1. Improving Performance Through
Object Lifetime Pro
fi
ling: the
DataFrame Case
Sebastian JORDAN MONTAÑO, Nahuel PALUMBO, Guillermo POLITO,
Stéphane DUCASSE and Pablo TESONE
Inria, Univ. Lille, CNRS, Centrale Lille, UMR 9189 - CRIStAL
August 2023 Evref
fervE

2. Memory management in software
int* ptr = (int*) malloc(sizeof(int));
free(ptr);
2

3. Pharo’s garbage collector
Eden S1 S2 Tenured
+
Young generation Old generation
3

4. GC parameters
4

5. Time spent on garbage collecting
5

6. Research question
How does approximate object lifetimes lead to GC performance
improvements?
6

7. An object’s lifetime
Object’s allocation
Object becomes
unreachable
GC collects
the object
Actual lifetime
Lifetime that we capture?
7

8. An object’s approximated lifetime
Object’s allocation
Object becomes
unreachable
GC prepares
the object
Actual lifetime
Lifetime that we capture
Object is
fi
nalized
(we take the measurement)
8

9. Capturing the allocations
Array new: 7
9

10. Capturing the allocations
Array new: 7
10
Capture the allocation
Register the
fi
nalization
MethodProxies [1]
Ephemerons [2]
[1] github.com/pharo-contributions/MethodProxies
[2] github.com/pharo-project/pheps/blob/main/phep-0003.md

11. An object’s finalization at a time m
11
Ephemeron
#111
Object: Array
#123
Model
allocationTime: n
finalizationTime: m
m
o
u
r
n
finalize
E
p
h
e
m
e
r
o
n
i
s
c
o
n
s
u
m
e
d
1 2
3
key
Finalization Queue
Object #123 is
garbage collected
4

12. An object’s allocation at a time n
12
OrderedCollection class >> new: anInteger
^ self basicNew setCollection:
(self arrayType new: anInteger)
Behavior >> basicNew
<primitive: 70>
Array class >> basicNew: size
lines := OrderedCollection new
...

13. Paper’s contributions
Challenges of lifetime pro
fi
ling
Illimani: a lifetime pro
fi
ler on stock Pharo VM
13

14. Methodology
14
Application’s Lifetime Profile
P
Application to Profile
3. Chose GC tuned
parameters based on
object lifetimes and
benchmark information
5. Did the performance improved?
GC tuned
parameters
2. Benchmark it with the
default GC parameters
4. Benchmark the application
again with the tuned GC parameters
1. Profile the application
Default GC parameters
benchmark
Tuned GC parameters
benchmark

15. The target application
https://github.com/PolyMathOrg/DataFrame
15

16. Benchmark the loading of DataFrame
16

17. Benchmark the loading of DataFrame
17

18. Object lifetimes for a 500MB DataFrame (memory)
-500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
0 B
9 B
116 B
1 KB
13 KB
145 KB
1 MB
16 MB
180 MB
1 GB
Lifetime in seconds
Memory
(log
scale)
60%
18

19. -500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
0
5
35
217
1,311
7,900
47,557
286,265
1,723,096
10,371,664
Lifetime in seconds
Number
of
objects
(log
scale)
75%
Object lifetimes for a 500MB DataFrame (# objects)
19

20. Common object lifetime distribution
20
Source: oracle.com

21. GC custom parameters
21

22. Pharo’s garbage collector
Eden S1 S2 Tenured
+
Young generation Old generation
22

23. Benchmarks results
23

24. Future work
Measure the precision of our approximate object lifetimes
Pro
fi
ling at VM level to reduce the overhead
Pre-tenuring
24

25. Summary
We developed a lifetime profiler
We profiled the object lifetimes and we validated our solution by
observing how lifetimes relate to performance improvements when
tuning the GC.
25
github.com/jordanmontt/illimani-memory-pro
fi
ler
Sebastian JORDAN MONTAÑO
sebastian.jordan@inria.fr