Guilherme Koslovski (INRIA – University of Lyon)
Wai-Leong Yeow (DoCoMo USA Labs)
Cedric Westphal ( DoCoMo USA Labs)
Tram Truong Huu (University of Nice – I3S)
Johan Montagnat (CNRS – I3S)
Pascale Vicat-Blanc Primet (INRIA - LYaTiss)
2 nd IEEE International Conference on Cloud Computing
Technology and Science, Indianapolis, 2010
RESO

•
Reliability : probability that a system will survive failures
•
Availability : fraction of time that a system is functional
99.95% availability 99.9% availability 99.95% reliability
100% uptime
•
Actually nothing more than SLAs.
–
Failure => credits
–
Lock-ins
–
No guarantees at all
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
100% network uptime
2

Convergence of computing and communication: Virtual Infrastructure is a concept emerging from Virtual Networks and Infrastructures as a Service
New models and tools to manage virtualized substrate & to help users in execution of their applications
Users
Network virtualization
Resources virtualization
Distributed & virtualized substrate
Grid computing experience
IaaS, PaaS, … XaaS concepts Complex applications
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
3

•
Network and IT resources are subject to random failures
•
Failures can be measured: mean time between failures (MTBF)
•
Impact of a failure on a distributed application:
• worker node failure: can affect the total execution time
• database and servers: can compromise the entire execution
•
Some applications can recover from failures but
•
This process usually affects the execution time
•
This complicates the application development
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
4

Reliability as a service offered by the infrastructure provider
Provide me a basic infrastructure
Provide me a reliable infrastructure
User
Application
VM 1 PM
VM 2 PM
VM n PM
BKP 1 PM
BKP 2 PM
BKP n PM
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
5

Reliability becomes a service offered by the infrastructure provider
Transparent realibility provisioning
Users (applications) have no knowledge about physical failures
User
Application
VM 1 PM
VM 2 PM
VM n PM
BKP 1 PM
BKP 2 PM
BKP n PM
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
6

Providing Transparent Reliability
Reliable Virtual Infrastructure description
Automatic generation of backup nodes and backup links
Allocation algorithm
Evaluation through a use case application
Conclusion & Future work
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
7

I.
Virtual Infrastructure description
II. Translation of reliability requirements into real backup nodes
III. Allocation of a reliable virtual infrastructure
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
8

VXDL file
General description
VXDL: Virtual private eXecution infrastructure
Description Language
– http://www.ens-lyon.fr/LIP/RESO/Software/vxdl/ vm1
Resources description
Network topology description
Timeline description
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc database
[1 GB, 2 GB]
2 GHz
2 cores
Location: lyon.fr
Reliability: 99.99% workers [100 nodes]
1 GB, 2 GHz
2 cores
Location: lyon.fr
Reliability: 99.9%
9

Translation of reliability requirements into replica nodes
Opportunistic Redundancy Pooling (ORP) mechanism
[W. Yeow et al, 2010]
:
Input:
Reliability level (user requirement)
Probability of physical failures (from MTBF)
Number of protected virtual nodes (user requirement)
Output: the number of backup nodes
– Backup nodes can be shared among different groups of critical nodes
– For example, two sets of backup nodes (k1 and k2) can be shared to protect two groups of critical nodes.
Thanks to ORP is required only the min(k1, k2)
[W. Yeow et al, 2010]: Designing and embedding reliable virtual infrastructures, VISA workshop 2010.
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
10

Backup links: consistent network topology
Step 1
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
Step 2
Step 3
11

An extended graph is composed by original description + backup components
Backup components can have specific constraints:
– For example, original node and backup node should be allocated on different physical racks
Subgraph-isomorphism detection
[Lischka et al, 2009]
Physical substrate
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
Embedded graph
12

The map provided by the allocation is interpreted and instantiate using the HIPerNet framework
[P. Primet et al, 2010]
•
Original VMs and replicas are synchronized by a modified version of the Remus live protection mechanism
[B. Cully et al, 2008]
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
13

Bronze Standard: distributed large-scale application
– Quantifies the maximal error resulting from medical-image analysis
– Large databases: more the data, more the accuracy
– 31 VMs: 512 MB,1 GHz
– 10 Mbps for each virtual link between the database and the workers
I) Translated into VXDL
II) Submitted to HIPerNet Two scenarios of reliability requirements:
– Database protection: a failure stops the application execution
– Workers protection: a failure increases the execution time
Testbed: Grid’5000
– Physical substrate is composed by 100 nodes:
– MTBF simulation values: 60000s, 30000s, 15000s
[D. Atwood et al., 2008]
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
14

