Department of Computer Science | Institute of Systems Architecture | Chair of Computer Networks

From Research To Transfer:User-Controllable Cloud Storage

14th TERENA TF-Storage MeetingFebruary 10-11, 2014, Zurich, Switzerland

Josef Spillner

mailto:josef.spillner@tu-dresden.de xmpp:josef.spillner@jabber.org

The Trouble with Cloud Storage

3 months limit;tariff choice

utility billing

backup/syncvs. storage

web accessonly

proprietary client

file size limit

institutional

pricing:progressive/degressive

trust?

quality?

Taming the Trouble: RAICs, RAOCs

controller/gateway

storage targets: services, devicesservice propertiesservice bundles

auto-selectionauto-configurationauto-repair

integrationschedulingoptimisation

proxy cloud

extensionbackupsyncsharing...

The Cloud Storage Lab @ TU Dresden

prototypicalsoftware

developmentresearch

experimentsand pilot

installationstransfer

widespread productiondeployments

Short history:

● 2010 ideas and service modelling

● 2011 NubiSave storage controller started UCC paper on Cloud storage controllers

● 2012 Usable software, Debian packaging

● 2013 NubiVis; π-Box personal Cloud VM FGCS article on optimal storage SUCRE summer school tutorial at KIT

● 2014 NubiGate VM

http://lab.nubisave.org/

Research

Storage Service Ecosystem

Excursus: Service Science-> Bundles -> RAOC

domain 1

domain 2domain

Cloud Storage

service interface

functionality non-functionalproperties

nfp 1nfp 2

nfpbackup

super-domainInfrastructure

sub-domainFile Storage

protocol description

service instance

service hosting

client

consumer

provider

service impl

engineer

SLAsmessages

Storage Service Descriptions

webService GoogleStorageimportsOntology { _"urn:ontology:conqo/CloudQoS.wsml#" }

capability ServiceCapability postcondition definedBy ?serviceType memberOf cloud#CloudStorage .

instance PricePerData memberOf { cloud#PricePerData, qos#ServiceSpec }

qos#value hasValue 0.17 qos#unit hasValue qos#Euro

concept MeasurementUnit conversionFactor impliesType _double

concept SpaceUnit subConceptOf MeasurementUnit

instance GB memberOf SpaceUnit conversionFactor hasValue 1024.0

Base Ontology (WSML)

Instance Ontology

Storage Service Bundles

Servicebundle

$$

xa

(k+m) erasure coding secret sharing

s1

s2

s3

s4

s5

s6

s1

s2

s3

s4

s5

s6

replication s1

Storage Service Markets

Storage Flows

Storage Targets Data Flow Integration

localdirectory

USB stick

NFSexport

WebDAV

Commercial(Dropbox)

localdirectory

CIFSproxy

Splitter/Multiplexer

Filemodifier

(enc,dedup)

blockmodifier

(compress)

blockmodifier(stego)

Splitter/Multiplexer

Splitter/Multiplexer

Splitter/Multiplexer web

application

sink sourcetransport

filemodifier

(enc)

Reconfigurable Storage Flows

... aliasSoftware-DefinedStorage

Modules (i.e.FUSE modules):- Splitter (1:n)- Modifiers (1:1) * encryption * compression * ...- Transports * Directory * CIFS, NFS * WebDAV * SSHfs * S3fs * ... * CloudFusion- SDS propagation

Multi-User Storage Gateways

FlexDDPL: Flexible Data Distribution Policy Language

scopes → applied in contexts (e.g. user database)~user !negation T:22:00­06:00@group fragment mime:message/rfc822

rules → applied to targets (e.g. storage provider)store control adapt

Feature: Algorithms Selection

Background● secret sharing● forward error correction / replication

Striping (→ RAID0)● capacity 1.0, safety 1.0, performance 2.0Mirroring (→ RAID1)● capacity 0.5, safety 2.0, performance 1.0XOR parity (→ RAID5, RAID6)● 4 disks: capacity 0.75, safety 1.33, performance 1.0

Erasure coding (→ RAIDn)● n storage targets, k significant, m redundant: capacity k/m, safety 1+m/k● (Rotated) Reed-Solomon:

maximum-[Hamming-]distance separable (MDS) property● Cauchy-Reed-Solomon, Vandermonde-Reed-Solomon:

practically secure; CRS faster due to XOR operations compared to GF(2w)

AONT: information-theoretically secureAONT-RS: blend of AONT with (C)RS; alternative: encrypted fragments

Feature: Nested Contexts

context »normal«

context »private«

Configuration Data

Feature: Weights

Assigning weights to the storage targets● to fully utilise the capacity● to exploit faster upload connections

Feature: Recursion

Saving the database which contains all file and fragment metadata...● to avoid single point of failure● to allow for selective sharing

Feature: Streaming (1:x)

Continuous operation during read or write of large file● buffers on controller device can be kept small● improved performance through parallel coding and transmission

Feature: Batching (x:1)

0 1000 2000 3000 4000 5000 6000 7000 8000 90000

10000

20000

30000

40000

50000

60000

70000

80000

90000

100000

Encoding speed of different batch sizes

15 nodes

M=0

M=4

M=8

M=12

Filechunk size [KB]

En

co

din

g [

MB

s]

Experiments

peaCS: Performance and EfficiencyAnalsis for Cloud Storage

Nubisave Read Performance - 1MB File - UseAllInParallel

Read in KB/s

12345678910

Storages

010

2030

4050

6070

8090

100

Redundancy

4000

6000

8000

10000

12000

14000

5000600070008000900010000110001200013000

Experiment WeekFeb 7-14 2014

Autonomous Storage Clients

Scenario:A robot wants to store some data...

