Model transformation (MT) of very large models (VLMs), with mil- lions of elements, is a challenging cornerstone for applying Model- Driven Engineering (MDE) technology in industry. Recent research efforts that tackle this problem have been directed at distributing MT on the Cloud, either directly, by managing clusters explicitly, or indirectly, via external NoSQL data stores. In this paper, we draw attention back to improving efficiency of model transformations that use EMF natively and that run on non-distributed environments, showing that substantial performance gains can still be reaped on that ground. We present Yet Another Model Transformation Language (YAMTL), a new internal domain-specific language (DSL) of Xtend for defining declarative MT, and its execution engine. The part of the DSL for defining MT is similar to ATL in terms of expressiveness, including support for advanced modelling contructs, such as multiple rule inheritance and module composition. In addition, YAMTL provides support for specifying execution control strategies. We experimentally demonstrate that the presented transformation engine outper- forms other representative MT engines by using the batch transformation component of the VIATRA CPS benchmark. The improvement is, at least, one order of magnitude over the up-to-now fastest solution in all of the assessed scenarios.

1. Expressive and Efficient Model Transformation
with an Internal DSL of Xtend
Artur Boronat
TL

2. Scale OUT Using Cloud Technology…?
TL
Very Large Models (VLMs):
Model Transformation (MT)
a key bottleneck
Challenge to application
in industry
Explicitly
Parallelization
[Tisi et al. MoDELS’13]
Distribution
[Benelallam et al. SLE’15]
Implicitly (NoSQL)
Models only
[Gwendal et al. Sci.
Comput. Program 2017]
MT too
[Gwendal et al. ASE’17]

4. Scale UP with…
Yet Another Model Transformation Language !
• Internal DSL of Xtend
• Compiled to Java and excellent interoperability with JVM programs
• High-level API to work with collections
• Succinct syntax to work with lambda expressions
• IDE support: type checker and debugger…
• Focus on
• expressiveness: advanced contructs to model MT using
ATL as reference
• efficiency: MT performance using VIATRA3 as reference
• productivity as a consequence of being an internal DSL
TL
Krikava et al. On the use of an internal DSL for enriching EMF models. OCL and Textual
Modelling. 2012.
Hinkel et al. Using internal domain-specific languages to inherit tool support and modularity for
model transformations. SoSyM. 2017.

5. Expressive Model Transformation

6. MTs in YAMTL: ORM Example
name: String
«stereotype»
NamedElt
«stereotype»
Classifier
DataType
Package
isAbstract: Boolean
Class
*
owner
attr
*
type
*super
*
classifiers
name: String
«stereotype»
Named
Type
Database
Column
Table
*
owner
col
*
type
*
types
*
tables
keyOf
key
*
0..1
name = "Item"
1:Class
multiValued: Boolean
Attribute
name = "product"
multiValued=false
2:Attribute
name = "Order"
4:Class
name = "date"
multiValued=false
3:Attribute
name = "items"
multiValued=true
5:Attribute
attr attr attr
1:Class 4:Class
name = "Item"
1:Table
name = "product"
2:Column
name = "Order"
6:Table
name = "date"
3:Column
name = "pk_Order"
7:Column
col
col
col
name =
"pk_Item"
4:Column
col
name = "fk_Item--items-->Order"
5:Column
col
name = "Item"
1:Table
name = "Invoice"
6:Table
typename = "String"
0:DataType
typetype
0:DataType
Source metamodel (class diagrams) Target metamodel (database schemas)
Initial source model
Propagation of delta a (case 4)
Ms MtInitial target model
TL

7. YAMTL Module Declaration
class cd2db extends YAMTLModule {
new () {
header().in('cd', CD).out('db', DB)
ruleStore(
// rule declarations
)
helperStore(
// helper declarations
)
}
val CD = CDPackage.eINSTANCE
val DB = RelationalPackage.eINSTANCE
// other helpers as Xtend methods
}
Façade interface:
loading models,
execution, statistics
MT name
MT signature for
traceable models
inOut()
for in-place MT
TL

8. new Rule('ClassToTable')
.in('c', CD.class_).build()
.out('t', DB.table, [
// environment vbles
val c = 'c'.fetch as Class
val t = 't'.fetch as Table
val pk_col = 'pk_col'.fetch as Column
// bindings
t.name = c.name
t.col += c.attr.fetch as List<Column>
t.col += pk_col
val list = c.incomingReferences
.fetch('ReferenceToFkColumn')
as List<Column>
if (list !== null)
t.col += list
]).build()
.out('pk_col', DB.column, [
// environment vbles
val c = 'c'.fetch as Class
val pk_col = 'pk_col'.fetch as Column
// bindings
pk_col.name = 'pk_' + c.name
]).build()
.build()
YAMTL Rule
rule
• matched | lazy | uniqueLazy
• transient
• filter(EXPR): global constraints
• using(VBLE, EXPR): local variables
• endWith(EXPR): imperative expression
name = "Item"
1:Class
name = "Item"
1:Table
name = "pk_Item"
4:Column
col
TL