Goal : quantify the cost of a reliable virtual infrastructure
Prices are based on Amazon EC2 for Europe
We do not include any specific link pricing cost without reliability support (short term lease): $2.95 / h
VM specifications Basic node
Short term lease $0.095
Long term lease $0.031
Prices for computing nodes protection (30 VMs, 99.9%):
MTBF
60000 s
30000 s
Backup
Nodes
5
8
15000 s
12
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
Total cost
$3.42
$3.71
$4.09
Short term
Reliability cost / total cost
16.1%
25.8%
38.7%
Total cost
$3.10
$3.19
$3.32
Long term
Reliability cost / total cost
5.3%
8.4%
12.6%
15

Goal: evaluate the application behavior when executing with reliability support
Application makespan without substrate failures: 1205s, used as baseline
– Database protection:
DB label : database is the unique component protected
Makespan increases proportionally to the number of failures
– Worker nodes protection:
WN label : only computing nodes are protected
Makespan slightly increases
1800
1600
MTBF DB
Increase
WN
Increase
NI
60000s 16.26%
30000s 26.47%
15000s 40.08%
0.2%
1.7%
3.2%
1400
1200
1000
800
600
400
200
0
NI 60000s 30000s 15000s
DB
WN
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
16

Goal: reliability service vs resubmission mechanism:
– Application is aware about substrate failures
– A task is resubmitted on a new computing node
– The makespan difference would have been more if backup nodes were not pre-allocated and configured
MTBF
60000s
30000s
15000s
Makespan
Increase
+13.08%
+19.67%
+22.19%
1600
1400
1200
1000
800
600
400
200
0
Reliability
Resubmission
60000s 30000s 15000s
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
17

Reliability becomes a service offered by the infrastructure provider
We have developed a framework to provide transparent reliability:
– A language to specify the reliability requirements;
– A mechanism to interpret these requirements and transform it in replicas (nodes and links)
– A map and allocation process to provisioning the reliability level required by the user
The framework was implemented on top of the HIPerNet framework, and validated over the Grid’5000 testbed
Future work includes:
– the implementation of a mechanism to protect virtual links
– a detailed investigation on the economical aspects
– Tomorrow there is a demonstration about the industry version of the HIPerNet framework (LYaTiss core) - http://www.lyatiss.com/
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
18

guilherme.koslovski@ens-lyon.fr, wlyeow@ieee.org, cwestphal@docomolabs-usa.com, tram@polytech.unice.fr, johan@i3s.unice.fr, pvb@lyatiss.com
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
19

Specifying and provisioning Virtual Infrastructures with HIPerNET . Fabienne Anhalt, Guilherme Koslovski, and
Pascale Vicat-Blanc Primet. ACM International Journal of Network Management (IJNM) - special issue on Network
Virtualization and its Management, 2010;
Joint elastic cloud and network framework for application performance optimization and cost reduction . Tram
Truong Huu, Guilherme Koslovski, Fabienne Anhalt, Pascale Vicat-Blanc Primet, and Johan Montagnat. Journal of
Grid Computing (JoGC) , 2010;
Reliability support in virtual infrastructures . Guilherme Koslovski, Wai-Leong, Cedric Westphal, Tram Truong Huu,
Pascale Vicat-Blanc Primet, and Johan Montagnat. In 2 nd IEEE CloudCom 2010, Indianapolis, USA;
A scalable security model for enabling Dynamic Virtual Private Execution Infrastructures on the Internet . Pascale
Vicat-Blanc Primet, Jean-Patrick Gelas, Olivier Mornard, Guilherme Koslovski, Vincent Roca, Lionel Giraud, Johan
Montagnat, and Tram Truong Huu. In IEEE/ACM CCGrid2009, Shanghai, May 2009;
Exploring the virtual infrastructure service concept in Grid'5000 . Pascale Vicat-Blanc Primet, Fabienne Anhalt, and Guilherme Koslovski. In 20th ITC Specialist Seminar on Network Virtualization, Hoi An, Vietnam, May 2009;
Executing distributed applications on virtualized infrastructures specified with the VXDL language and managed by the HIPerNET framework . Guilherme Koslovski, Tram Truong Huu, Johan Montagnat, and Pascale Vicat-Blanc
Primet. In CLOUDCOMP 2009, Munich, Germany, October 2009;
Virtual Resources and Interconnection Networks Description Language . Guilherme Koslovski, Pascale Vicat-Blanc
Primet, and Andrea Schwertner Charão. In GridNets 2008, Oct. 2008;
HIPernet: A Decentralized Security Infrastructure for Large Scale Grid Environments . Julien Laganier, Pascale
Vicat-Blanc Primet. In 6th IEEE/ACM International Conference on Grid Computing (GRID 2005), November 13-14,
2005, Seattle, Washington, USA, Proceedings, pages 140-147, 2005. IEEE;
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
20

