The document describes a run-fail-grow approach for tailoring object-oriented runtimes to create very small customized runtimes. A virtualization infrastructure called Tornado uses a run-fail-grow algorithm to run an application, detect failures when missing dependencies are encountered, copy over the missing code from a reference application, and resume execution. Experiments show the approach can reduce runtime sizes by up to 99% compared to traditional runtimes, resulting in tailored runtimes as small as 10KB that are still able to support features like reflection. The approach works for dynamic languages and offers advantages over traditional static/dynamic tailoring solutions in supporting a wide range of applications and runtime configurations.

1. Run-Fail-Grow: Creating
Tailored Object-Oriented
Runtimes
G. Polito, L. Fabresse, N. Bouraqadi, S. Ducasse,
in Journal of Object Technology
Guille Polito
@guillep
https://guillep.github.io/

2. A(gain) need for
small and fast startup
• Micro Services
• Container architectures
• Small devices
2

3. Lets look at an application...
3

4. MainApp>>start
logger := StdoutLogger new.
logger log: ’Application has started’.
"do something"
logger log: ’Application has finished’.
StdoutLogger»newLine
stdout newLine.
StdoutLogger»log: aMessage
stdout nextPutAll: Time now printString.
stdout nextPutAll: aMessage.
stdout newLine.
RemoteLogger»log: aMessage
| socket |
socket := self newSocket.
socket nextPutAll: Time now printString.
socket nextPutAll: aMessage.
socket newLine.
RemoteLogger»newSocket
"...."
"creates an instance of socket given some configuration"
Log Lib
Base Libs
Logger <<interface>>
StdoutLogger
log:
newLine
App
MainApp
start RemoteLogger
log:
newSocket
Date Time :stdout Socket
4

5. There is extra stuff!
5

6. Log Lib
Base Libs
Logger <<interface>>
StdoutLogger
log:
newLine
App
MainApp
start RemoteLogger
log:
newSocket
Date Time :stdout Socket
MainApp>>start
logger := StdoutLogger new.
logger log: ’Application has started’.
"do something"
logger log: ’Application has finished’.
StdoutLogger»newLine
stdout newLine.
StdoutLogger»log: aMessage
stdout nextPutAll: Time now printString.
stdout nextPutAll: aMessage.
stdout newLine.
RemoteLogger»log: aMessage
| socket |
socket := self newSocket.
socket nextPutAll: Time now printString.
socket nextPutAll: aMessage.
socket newLine.
RemoteLogger»newSocket
"...."
"creates an instance of socket given some configuration"
6

7. MainApp>>start
logger := StdoutLogger new.
logger log: ’Application has started’.
"do something"
logger log: ’Application has finished’.
StdoutLogger»newLine
stdout newLine.
StdoutLogger»log: aMessage
stdout nextPutAll: Time now printString.
stdout nextPutAll: aMessage.
stdout newLine.
RemoteLogger»log: aMessage
| socket |
socket := self newSocket.
socket nextPutAll: Time now printString.
socket nextPutAll: aMessage.
socket newLine.
RemoteLogger»newSocket
"...."
"creates an instance of socket given some configuration"
Log Lib
Base Libs
Logger <<interface>>
StdoutLogger
log:
newLine
App
MainApp
start RemoteLogger
log:
newSocket
Date Time :stdout Socket
7

8. MainApp>>start
logger := StdoutLogger new.
logger log: ’Application has started’.
"do something"
logger log: ’Application has finished’.
StdoutLogger»newLine
stdout newLine.
StdoutLogger»log: aMessage
stdout nextPutAll: Time now printString.
stdout nextPutAll: aMessage.
stdout newLine.
RemoteLogger»log: aMessage
| socket |
socket := self newSocket.
socket nextPutAll: Time now printString.
socket nextPutAll: aMessage.
socket newLine.
RemoteLogger»newSocket
"...."
"creates an instance of socket given some configuration"
Log Lib
Base Libs
Logger <<interface>>
StdoutLogger
log:
newLine
App
MainApp
start RemoteLogger
log:
newSocket
Date Time :stdout Socket
8