9. new Rule('ClassToTable')
.in('c', CD.class_).build()
.out('t', DB.table, [
// environment vbles
val c = 'c'.fetch as Class
val t = 't'.fetch as Table
val pk_col = 'pk_col'.fetch as Column
// bindings
t.name = c.name
t.col += c.attr.fetch as List<Column>
t.col += pk_col
val list = c.incomingReferences
.fetch('ReferenceToFkColumn')
as List<Column>
if (list !== null)
t.col += list
]).build()
.out('pk_col', DB.column, [
// environment vbles
val c = 'c'.fetch as Class
val pk_col = 'pk_col'.fetch as Column
// bindings
pk_col.name = 'pk_' + c.name
]).build()
.build()
YAMTL Rule
input pattern elements
• with(VBLE) dependencies
• filter(EXPR): local constraint (optional)
• derivedWith(EXPR)
name = "Item"
1:Class
name = "Item"
1:Table
name = "pk_Item"
4:Column
col
TL

10. new Rule('ClassToTable')
.in('c', CD.class_).build()
.out('t', DB.table, [
// environment vbles
val c = 'c'.fetch as Class
val t = 't'.fetch as Table
val pk_col = 'pk_col'.fetch as Column
// bindings
t.name = c.name
t.col += c.attr.fetch as List<Column>
t.col += pk_col
val list = c.incomingReferences
.fetch('ReferenceToFkColumn')
as List<Column>
if (list !== null)
t.col += list
]).build()
.out('pk_col', DB.column, [
// environment vbles
val c = 'c'.fetch as Class
val pk_col = 'pk_col'.fetch as Column
// bindings
pk_col.name = 'pk_' + c.name
]).build()
.build()
YAMTL Rule
output pattern elements
• action(EXPR) dependencies
• drop()
name = "Item"
1:Class
name = "Item"
1:Table
name = "pk_Item"
4:Column
col
TL

11. new Rule('ClassToTable')
.in('c', CD.class_).build()
.out('t', DB.table, [
// environment vbles
val c = 'c'.fetch as Class
val t = 't'.fetch as Table
val pk_col = 'pk_col'.fetch as Column
// bindings
t.name = c.name
t.col += c.attr.fetch as List<Column>
t.col += pk_col
val list = c.incomingReferences
.fetch('ReferenceToFkColumn')
as List<Column>
if (list !== null)
t.col += list
]).build()
.out('pk_col', DB.column, [
// environment vbles
val c = 'c'.fetch as Class
val pk_col = 'pk_col'.fetch as Column
// bindings
pk_col.name = 'pk_' + c.name
]).build()
.build()
YAMTL Output Action
fetch
to access execution environment
name = "Item"
1:Class
name = "Item"
1:Table
name = "pk_Item"
4:Column
col
TL

12. new Rule('ClassToTable')
.in('c', CD.class_).build()
.out('t', DB.table, [
// environment vbles
val c = 'c'.fetch as Class
val t = 't'.fetch as Table
val pk_col = 'pk_col'.fetch as Column
// bindings
t.name = c.name
t.col += c.attr.fetch as List<Column>
t.col += pk_col
val list = c.incomingReferences
.fetch('ReferenceToFkColumn')
as List<Column>
if (list !== null)
t.col += list
]).build()
.out('pk_col', DB.column, [
// environment vbles
val c = 'c'.fetch as Class
val pk_col = 'pk_col'.fetch as Column
// bindings
pk_col.name = 'pk_' + c.name
]).build()
.build()
YAMTL Output Action
fetch
to resolve references in…
name = "Item"
1:Class
name = "Item"
1:Table
name = "pk_Item"
4:Column
col
TL

13. new Rule('ClassToTable')
.in('c', CD.class_).build()
.out('t', DB.table, [
// environment vbles
val c = 'c'.fetch as Class
val t = 't'.fetch as Table
val pk_col = 'pk_col'.fetch as Column
// bindings
t.name = c.name
t.col += c.attr.fetch as List<Column>
t.col += pk_col
val list = c.incomingReferences
.fetch('ReferenceToFkColumn')
as List<Column>
if (list !== null)
t.col += list
]).build()
.out('pk_col', DB.column, [
// environment vbles
val c = 'c'.fetch as Class
val pk_col = 'pk_col'.fetch as Column
// bindings
pk_col.name = 'pk_' + c.name
]).build()
.build()
YAMTL Output Action
a) declative rules
new Rule('AttributeToColumn')
.in('att', CD.attribute).filter([ ... ]).build()
.out('col', DB.column, [ ... ]).build()
.build(),
fetch
to resolve references in…
name = "Item"
1:Class
name = "Item"
1:Table
name = "pk_Item"
4:Column
col
TL

