Download

ORCA-BEN: a distributed control 

framework for GENI  

- A Inter-Cloud IaaS Provisioning System for Scientific Workflows   

Yufeng Xin yxin@renci.org!

The Project Team! 

Co-‐PI: Jeffrey Chase 

Duke University 

Students: 

Prateek Jaipuria 

Muzhi Zhao 

PI: Ilia Baldine 

RENCI, UNC-‐Chapel 

Hill 

Anirban Mandal 

GENI Controller, 

applica>on 

interface, actor 

registry 

Aydan Yumerefendi 

Core code 

enhancements 

Chris Heermann 

BEN Testbed 

opera>ons, 

interconnec>ons with 

na>onal/interna>onal 

networks 

Staff: 

Victor Orlikowski 

Sincere thanks to our sponsors: 

GPO/GENI, ASCR/DOE, SDCI/NSF 

2!

Background! 

• ORCA-BEN: one of several GENI Control Framework Candidates" 

– Lead of ʻCluster Dʼ of GENI projects" 

• Project goal: extend ORCA (Open Resource Control Architecture) to" 

– Cross cloud provisioning: connections, clusters, and topologies " 

– Provision multi-layered network domains (BEN, NLR Sherpa, I2 ION)" 

– Leveraging Semantic Web technologies" 

– Interoperate with other CF candidates (XMLRPC controller)" 

• Connectivity" 

• API and Resource description translations" 

• Big Idea: create a variety of application-specific topologies out of 

collection of independent cloud providers connected by one or more 

multi-layered network providers" 

– GENI is a kind of IaaS." 

– 40GB/100GB network will bring everything closer." 

– Other uses: large-scale scientific applications on clouds interconnected by 

dynamic high speed networks" 

3!

Outline! 

• ORCA Architecture and Implementation! 

– Distributed actors and lease-based operation" 

• Substrate Stitching" 

• Semantic resource representations" 

– Substrate components and interfaces" 

• BEN and beyond" 

• NEuca + ImageProxy" 

• Interoperability" 

– Connectivity" 

– Resource description translations" 

• Scientific Application Support" 

– Wide area Hadoop based applications" 

– Workflows"

ORCA-BEN Overview 

Goal: provision built-‐to-‐order slices of networked virtual infrastructure from 

mul>ple cloud or network providers, dynamically on the fly. 

AM 

AM 

SM 

Brokering Services 

Image/Appliance Services 

Slide courtesy of Jeff Chase

ORCA Design and Implementation! 

• Pervasive virtualiza>on 

• Brokering 

• Pluggable 

– Policy 

– Controller 

– Interface 

• Cloud technologies 

• Seman>c web resource representa>on 

• Federated ID management

NEUCA 

NEuca – networked extension to 

Eucalyptus! 

Provide publicly routable 

IP with port number(s) 

to reach VM 

Donwload/register 

user images at multiple 

sites 

DNAT PROXY 

IMAGE PROXY 

Instantiate VM 

topologies 

EUCALYPTUS 

ORCA 

HEAD NODE 

WORKER 

WORKER 

WORKER 

WORKER 

7!

Eucalyptus/XCat ImageProxy! 

Map URL to EMI 

IMAGE PROXY 

ORCA SITE 

AUTHORITY 

EUCALYPTUS 

Register 

Download 

Remotely hosted 

image 

Download 

OS Image 

Map URL to EMI 

IMAGE PROXY 

ORCA SITE 

AUTHORITY 

Register 

EUCALYPTUS 

8!

BEN Topology and Inter-Domain 

Connectivity! 

6509 

ATDNet, MAX, 

BOSSnet 

Infinera 

DTN 

I2 ION 

Dark fiber to 

McLean, VA 

Duke 

Polatis 

Cisco 

RTP 

Cisco Research 

Wave 

C-Wave 

L3 PoP, 

Raleigh 

RENCI 

NLR FrameNet 

Cisco 

Research 

Wave 

FrameNet 

I2 

NLR 

7606 

Polatis 

OXC 

To BEN 

10Gbps 

UNC-CH 

NCSU 

RENCI 

6509 

StarLight 

C-wave 

6506 

NCNI BEN Fiber 

UNC/RENCI BEN Fiber 

C-Wave 

Potential connection 

National/International 

research networks 

Existing connection 

Presentation title goes here" 9!

Distributed Stitching in ORCA! 