9. MainApp>>start
logger := StdoutLogger new.
logger log: ’Application has started’.
"do something"
logger log: ’Application has finished’.
StdoutLogger»newLine
stdout newLine.
StdoutLogger»log: aMessage
stdout nextPutAll: Time now printString.
stdout nextPutAll: aMessage.
stdout newLine.
RemoteLogger»log: aMessage
| socket |
socket := self newSocket.
socket nextPutAll: Time now printString.
socket nextPutAll: aMessage.
socket newLine.
RemoteLogger»newSocket
"...."
"creates an instance of socket given some configuration"
Log Lib
Base Libs
Logger <<interface>>
StdoutLogger
log:
newLine
App
MainApp
start RemoteLogger
log:
newSocket
Date Time :stdout Socket
9

10. 10
• Core Libraries take place and are distributed with the language
• Runtime objects take place also
• And some of those are runtime meta-data, used for reflection
But there is not only bloat in
application code
Mariano Martinez Peck, Noury Bouraqadi, Marcus Denker, Stéphane
Ducasse, and Luc Fabresse.
Visualizing objects and memory usage. In Smalltalks 2010

11. Q: Can we tailor the runtime?
Q: How small can they be?
11
Research questions

12. a runtime virtualisation infrastructure
+
a run-fail-grow algorithm
12
Our solution in a nutshell

13. Reference
Application
Nurtured
Application
1. run
3. copy missing
2. fail 4. resume
Code Units
Tornado Run-Fail-Grow
Virtualized Runtime
hypervisor
(empty)
13

14. Reference
Application
Nurtured
Application
1. run
3. copy missing
2. fail 4. resume
Code Units
Tornado Run-Fail-Grow
hypervisor
Virtualized Runtime
14

15. Reference
Application
Nurtured
Application
1. run
3. copy missing
2. fail 4. resume
Code Units
Tornado Run-Fail-Grow
hypervisor
Virtualized Runtime
15

16. Reference
Application
Nurtured
Application
1. run
3. copy missing
2. fail 4. resume
Code Units
Tornado Run-Fail-Grow
hypervisor
Virtualized Runtime
16

17. Reference
Application
Nurtured
Application
1. run
3. copy missing
2. fail 4. resume
Code Units
Tornado Run-Fail-Grow
hypervisor
Virtualized Runtime
17

18. Reference
Application
Nurtured
Application
1. run
3. copy missing
2. fail 4. resume
Code Units
Tornado Run-Fail-Grow
hypervisor
Virtualized Runtime
18

19. How small can we get?
Experiment Size (KB) %Saved
Addition 11 99.99%
Reflective App 32 99.83%
Factorial 100 + I/O 89 99.39%
Seaside Counter 573 96.73%
19

20. Comparison with
Traditional Tailoring Solutions
criteria
Dedicated
Platforms
Static/Hybrid
Analysis
Dynamic
Analysis
Tornado
20

21. Comparison with
Traditional Tailoring Solutions
criteria
Dedicated
Platforms
Static/Hybrid
Analysis
Dynamic
Analysis
Tornado
Base libraries
Third-party libraries
Legacy Code
Reflection Support
General Purpose
Infrastructure
Configurability
Dynamic Typing
Minimality
Completeness
21

22. Comparison with
Traditional Tailoring Solutions
criteria
Dedicated
Platforms
Static/Hybrid
Analysis
Dynamic
Analysis
Tornado
Base libraries
Third-party libraries
Legacy Code
Reflection Support
General Purpose
Infrastructure
Configurability
Dynamic Typing
Minimality
Completeness
22

23. • Run-fail-grow approach for tailoring
•Works on dynamic languages
• Small memory footprint results (10k!)
Run-Fail-Grow: Creating
Tailored Object-Oriented
Runtimes
G. Polito, L. Fabresse, N. Bouraqadi, S. Ducasse,
in Journal of Object Technology
Guille Polito
@guillep
https://guillep.github.io/23