public:...std::vector<std::string> stVector;...struct ccdbState{

std::string uri;std::string published;std::string node;std::string user;

};std::vector<ccdbState> ccdbMutexes;



PrototypeLimitations

I No Indexes, Foreign Keys

I Sequential Scan Only

I Single Instance (compare Oracle)

I Single Schema (database.schema.table)

I No DB Owner (Namespace)



PrototypeWalkthrough Excerpts

I SQL SHOW ENGINES;




I SQL Check Publish Status (CCNx Repo)ccdb publish status ccvldb$ cryosat2 altimeter track1;




I SQL Open CCDB Tableccdb opentable ccvldb$ cryosat2 altimeter track1




I SQL Lock/Unlock Tableccdb lock cctable ccvldb$ cryosat2 altimeter track1




I SQL Check Status:SHOW CCDB STATUS



DemonstratorSatellite Remote Sensing Altimeter (Vector) Data

I Croysat-2::Synthetic Interferometric Altimeter(SIRAL)::Low Resolution Mode (LRM)

I Level 1 Processed

I Data File Split for Relational Join using Matlab Tool (seeAcknowledgements)C OFFL SIR FDM 1B 20130419T084613 20130419T084845 B001.DBL



Testbed Walkthrough 1I Open DB

I use cryosat2 altimeter;

I Check CCDB Status

I show ccdb status;

I Open CCDB Table

I ccdb opentable ccnx$ ccvldborg cryosat2 altimeter lrm;

I Lock CCDB Table

I ccdb lock cctable ccnx$ ccdb lock cctableccnx$ ccvldborg cryosat2 altimeter $



Testbed Walkthrough 2I Selects on Local Table doppler and CCDB Table lrm

I select * from lrm;

I select * from doppler;

I Join on them

I select lrm.a2, doppler.a3 from lrm, doppler wherelrm.i=doppler.i;

I Publish Table back to the Network

I ccdb publish ccnx$ ccvldborg cryosat2 altimeter lrm;

I Check Publish Status

I ccdb publish statusccnx$ ccvldborg cryosat2 altimeter lrm;



Possible ExtensionsDomain Use Case Served Well with Table Granularity

I Lower Granularity

I Make Clever Use of Segmentation & Seeks?

I Compare Legacy ISAM Storage Access Method

I Sharding

I Optimistic Concurrency Control



Conclusion

I Proposal of a Large-Data DWH Architecture based onContent Centric Networking

I Fits OLAP

I Wide-Area Data Distribution

I Explicit Read Locality, Write Concurrency Control

I MariaDB/MySQL Integrated Demonstrator



Acknowledgements

I Special thanks to Herbert Findeis

I Salvatore Dinardo (ESA) - Cryosat Matlab ReaderPackage



References[]

Alexander Afanasyev.

NS-3 based named data networking (NDN)simulator, March 2013.

Alexander Afanasyev, Ilya Moiseenko, and LixiaZhang.

ndnSIM: NDN simulator for NS-3.

Technical Report NDN-0005, NDN, October 2012.

Vikas Agrawal, Udayan Nandkeolyar, P.S.Syndararaghavan, and Mesbah Ahmed.



Simulation model and analysis of a datawarehouse.

15(3):31–45, 2006.

Marcos K. Aguilera, Wojciech Golab, and Mehul A.Shah.

A practical scalable distributed b-tree.

Proc. VLDB Endow., 1(1):598609, August 2008.

Charles Bell.

Expert MySQL.

Expert’s voice in databases. Apress, 2012.

Dan Boneh, Amit Sahai, and Brent Waters.



Functional encryption: a new vision for public-keycryptography.

Commun. ACM, 55(11):5664, November 2012.

Jerome Darmont, Fadila Bentayeb, and OmarBoussaid.

DWEB: a data warehouse engineering benchmark.

CoRR, abs/0705.1453, 2007.

B. Defourny, D. Ernst, and L. Wehenkel.

Scenario trees and policy selection for multistagestochastic programming using machine learning.

ArXiv e-prints, December 2011.



Boris Defourny, Ernst Damien, and LouisWehenkel.

forthcoming: Scenario trees and policy selectionfor multistage stochastic programming usingmachine learning.

INFORMS Journal on Computing, 2012.

Mark Gritter and David R. Cheriton.

An architecture for content routing support in theinternet.

In Proceedings of the 3rd conference on USENIXSymposium on Internet Technologies andSystems - Volume 3, USITS’01, page 44, Berkeley,CA, USA, 2001. USENIX Association.



Patrick Hunt, Mahadev Konar, Flavio P. Junqueira,and Benjamin Reed.

ZooKeeper: wait-free coordination forinternet-scale systems.

In Proceedings of the 2010 USENIX conference onUSENIX annual technical conference,USENIXATC’10, page 1111, Berkeley, CA, USA,2010. USENIX Association.

Van Jacobson, Marc Mosko, Diana Smetters, andJ.J. Garcia-Luna-Aceves.

Content-centric networking.

Whitepaper, Palo Alto Research Center, January2007.



Van Jacobson, Diana K. Smetters, James D.Thornton, Michael F. Plass, Nicholas H. Briggs,and Rebecca L. Braynard.

Networking named content.

In Proceedings of the 5th international conferenceon Emerging networking experiments andtechnologies, CoNEXT ’09, page 112, New York,NY, USA, 2009. ACM.

{Palo Alto Research Center}.

CCNx reference implementation.

Technical report, PARC, Palo Alto, 2013.

Diego Perino and Matteo Varvello.



A reality check for content centric networking.

In Proceedings of the ACM SIGCOMM workshopon Information-centric networking, ICN ’11, page4449, New York, NY, USA, 2011. ACM.

John Sherwood.

An implementation of content-centric networkingsocket for use with hadoop.

Technical report, Stetson University, 2011.

Jeff Shute, Mircea Oancea, Stephan Ellner, BenHandy, Eric Rollins, Bart Samwel, RadekVingralek, Chad Whipkey, Xin Chen, BeatJegerlehner, Kyle Littleeld, and Phoenix Tong.



F1 - the fault-tolerant distributed RDBMSsupporting google’s ad business.

In SIGMOD, 2012.

Talk given at SIGMOD 2012.

A Reconciled Data Warehouse Layer based on CCNxA Reconciled Data Warehouse Layer based on CCNx...

Documents

Transcript of A Reconciled Data Warehouse Layer based on CCNxA Reconciled Data Warehouse Layer based on CCNx...