Labcast available:http://www.youtube.com/watch?v=NDlN0fG9Okk

Autonomous Storage Clients

Scenario:Pervasive storage on all devices

Credentials may be auto generated

Software

Comparison of Storage Controllers

Prototype Origin Authors Approach Integration PublicAvailability

Dependable Sky ('11)

T-Clouds,Univ. of Lisbon

Bessani, Corr-eia, Quaresma, André, Sousa

data units, byzantine fail, Rabin erasure

application library

no

('11) HPI,Potsdam

Maxim Schnja-kin, Christoph Meinel

NFPs, propr. platform, Libe-ration erasure

web interface no

TrustedSafe ('10)

eGovCD, Berlin/FhG FOKUS

(no publication) n fragments: one on USB, others remote

Windows plugin no

CloudShredder ('11)

Chinese Academy of Sciences

Nan Zang, Jiwu Jing, Peng Liu

Two fragments: one local, one remote

(not yet known) no

SecCSIE ('11) T-Systems MMS, Dresden

Ronny Seiger, Stephan Groß

Cachy Reed-Solomon erasure

CIFS proxy no

RACS ('10) CornellUniversity

Abu-Libdeh, Princehouse,Weatherspoon

erasure coding: Zfec Reed-Solomon

S3 proxy, can be distributed w/ Zookeeper

yes (www.cs.cornell.edu)

plus numerous commercial implementations:NetApp StorageGRID, Cleversafe, Trusted Safe etc.

since 2012, community projects:Tahoe-LAFS externals, TCD,OpenStack/Intel - PyEC2Lib

OpenStack Intermezzo

NubiSave Cloud Storage Controller

Installation of NubiSave

● Git repository● git://nubisave.org/git/nubisave

● Debian packages● http://nubisave.org/packages/

● Integrated into π-Box or NubiGate VMs● http://nubisave.org/downloads/

NubiSave Storage Flow Editor

NubiSave Configuration

NubiSave Configuration

NubiSave Tools

Controller operation

$ nubisave [<instance>]

Master script which starts bothof the below combined.

$ nubisave headless

Starts a new instance of thesplitter/dispersion file system.

$ nubisave gui

Starts the storage flow editor withstorage integration configuration.

$ nubisave stop

Stops the splitter.

Storages and database

$ nubisave-status

Lists all splitter instances andstorage targets attached to themor as part of any storage flow.

$ nubisave-mounter [<module>]

Mounts all splitters and/or storagetargets.

$ nubisave-unmounter [<module>]

Unmounts; inverse of the above.

$ nubisave-database [...]

Inspection of storage metadata.

Splitter-NG Framework

Plugins: Jerasure, JSharing, RAID-1, more coming...

Transport Module: CloudFusion

Additional Implementations

NubiVis: Distributed Data Visualisation

NubiVis Map View

NubiGate VM: Everything in one Box

NubiGate VM: OwnCloud Frontend

Our instance „RN-Dropbox“141.76.41.232/owncloud

Your instance...?

141.76.41.206 141.76.41.207

Future Vision

Future Work

What's next in our lab

● Software improvements● Configuration through the web interface (student thesis)● Git-based service registry query

● Research: MapReduce-like processing over dispersed data

● Scaling large with more users and storage providers, HA setups etc.

● Federation between research institutions:reciprocative/collaborative storage resource sharing <-- role of NRENs

Publications

[SS13] Josef Spillner, Alexander Schill:Orchestration of Distributed Storage Targets through Storage Flows.5th IEEE International Conference on Cloud Computing Technology and Science(CloudCom), Bristol, UK, December 2013.

[SQ+13] Josef Spillner, Maximilian Quellmalz, Martin Friedrich, Alexander Schill:peaCS - Performance and Efficiency Analysis for Cloud Storage.Workshop on Cloud Storage Optimisation (CLOUSO) @ 2nd ESOCC, Málaga, Spain, September 2013.

[SMS13] Josef Spillner, Johannes Müller, Alexander Schill: Creating Optimal Cloud Storage Systems. Elsevier Future Generation Computer Systems (FGCS), Issue 29(4), p. 1062-1072, June 2013. DOI 10.1016/j.future.2012.06.004.

[SS12] Josef Spillner, Alexander Schill: Flexible Data Distribution Policy Language and Gateway Architecture. 1st IEEE Latin American Conference on Cloud Computing and Communications (LatinCloud), Porto Alegre, Rio Grande do Sul, Brazil, November 2012.

[SP+12] Josef Spillner, Christian Piechnick, Claas Wilke, Uwe Aßmann, Alexander Schill: Autonomous Participation in Cloud Services. 2nd International Workshop on Intelligent Techniques and Architectures for Autonomic Clouds (ITAAC) @ 5th IEEE/ACM UCC, Chicago, Illinois, USA, November 2012.

[SG+11] Josef Spillner, Gerd Bombach, Steffen Matthischke, Johannes Müller, Rico Tzschichholz, Alexander Schill: Information Dispersion over Redundant Arrays of Optimal Cloud Storage for Desktop Users. 4th IEEE/ACM International Conference on Utility and Cloud Computing (UCC), Melbourne, Australia, December 2011. [Best Paper Award]