• Stitching is the process of connecting resources of different types of 

belonging to different providers to each other" 

• Stitching involves a process of agreeing on labels between two 

parties" 

– Labels may denote bandwidth (VLAN tag, MPLS label, DWDM wavelength)" 

• Can be done through multi-step negotiation" 

– Complex protocols may be required" 

• ORCA approach is based on analyzing dependencies of providers 

(label producer vs. label consumer; can you translate labels?)" 

– A DAG of dependencies is formed based on the requested topology" 

– Properties carrying labels are passed from the root(s) of the DAG along the 

branches " 

– Certain resources wait to be instantiated until the label information needed 

to stitch them is available" 

10!

Outline! 

• ORCA Architecture" 



– Substrate Stitching" 

– Semantic resource representations! 

• Semantic graph ! 

• Substrate: NDL-OWL, Cloud! 

• Domain: Abstraction! 

• Request: Cluster, connection, topology! 






– Workflows" 

11!

Outline! 

• ORCA Architecture and Implementation! 


• Substrate Stitching" 

• Semantic resource representations" 


• BEN and beyond" 

• NEuca + ImageProxy" 



– Resource description translations"

Substrate Resource Description! 

• Multilayer Network description" 

– Semantic graph: dynamical model with new RDF 

resources and properties (xc, virtual interface, 

connection hierarchy, reservations)" 

– OWL vs. RDF" 

– DTN network description" 

– Collection (bag, set, list), labelSet, etc" 

• Cloud resource description" 

– Virtualization as an adaptation" 

– Substrate hosting capability" 

– Shortcut to classified type" 


Compute.owl! 

 

MacOS 

Note: Features mostly represent known constants. 

 

Feature 

1 

 

Vendor 

foaf:xxx() 

 

Debian 

 

Microsoft 

 

LinuxFileSystem 

 

ISO9660 

 

 

0..1 

 

 

Linux 

OS 

VServer 

vendor 

VirtualBox 

osDistribution(xsd:String) 

 

 

 

 

osVersion(xsd:String) 

LinuxRamdisk 

RAMType 

Xen 

ImageType 

 

DebianLenny 

clockFrequency 

 

 

 

version(xsd:String) 

(xsd:Float) 

OpenVZ 

VMM 

KVM 

 

1 

osDistribution(Lenny) 

 

LinuxKernel 

osVersion(5.0.8) 

Windows 

1 

1 

 

vendor(url:Debian) 

None 

 

 

'None' virtualization means bare metal 

 

 

DDRRam 

VMWare 

ImageProxyXML 

OVF 

 

 

Windows7 

Note: this goes to storage ontology 

 

 

Laptop 

osDistribution 

DDR1066 

DDR266 

 

 

(Windows7) 

 

clockFrequency clockFrequency 

RawStorageVolume 

ComputeElementType 

osVersion(7) 

RackedServer 

(1.066) 

(0.266) 

 

vendor(url:Microsoft) 

storageCapacity 

RamDisk 

(xsd:Integer) 

 

 

Blade 

 

DDR2-1066 

 

MoteType 

 

clockFrequency 

 

 

 

 

SCSI146G 

TelosB 

(1.066) 

IDE40G 

ARM2 

ARM6 

AtmelAVR8 

IDEDisk 

SCSIDisk 



(146GB) 

 

 

(40GB) 

 

 

IMote 

XSM 

 

 

ComputeElementComponent represent 

SATADisk 

VirtualDisk 

x86 

CPUArchitecture 1 

quantifiable elements of a running compute system 

 

1 

SATA250G 

Semantic shortcut hierarchy for well-known server types. 

 

 

 

For convenience, NDL-OWL declares these as both 

 

ComputeElementComponent 


FlashDisk 

SSD 

classes and individuals. Individuals are used in substrate descriptions, 

x86-64 

classes are used in requests and manifests. 