14. new Rule('ClassToTable')
.in('c', CD.class_).build()
.out('t', DB.table, [
// environment vbles
val c = 'c'.fetch as Class
val t = 't'.fetch as Table
val pk_col = 'pk_col'.fetch as Column
// bindings
t.name = c.name
t.col += c.attr.fetch as List<Column>
t.col += pk_col
val list = c.incomingReferences
.fetch('ReferenceToFkColumn')
as List<Column>
if (list !== null)
t.col += list
]).build()
.out('pk_col', DB.column, [
// environment vbles
val c = 'c'.fetch as Class
val pk_col = 'pk_col'.fetch as Column
// bindings
pk_col.name = 'pk_' + c.name
]).build()
.build()
YAMTL Output Action
b) lazy rules
new Rule('ReferenceToFkColumn')
.uniqueLazy
.in('ref', CD.attribute).build()
.out('fk_col', DB.column, [ ... ]).build()
.build()
a) declative rules
new Rule('AttributeToColumn')
.in('att', CD.attribute).filter([ ... ]).build()
.out('col', DB.column, [ ... ]).build()
.build(),
fetch
to resolve references in…
name = "Item"
1:Class
name = "Item"
1:Table
name = "pk_Item"
4:Column
col
TL

15. Reuse in YAMTL
• Multiple Rule Inheritance and abstract rules
• Dispatch semantics: leftmost top-down search for most concrete applicable
rule for each object
• Execution semantics of output element actions (inherits/overriding)
• Module composition by class extension
• Inheritance: by default
• Redefinition: when a rule (or a helper) in the extending module have a
compatible interface with one rule (or helper) in the extended module
• Module reuse in other JVM programs using imports
TL

16. Efficient Model Transformation

17. VIATRA CPS Benchmark
YAMTL solution:
1. Helper to determine receiving transitions (waitFor)
2. Build the structure of deployment model using rules with high priority
3. Initialize triggers with a transient rule (only updates)
4. Extract CPS traceability model from YAMTL’s internal traceability model
action triggers to be
set after behaviours
are instantiated
behaviours are
instantiated (copy)
traceability model
TL

18. VIATRA CPS Benchmark
• Adaptation
• Model generation with VIATRA CPS model generator
• Each experiment (size): 13 iterations
• Median of inner results (excluding 1st – warm up)
1
2
4
8
16
32
64
128
256
512
1024
2048
4096
8192
16384
32768
65536
131072
262144
1 2 4 8 16 32 64 128 256 512 1024 2048 4096 8192 16384 32768
time (ms)
scale
Client-server
VIATRA EMFTVM Xtend ATL YAMTL EMFTVM (lazy) ATL (lazy)
scale factor
time (ms)
YAMTL
EMFTVM
(imp.)
EIQEMFTVM
(decl.)
ATL2010
(imp.)
ATL2010
(decl.)
Xtend
(optimized)
8 3k 7k 7 45 36 35 120 124 258
16 5k 15k 14 74 56 65 195 201 979
32 10k 29k 26 113 94 102 445 438 3579
64 20k 59k 49 226 190 212 1400 1346 26865
128 40k 123k 102 480 408 473 5316 4916 54594
256 80k 229k 216 768 786 827 20806 20595 223874
512 160k 453k 450 1442 1607 1887 107526 120443
1024 319k 940k 927 2935 3563 5288
2048 636k 1,940k 1889 6019 7800 15942
4096 1.27M 3.69M 3724 10141 15735 50886
8192 2.54M 7.25M 7610 19145 32866 184803
16384 5.08M 14.26M 15838 36490 72377
32768 10.16M 27.53M 32680 67969 oom
• Findings
• YAMTL at least one
orders of magnitud
faster than EIQ
• YAMTL also uses
resources
efficiently
• ATL/EMFTVM with
lazy rules scales
well but trafo not
complete
TL

19. Final Remarks
https://yamtl.github.io/
• YAMTL brings declarative M2M transformations to the Xtend world
• Advanced model transformation constructs
• Support for VLMs/good performance
• Productivity and analysis tools available via Xtend/Java
• Next: Incremental MT - most scalable solution of TTC’18
• Interested in exploring applications
• Thanks to Ábel Hegedüs for integrating YAMTL
solution in official VIATRA CPS Benchmark
Thank you for your attention
Artur Boronat