2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
21

VXDL file
General description
Resources description
Network topology description
Timeline description
VI identification (name, owner, users)
Reservation (start and end time)
General properties:
– Location
– Security
– Monitoring
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
22

VXDL file
General description
Resources description
Network topology description
Timeline description
Resources and groups of resources
Resources types:
Virtual routers/switches
Virtual machines
Virtual storages
Virtual access points
Cross-layers parameters:
– Exclusivity
– Specific devices
– Location
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
23

VXDL file
General description
Resources description
Network topology description
Timeline description
Topology description:
- (Provisioned) virtual links
Specification of QoS constraints:
– Bandwidth (forward and reverse)
– Latency
– Reliability
– Security
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
24

VXDL file
General description
Resources description
Network topology description
Timeline description
often components are not used simultaneously or all along the VI lifetime
an internal timeline for each VI can help optimizing the allocation, scheduling, and provisioning processes
a timeline is composed by stages, delimited by temporal marks
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
25

Total cost is: original cost + reliability cost
Original cost
Reliability cost
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
26

Bronze Standard workflow is analyzed and translated into a VXDL file:
– 31 VMs: 512 MB,1 GHz
– 10 Mbps for each virtual link between the database and the workers
Two scenarios of reliability requirements:
– Database protection: a failure stops the application execution
– Workers protection: a failure increases the execution time
Physical substrate: Grid’5000 testbed
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
I) Translated into VXDL
II) Submitted to HIPerNet
27

Reliability is informed by the user considering the information exposed by the physical substrate; vm3 vm1 Link - l1
[20 Mb/s, 200 Mb/s] with monitoring
Links – l2, l3
[10 Mb/s, 100 Mb/s]
Node - vm2
[1 GB, 2 GB]
2 GHz
2 cores
Location: lyon.fr
Reliability: 99.9%
Routers – r1, r2
Layer: Ethernet
Dynamically configured
Link – l4, l5
{10 Mb/s, 50 Mb/s, 100 Mb/s}
Reliability: 99.99% vm4
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
28

Goals:
Evaluate the application behavior when executing with reliability support
Quantify the cost of a reliable ViPXi
Application makespan without substrate failures: 1205s, used as baseline
Simple cost model for the pricing:
– ViPXi cost + replicas cost
– Prices are based on Amazon EC2 for Europe
VM specifications Basic node
(1.7GB RAM)
Short term lease $0.095
– 1 hour reservation
– We do not include any specific link pricing
Long term lease $0.031
ViPXi cost without reliability support (short term lease): $2.95, used as baseline
Physical substrate is composed by 100 nodes:
MTBF simulation values: 60000s, 30000s, 15000s
2 nd IEEE CloudCom – 2010
G. Koslovski, W. Yeow, C. Westphal,
T. Huu, J. Montagnat, P. Vicat-Blanc
29

Goal : Quantify the cost of a reliable ViPXi
Prices are based on Amazon EC2 for Europe
We do not include any specific link pricing cost without reliability support (short term lease): $2.95
Prices for database protection (1 VM, 99.99%):
MTBF
60000 s
30000 s
15000 s
MTBF
60000 s
Fail prob.
0.03
0.06
Prices for computing nodes protection (30 VMs, 99.9%):
0.12
Fail prob.
0.03
Backup
Nodes
2
3
4
Backup
Nodes
5
2 nd
30000 s
15000 s
0.06
0.12
8
12
Total cost
$3.13
$3.23
$3.33
Total cost
$3.42
$3.71
$4.09
Short term
Reliability cost / total cost
6%
10%
Reliability cost / total cost
16.1%
25.8%
38.7%
Total cost
$3.01
$3.04
$3.07
Total cost
$3.10
$3.19
$3.32
VM specifications Basic node
Short term lease $0.095
Long term lease $0.031
Long term
Reliability cost / total cost
2%
3%
Long term
4%
Reliability cost / total cost
5.3%
8.4%
12.6%
30