(250GB) 

IA-64 

We expect the RSpecs to support these by name. 

architecture 

 

 

 

 

architecture vendor 

os 

 

EC2CPUCore 

0..1 0..1 

0..1 ramType storageVolume 

StarGateCE 

LaptopKanseiCE 

EC2C1XLarge 

EC2M1XLarge 

0..1 

0..1 0..1 0..1 

vmm(url:None) 

vmm(url:None) 

vmm(url:Xen) 

vmm(url:Xen) 

clockFrequency(1.2) 

 

 

 

imageType 

cpu(url:PXA55) 

cpu(url:IntelCeleromM) cpu(url:EC2CPUCore) cpu(url:EC2CPUCore) 

vendor(url:AMD) 

CPU 

Memory 

Storage 

DiskImage 

memoryCapacity(64MB) numCPUCores(1) memoryCapacity(7000) memoryCapacity(15000) 

family("OPTERON") 

storageCapacity(32MB) memoryCapacity 

numCPUCores(8) 


numCPUCores(xsd:Integer) memoryCapacity storageCapacity imageLocation(xsd:URL) 

architecture(url:x86-64) 

(200MB) 

architecture(url:x86-64) architecture(url:x86-64) 

cpuFamily(xsd:String) 

(xsd:Integer) 

(xsd:Integer) 

imageSize(xsd:Integer) 

storageCapacity(80GB) storageCapacity 


cpuModel(xsd:String) hasUnit(url:Unit) 

1 

hasUnit(url:Unit) 

1..* 

(1690GB) 

(1690GB) 

cpuStepping(xsd:string) 

ioSpeed(xsd:Integer) 

 

clockFrequency(xsd:Float) 

raidLevel(xsd:Integer) 

QuadNehalem24 

numVolumes 1..* 

1..* 

 

(xsd:Integer) 

XSMCE 

clockFrequency(2.4) 

 

vendor(url:Intel) 

vmm(url:None) 

EC2M1Large 

Note: this goes to storage ontology 

family("Xeon") 

cpu(url:ATmega128L) 

 

Note: for systems including heterogeneous cores 

vmm 

vmm(url:Xen) 

cpuModel("Nehalem") 

you can add several CPUInfo instances 

memoryCapacity(4kB) 

ClassifiedComputeElement 

storage 

diskImage 

cpu(url:EC2CPUCore) 


describing different CPU pools 

cpu 

memory 

storageCapacity(512kB) 

memoryCapacity(7500) 



 


 

 

virtualize 

storageCapacity(850GB) 

computeElementType 

TelosBCE 

Xeon3 

ATmega128L 

0..1 0..1 0..1 

vmm(url:None) 

 

 

clockFrequency(3.0) clockFrequency(0.016) 

1..* 

cpu(url:MSP430) 

NetworkElement 

ComputeElement 

 

 

vendor(url:Intel) 

vendor(url:Atmel) 


 

EC2M1Small 

0..1 

EC2C1Medium 

family("Xeon") 

architecture 

memoryCapacity(xsd:Integer) 


PlanetLabCE 

architecture(url:x86-64) (url:AtmelAVR8) 

numSockets(xsd:Integer) 

(1000kB) 

vmm(url:Xen) 

vmm(url:Xen) 

vmm(url:VServer) 

numCPUCores(1) numCPUCores(1) 

storageCapacity(xsd:Integer) 

cpu(url:EC2CPUCore) cpu(url:EC2CPUCore) 

imageLocation(xsd:URL) 


 



IMoteCE 


architecture(url:x86) 

architecture(x86) 

Note: Individuals of these classes can describe resources of an AM. 

vmm(url:None) 

storageCapacity(160GB) storageCapacity(350GB) 

'virtualize' property is a kind of adaptation. 

 

 

cpu(url:PXA271) 

ServerCloud 

Testbed 


storageCapacity(32MB) 

Note: need to make these types/individuals 

 

StarGate 

 

EucaM1Large 

vmm(url:Xen) or vmm 

(KVM) 






 

EucaC1Medium 


(KVM) 






 

EucaM1Small 


(KVM) 






 

EucaM1XLarge 


(KVM) 






 

EucaC1XLarge 


(KVM) 






 

CoreLabCluster 

virtualize 

(url:CoreLabNode) 

 

EucalyptusCluster 

virtualize 

(url:EucaM1Small) or 

virtualize 

(url:EucaC1Medium) or 

virtualize 

(url:EucaC1XLarge) or 

virtualize 

(url:EucaM1Large) or 

virtualize 

(url:EucaM1XLarge) 

 

PlanetlabCluster 

virtualize 

(url:PlanetLabNode) 

 

ProtogeniCluster 

virtualize 

(url:ProtoGeniNode) 

Note: More info needed 

 

ORBITTestbed 

virtualize(???) 

