cost-efficient deployment of distributed software services · 2015-02-17 · cost-efficient...
TRANSCRIPT
1/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Máté J. Csorba
Cost-efficient deployment of
distributed software services
2/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o The deployment problem
• MAPE, notation, target environment
o Intro to CEAS
o How to encode pheromones
o Scenario 1 : Collaborating software components
o Scenario 2 : Replication management
o Scenario 3 : Cloud computing
o Validation using ILP
o Summary & not covered
Short introduction & contents
Cost-efficient deployment of distributed software services
Cost functions
Bio-inspired decentralized optimization strategies
Cross-validation
3/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
The deployment problem and its complexity
#S
#N
Deployment
mapping
Focus on:
- load-balancing,
- remote comm.,
- replica management,
- and financial costs
4/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o REQs 1 – 6 :
1. Efficiency
2. Robustness
3. Autonomicity
4. Adaptivity
5. Scalability
6. Generality &
Extensibility
In the context of the MAPE cycle
Deployment
mapping
#
S
#N
5/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Target environment
6/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o Constants:
• N = {n1, n2};
• D = {d1}; d1 = {n1, n2};
• S = {S1}; C = {ci, cj}; K = {kk};
o Variables:
• D1 = {d1}; H1 = {n1, n2};
• L1 = {30 => n1, 20 => n2};
• M1 = {ci => n1, cj => n2};
Notation example
7/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Cross-Entropy Ant System
o Originally: robust, distributed path management in network nodes
• Schoonderwoerd et al. – ant based routing scheme (1997)
• Rubinstein – Cross-Entropy for stochastic optimisation (1999)
• Helvik & Wittner – distributed, autoregressive variant (2003)
o 2 types of ants:
• Explorer => cover up the search space, i.e. do random search
• Normal => optimize mapping
o 2 phases in an iteration
• Forward search => ants looking for a deployment mapping
• Backtracking => ants update the pheromone database
Introduction to CEAS
8/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
cost F()r updatetemp(cost)
CEAS deployment strategies – how it works
n1
n2
n3
n4
n5
n6
?
?
??
select : mn,rget : ln,r
Ci = Ø ?
.
..
.
...
....
...
..
.....
...
Fwd. searchPheromone update
9/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o Parallel nests
• In joining / splitting networks
o Load-sampling
• To facilitate load-balancing and gather information
o Guided random hop-selection
• Taboo-listing
• Cover domains and then nodes
o Binding & release of instances / maintenance
• Ease convergence
• Conditional
CEAS – auxiliary adjustments
10/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
CEAS – in a class–diagram
o Implemented in Simula/DEMOS
o Peersim (Java) ongoing
11/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Pheromone encodings
o Tables distributed over the possible hosts
Instances
Ye
s / N
o
12/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o Deployment of collaborating components – UML 2.0 collaborations
o PAPER A – C
o Target
• Load-balancing (i.e. min. deviation from average load)
• Minimization of remote communication
Scenario 1 : Collaborating software components
o E.g.
S1 : a simple client-server
S2 : authorization and
authentication server
13/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
with global load-balance estimate
Scenario 1 (cont’d)
and communication cost
o Deployment cost function
14/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
27 = max DB size
exploration
117 =
optimal cost
Scenario 1 (cont’d)
o Example run from PAPER A
Iterations
Co
st
/ D
B s
ize
15/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Scenario 1 (cont’d)
o Deployment cost function
where
o Deployment costs for multiple services
• Eliminating global knowledge (~T )
• Multiplicative (~better convergence)
16/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o Examples with 3 services from PAPER C
Scenario 1 (cont’d)
Node error & repair New node
Iterations
Co
st
17/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o Deployment of replicas
o PAPER D – E
o Target
• Load-balancing
• Node- & domain-disjointness
Scenario 2 : Replication management
o E.g.
S3 : a simple
DB replica
S4 : 3-way
replication
18/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Scenario 2 (cont’d)
o Deployment cost function
where
and
Domain-disjointness Node-disjointness Load-balancing
19/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Scenario 2 (cont’d)
o Example from PAPER E
• 65 replicas, 11 nodes in 5 domains
20/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
• Costs of service S10
• 2000 explorer iterations
• Splitting of d1 after 4000 iterations
• Merging the regions after 5000 iterations
Scenario 2 (cont’d)
o Example from PAPER E
21/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o Deployment of VM replicas in a cloud-like environment
o PAPER F
o Target
• Load-balancing
• Node- & domain-disjointness
• Minimize financial costs
Scenario 3 : Cloud computing
• 5 private (costs 0), 2
public clouds (costs 1
and 10)
• 18 clusters, 130
nodes (cluster sizes
of 10 or 5 nodes)
• 5 x 25 services,
replication level 5,
=> 625 VMs
22/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Scenario 3 (cont’d)
o Deployment cost function
o where the weighting function is defined as
o and the sum of financial costs is
23/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o Combination of cost function terms in F4
Scenario 3 (cont’d)
24/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o Results of 100 simulation runs each setting
• No cluster weights
• Linear weighting
• Exponential weighting
Scenario 3 (cont’d)
25/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Validation using ILP
o PAPER G
o ILP (AMPL/CPLEX)
Examples
Cost
26/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Validation using ILP (cont’d)
o ILP built with objective:
o and corresponding constraints:
o with the variables:
o Cross-validation of exact solutions vs. heuristics
• Global-view vs. decentralized
||
1
||
1
minK
i
ii
N
j
j colf
jijiji colmbT ,,,, ,,
212,1,
,,
||
1 ,
||
1 ,
,,
||
1 ,
, ,, ),,( ; 1
,, ),,( ; )2(
;
;
, ;
;1
jjkikilljkji
jkikilljkji
jj
C
i jii
jj
C
i jii
jijiji
N
j iji
nncccckcolmm
ncccckcolmm
nmeT
nTme
ncbm
cm
~ F1()
Impediments:
• global view in memory,
• no support for dynamicity,
• mi,j size explosion
27/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Summary
o CEAS
• Heuristic optimization
• Decentralized algorithm, ant-like agents
o Application domains
• Deployment of collaborating sw components
• Deployment of replicas
• VM instance placement in private and public clusters
1) Cost functions
2) Heuristic algorithms
3) Modeling scenarios
4) Cross-validation
28/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Publications
o Category I. – Validation of mappings
o Category II. – Load-balancing and communication costs
o Category III. – Load-balancing and replication costs
o Category IV. – Cluster costs
29/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
o Translation of the simulations to PeerSim
o Additional ILP models
o The power saving dimension
o Migration costs
o Scaling and larger problem sizes
o Incremental scaling
o Coordination and designated nest selection
o Better and broader service models
o Introducing clients
o Costs derived from service models
o Deployment diagrams
o Feedback to functional design
o Experiments using a middleware platform, e.g. DARM
Not covered, not implemented, not yet given up
30/30
PhD Thesis defense, 12.05.2011Máté J. Csorba
Thank You for your attention!*
* and many thanks to
Poul E. Heegaard,
Peter Herrmann, and
Hein Meling