A presentation of the paper developed in the SAFIRE project titled "Value-based manufacturing optimisation in serverless clouds for Industry 4.0", delivered at the Genetic and Evolutionary Computation Conference (GECCO) at Kyoto, Japan in July 2018.
2. Serverless computing
SAFIRE project
Industry 4.0
System architecture
Problem description
Implementation issues
Experimental results
Conclusion
GECCO 2018 Kyoto, July 15th-19th 2018 2
Agenda
3. A misnomer - still requires servers.
A cloud-computing execution model in
which the cloud provider dynamically
manages the allocation of machine
resources.
Statelessness.
Pricing is based on the actual amount
of resources consumed by a function.
GECCO 2018 Kyoto, July 15th-19th 2018 3
Serverless computing (FaaS)
4. GECCO 2018 Kyoto, July 15th-19th 2018 4
Limits of serverless services in the
most popular public clouds
Limit IBM Cloud
Function (Apache
OpenWhisk)
Amazon lambda Google Cloud
Functions
Azure Functions
(Consumption
plan)
languages &
environments
Docker, JavaScript
(Node.js runtime),
Python, Java, and
Swift
JavaScript
(Node.js runtime),
Python, Java, C#,
Go
JavaScript
(Node.js runtime)
C#, F#, JavaScript
(Node.js runtime),
Java, or PHP
max code size 48MB 50MB 100MB
(compressed)
not documented
max no of
concurrent
execution
1000 1000 (per region) 1000 200
max memory
footprint
512MB 3008MB 500MB 1500MB
max no. of
function
invocations per
time unit
5000 per 60s not documented 1,000,000
per 100 s
not documented
timeout 600s 300s 540s 600s
5. Funded under: H2020 -
EU.2.1.1. - Industrial
Leadership
Reconfiguration-as-a-
Service for dynamic smart
factories and manufactured
smart products.
Exploits cloud-based
services to continuously
optimise the performance of
production systems and
products.
GECCO 2018 Kyoto, July 15th-19th 2018 5
SAFIRE Project
7. GECCO 2018 Kyoto, July 15th-19th 2018 7
On-demand manufacturing
scheduling optimisation
optimisation of manufacturing
plant that is specified via the
(max,+) algebra
8. a ⊕ b = max ( a , b )
a ⊗ b = a + b
⊗ has a higher precedence than ⊕
GECCO 2018 Kyoto, July 15th-19th 2018 8
(max, +)-algebra
9. semiring
associativity of ⊕, ⊗
commutativity of ⊕, ⊗
distributivity over ⊕
zero element
ε = -∞ (for ⊕) – ”never before”
unit element
e = 0 (for ⊗)
idempotency of ⊕
GECCO 2018 Kyoto, July 15th-19th 2018 9
(max, +)-algebra properties
10. A manufacturing process begins at time t1 in state 1.
t2 = t1 ⊗ dA = t1 + dA
t3 = t2 ⊗ dB = t2 + dB
t4 = t3 ⊗ dD = t3 + dD
t5 = (t2 ⊗ dC) ⊕ (t4 ⊗ dE) = max(t2 + dC, t4 + dE)
t6 = t5 ⊗ dF = t5 + dF
GECCO 2018 Kyoto, July 15th-19th 2018 10
(max, +)-algebra for Activity on Arrow
(AoA) plant representation - example
11. Assumption: both the solution quality and
optimisation time are relevant to the end-user.
Optimisation time impact: described by a value
curve in a monetary unit.
GECCO 2018 Kyoto, July 15th-19th 2018 11
Value curve of manufacturing order
13. GECCO 2018 Kyoto, July 15th-19th 2018 13
Stopping criterion to maximise
overall profit
14. Optimisation cost upperbound: 𝑐𝑖 = 𝛽 ∙ 𝑡𝑖 ∙ 𝑝𝑖
Manufacturing cost: 𝑓𝑖
Profit: 𝑃𝑖 = 𝑉𝐶 σ 𝑗=1
𝑖
𝑡𝑗 − σ 𝑗=1
𝑖
𝑐𝑗 − 𝑓𝑖
GECCO 2018 Kyoto, July 15th-19th 2018 14
Time and cost of stage execution
t
i-th stage
slave1
slave2
slave3
slavepi
ti
cost of a single slave
execution per 1 time unit
the lowest value of FF
after the i-th stage
16. Uses jMetal framework
Stateless
Placed inside a Docker container
REST-compliant Web service
Invoked with Apache OpenWhisk
Fully documented API
Communication between the master
and slave nodes is performed via JSON
over HTTP
GECCO 2018 Kyoto, July 15th-19th 2018 16
GA-based optimiser - implementation
issues
17. AoA with
22 nodes
6 levels
43 arrows
Each manufacturing process can be
executed on
one of 8 machine types
each machine - from 1 to 9 operating
modes
GECCO 2018 Kyoto, July 15th-19th 2018 17
Experiment
18. Value curve parameters
Initial number of containers run in
parallel is set to p1 = 10
GECCO 2018 Kyoto, July 15th-19th 2018 18
Experiment – parameters
AT = 0 D = 500s Z = 1000s
Vmax = 5000GBP
19. the Standard Deviation of population
fitness after the i-th stage, sdi , ≤ 10−6
or
minimal fitness function value fi was not
improved during the previous 20 stages
or
the number of stages == 100
GECCO 2018 Kyoto, July 15th-19th 2018 19
Baseline stopping criterion
20. Baseline: 71 stages, p71 = 2
Proposed: 6 stages (the second best, only 2%
worse than the highest possible profit)
The average profit prediction error - ca. 2%.
GECCO 2018 Kyoto, July 15th-19th 2018 20
Results for β = 0.5 GBP
Max profit
21. Baseline: 71 stages, p71 = 2
Proposed: 41 stages (the highest possible profit)
GECCO 2018 Kyoto, July 15th-19th 2018 21
Results for β = 0 GBP
Max profit
22. 10 different manufacturing orders (18 to 59
manufacturing processes)
No influence on the profit yielded by algorithm
Execution times differ significantly
GECCO 2018 Kyoto, July 15th-19th 2018 22
#GA generations per stage
23. AoA instance with 22 nodes, 6 levels
and 43 arrows
GECCO 2018 Kyoto, July 15th-19th 2018 23
Scalability
Profit [GBP]
[GBP]
24. 30 manufacturing orders (18 to 59 steps),
the maximal possible income – 5000 GBP
Proposed vs baseline: Optimisation
process 18.5 times faster
The obtained fitness
function values are
34% worse on
average
GECCO 2018 Kyoto, July 15th-19th 2018 24
Stopping criteria – time and ff value
(lower is better)
25. Baseline: 86% of the considered
manufacturing orders lead to financial
loss (cumulative profit: -124497 GBP)
Proposed: all of the
orders are profitable
(cumulative
profit: 83877 GBP)
GECCO 2018 Kyoto, July 15th-19th 2018 25
Stopping criteria - profit
(higher is better)
26. This article describes a serverless, cloud-based
architecture that provides general and scalable
support for the ‘Just in Time’ manufacturing
process envisioned for ‘Industry 4.0’.
The architecture is equipped with a novel adaptive
stopping criterion for optimising Overall Equipment
Effectiveness (OEE), in which the predicted
cost/benefit ratio of performing further
optimisation is grounded in monetary units.
The method was applied to a collection of case
studies for optimal configuration of manufacturing
plants specified via the (max, +) algebra.
We deployed stateless, Dockerised containers and
obtained near maximum profit from the resulting
optimisation.
GECCO 2018 Kyoto, July 15th-19th 2018 26
Conclusion