 

KanseiGenieTestbed 

virtualize(url:MicaNode) 

or virtualize 

(url:TelosBNode) or 

virtualize(url:IMoteNode) 

 

ProtoGeniCE-d710 

vmm(url:None) or vmm 

(url:OpenVZ) or vmm(url:KVM) 

cpu(url:QuadNehalem24) 

memory storage 

 

d710Storage 

type(url:Storage) 

storageVolume(url:SATA250G) 

numVolumes(2) 

 

ProtoGeniCE-pc3000 



cpu(url:Xeon3) 

memory 

storage 

 

pc3000Storage 


storageVolume(url:SCSI146G) 

numVolumes(2) 

 

ProtoGeniCE-pc850 



cpu(url:Pentium3) 

memory 

storage 

 

pc850Storage 


storageVolume(url:IDE40G) 

numVolumes(1) 

 

CoreLabCE 

vmm(url:KVM) 

 

d710Memory 

type(url:Memory) 

ramType(url:DDR2-1066) 

memoryCapacity(12GB) 

 

pc3000Memory 


ramType(url:DDR2-400) 

memoryCapacity(2GB) 

 

pc850Memory 


Presentation title goes here" 14! 

ramType(url:PC133-SDRAM) 

memoryCapacity(512MB)

Request and Embedding! 

– Request Type" 

• Bound vs. Unbound" 

• Single cloud site vs. multiple cloud sites" 

– Connection:" 

– Virtual Topology:" 

– Virtual cluster:Node group and dependency: Master/Slave groups, " 

– Properties:" 

• Reservation ID" 

• Reservation term: time.owl" 

• VM type: EC2Small" 

– Predefined (CPU, Memory, Disk)" 

• IP addresses" 

– Range per Node Group " 

• VM Image" 

– Image URL (XML metafile e.g. ImageProxy or OVF)" 

• Post-boot NEuca script" 

• Network QoS" 

– Bandwidth" 

– Embedding Policy! 

• Path computing! 

• Topology embedding! 

• Cluster and topology partitioning and embedding! 

" ""

ORCA Architecture and Information Life 

Cycle! 

"#$%&'(%)*! 

+,)*&,-!.,,-/!!!!! 

9%:5%/*! 

9;

Outline! 

• ORCA Architecture" 



– Substrate Stitching" 

– Semantic resource representations" 

• Semantic graph ! 

• Substrate: NDL-OWL, Cloud! 

• Domain: Abstraction! 

• Request: Cluster, connection, topology! 




– User Interface: Web, XMLRPC API" 



– Workflows" 

17!

ProtoGENI Interoperability! 

• Resource description conversion" 

– From NS2/RSpec v1/v2 to NDL-OWL requests" 

– http://geni-test.renci.org:11080/ndl-conversion/convert.jsp" 

• GENI AM API and ProtoGENI API support via 

ORCA XMLRPC controller in the SM" 

– Validated via GPOs ʻomniʼ" 

– ListResource(),createSliver(),sliverStatus(),deleteSliver()" 

• Next steps" 

– Advertizing ORCA clouds to ProtoGENI" 

– Creating ProtoGENI-friendly manifests" 

18!

Cluster-D connectivity! 

19!

Ongoing inter-cluster connectivity, 

interoperability, and application integration! 

• GENI AM API and ProtoGENI API support via 

ORCA XMLRPC controller in the SM. " 

• Developing I2 ION, and OSCAR interfaces" 

• Connectivity to ESNet" 

• Monitoring integration" 

• Hadoop based analytics" 

• Scientific workflow" 

20!

Use case 1: Wide area Hadoop based 

scientific applications over clouds! 

• Map-reduce is an example of a step within a 

complex scientific workflow." 

• Hadoop VM image, script to launch Hadoop 

cluster on Neuca/EC2 clouds." 

• ORCA controller to provision a Hadoop cluster 

on multiple Neuca/EC2 clouds interconnected 

by multi-domain high speed network." 

• Performance study with different distributions of 

Hadoop nodes." 

• Application: Hadoop-Blast." 

21!

Use case 2: Scientific Workflow! 

• Multi-scale chemistry simulation pipeline (UNC 

– EFRC, DOE)" 

• MotifNetwork (an NSF CDI-II project)" 

• Integrated Microbial Genomes or IMG(DOE 

environmental genomics) " 

• Nearby Supernova Factory or SNfactory (DOE 

cosmological physics)," 

• Workflow for Collaborative Adaptive Sensing of 

the Atmosphere or CASA (an NSF ERC)." 


More Information! 

• http://geni-orca.renci.org" 

• http://www.networkedclouds.org"

Download

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?