Lecture Notes in Computer Science 4917

330 F. Vandeputte and L. Eeckhout
PhasePredictability Phase Predictability
1.0
0.8
0.6
0.4
0.2
0.0
0
0.5
1
1.5
2
Number of Phases
2.5
3
(log10)
3.5
4
1.0
0.8
0.6
0.4
0.2
0.0
0
0.5
1
1.5
2
Number of Phases
2.5
3
(log10)
3.5
4
gcc-scilab eon-kajiya
20
21
22
23
24
25
26
27
28
29
30 Interval Size (log2)
1.0
0.8
0.6
0.4
0.2
0.0
Phase Predictability
1.0
0.8
0.6
0.4
0.2
0.0
0
0.5
1
1.5
2
Number of Phases
2.5
3
(log10)
3.5
4
gzip-program gap
20
21
22
23
24
25
26
27
28
29
30 Interval Size (log2)
1.0
0.8
0.6
0.4
0.2
0.0
Phase Predictability
1.0
0.8
0.6
0.4
0.2
0.0
0
0.5
1
1.5
2
Number of Phases
2.5
3
(log10)
3.5
4
20
21
22
23
24
25
26
27
28
29
30 Interval Size (log2)
20
21
22
23
24
25
26
27
28
29
30 Interval Size (log2)
Fig. 3. Phase complexity surfaces for gcc-scilab (top left), eon-kajiya (top right),
gzip-program (bottom left) and gap (bottom right)
Figure 3 shows the phase complexity surfaces for gcc-scilab, eon-kajiya,
gzip-program and gap which combine the phase count and predictability surfaces.
These examples clearly show two extreme phase behaviors. The phase behavior for
eon-kajiya is much less complex than for gcc-scilab: eon-kajiya has fewer
program phases and shows very good phase predictability; gcc-scilab on the other
hand, exhibits a large number of phases and in addition, phase predictability is very
poor.
6 Classifying Benchmarks
Having characterized all the benchmarks in terms of their phase behavior using phase
complexity surfaces, we can now categorize benchmarks according to their phase behavior.
To this end we employ the methodology proposed by Eeckhout et al. [10] to find
similarities across benchmarks.
6.1 Methodology
As input to this methodology we provide a number of characteristics per benchmark:
we provide phase predictability and (the logarithm of) the number of phases at three
threshold values (θ =5%, θ = 10% and θ = 25%) at four time scales (1M, 8M, 64M
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0