VCOC experiment in BonFIRE European Project (http://www.bonfire-project.eu) results. It shows a general architecture with fault-tolerance to use in fistributed Cloud environments and the usage of application performance indicators to trigger cluster elasticity. More information at www.cesga.es.
ICT role in 21st century education and its challenges
VCOC BonFIRE presentation at FIRE Engineering Workshop 2012
1. Building service testbeds on FIRE
Experimenting Virtual Clusters on distributed
Cloud environments using BonFIRE
A. Gómez, L.M. Carril, R. Valin, J.C. Mouriño, C. Cotelo
FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012
3. Objective of Experiment
Research the feasibility of using multiple Cloud environments
for the provision of Services which need a large pool of
CPUs (as High Throughput Computing).
Why
3 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 3
4. Objective of Experiment (II)
Experiments about time-to-deploy the virtual
infrastructure.
Experiments about elasticity triggered by application
performance.
Experiments about fault-tolerant virtual infrastructure.
4 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 4
9. Elasticity Engine
Takes information about
application performance
Get data about queues
Calculates Nr. VMs to
fulfill demanded
performance
Adds new VMs to
Experiment
9 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 9
10. VCOC Interfaces
Monitoring
GUI
dashboard Identity
LDAP
Portal Server
(Used by Portal,
VCOC Experiment Agent Experiment Manager,
API Resource Manager
uses Experiment Manager and Testbeds)
Experiment Manager
Portal
Usage
Elasticity Engine uses OCCI Monitoring Reservation Accounting API AuthZ
Resource Manager (Used by RM to
authorize resource
Resource Manager creation)
OCCI Reservation Message
Read/
Write Queue
Enactor
OCCI Reservation
SSH
Testbed
SSH Monitoring Monitoring
SSH
Gateway GUI API
VM
Gateway (Monitoring VM VM
Aggregator)
10 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 10
16. Name of the Managed
Experiment JSON VC Example
{
"name": "Virtual Cluster Experiment",
Maximum elapsed time for the Experiment in minutes
"description": "Virtual Cluster",
"duration": 180, Description of the experiment
"resources": [
{
"compute": {
"name": "master", Location for deployment
"description": "Virtual Cluster Master",
"instanceType": "small",
"locations": ["fr-inria"],
Main Storage Resource
"resources": [
{ "storage": "@VirtualClusterMaster-vcocdemo-v0.1"},
{ "storage": "@vcocdemo_volume-1GB-empty"}, Master Virtual Machine Configuration
{ "network": "@BonFIRE WAN"}
],
"contexts": [ Additional Storage Resource
{
}
] Network
}
},
{
"compute": {
"name": "client-1",
"description": "Virtual Cluster Client",
"instanceType": "small",
"locations": ["fr-inria"],
"resources": [
{ "storage": "@VirtualClusterClient-vcocdemo-v0.1"},
{ "network": "@BonFIRE WAN"}
Client Virtual Machine Description
],
"contexts": [
{
}
]
}
}
16 ]
} VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 16
17. EXPERIMENTS ABOUT
TIME-TO-DEPLOY
17 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 17
18. Experiment data
5 repetitions Maximum
Representation: Box plot
Average
Median
Minimum
18 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 18
19. Deployment time vs VC size
BonFIRE site: INRIA Experiment Manager Start VMs
Reserved: 2 physical nodes with
2x12 cores AMD Opteron 1.7 GHz
and 48 GB RAM
VC Configuration:
• Simple VC
• One master (small + 4GB
storage)
• n-1 CEs (small) Conf. OGS Undeployment
19 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 19
20. Deployment time vs storage size
Without cache
BonFIRE site: INRIA
Reserved: 2 physical nodes with
2x12 cores AMD Opteron 1.7 GHz
and 48 GB RAM
VC Configuration:
Simple VC 16VM
One master (small + x GB
storage)
15 CEs (small)
With Cache
20 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 20
21. Two clusters in one site
Second cluster delayed because the deployment of first VC.
BonFIRE site: INRIA
Reserved: 2 physical nodes
with 2x12 cores AMD Opteron First Virtual Cluster
1.7 GHz and 48 GB RAM
VC Configuration: 4 VM 8 VM
Simple VC
One master (small + 4GB
storage)
n-1 CEs (small)
Second Virtual Cluster
21 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 21
22. Application execution. One vs Two
sites
VC Conf.: Distributed VC (_dist)
BonFIRE sites:
INRIA: Master + CEs
HLRS: CEs
Deployment time decreases.
Two sites faster than one site.
But because second site has better
CPUs.
Impact of deployment ~ 10% total
time.
22 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 22
23. Adding a new VM Client
VC Conf.: Simple VC
BonFIRE sites:
INRIA: Master + CEs
Using BonFIRE Resource
Manager
Very fast due to cache
23 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 23
25. Horizontal elasticity
Monitoring application
performance works.
We have modified software
to produce information more
frequently.
Execution with deadline.
Elasticity works.
25 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 25
26. FAULT TOLERANCE
EXPERIMENT WITH ELASTICITY
26 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 26
30. Conclusions
BonFIRE provides infrastructure for experiments about
new concepts and services on Cloud.
Distributed VC can be used to speed up HTC applications.
Elasticity engine based on application performance for
HTC works.
High QoS can be provided in VC using distributed VC +
elasticity.
30 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 30
31. Acknowledgment
To all BonFIRE team, specially to:
Maxence Dunnewind (INRIA)
Michael Gienger (HLRS)
Ally Hume (EPCC)
Kostas Kavoussanakis (EPCC)
David García Pérez (Atos Research)
31 VCOC, FIRE Engineering Workshop, Ghent, Nov. 6th – 7th 2012 BonFIRE 31
32. Building service testbeds on FIRE
Thank you for your attention
Project contact: bonfire@bonfire-project.eu
Experiment Contact: agomez@cesga.es