Conference PaperPDF Available

A SIMULATED ANNEALING ALGORITHM FOR REGISTER ALLOCATION

Authors:

Abstract and Figures

In this paper a simulated annealing algorithm for register allocation is presented. The algorithm is based on the Metropolis procedure. The procedure accepts a new solution with less profit based on a probabilistic function. The objective of the heuristic is to color a graph representing the register allocation problem to maximize the profit function. The profit function is defined using the profit gained from coloring a subset of the nodes. The algorithm presented in the paper has a linear time asymptotic complexity. The experimental results show optimal solutions in many of the graphs used for testing.
Content may be subject to copyright.
A SIMULATED ANNEALING ALGORITHM FOR REGISTER
ALLOCATION
K. M. Elleithy* and E. G. AbdelFattah**
* Computer Engineering Department, King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia
** Computer Department, Moshmit Biherah Company, Alexnadria, Egypt
ABSTRACT: In this paper a simulated annealing algorithm for register allocation is
presented. The algorithm is based on the Metropolis procedure. The procedure
accepts a new solution with less profit based on a probabilistic function. The
objective of the heuristic is to color a graph representing the register allocation
problem to maximize the profit function. The profit function is defined using the
profit gained from coloring a subset of the nodes. The algorithm presented in the
paper has a linear time asymptotic complexity. The experimental results show
optimal solutions in many of the graphs used for testing.
1. INTRODUCTION
Register allocation may be viewed as a graph coloring problem, where coloring a graph is an
assignment of a color to each of its nodes such that two connected nodes have different
colors. Each node in the graph stands for a computed quantity that resides in a machine
register. Two nodes are connected by an edge if the quantities are simultaneously alive at
some point in the object program.
The graph coloring problem is NP-C problem. Exhaustive search approaches for coloring a
graph imply an exponential time complexity. Heuristics of reasonable time complexity are
required for large size problems. A number of factors can be considered in designing a graph
coloring heuristics, e.g., a vertex of high degree is harder to color than a vertex of low degree,
vertices with the same neighborhood should be colored alike, and coloring many vertices with
the same color is a good idea.
In the last decade simulated annealing (SA) has been used extensively in solving various
combinatorial problems, e.g., traveling salesman problem [1], computer generated
holograms[2], RF power efficiency[3], heat exchangers[4], and allocation[5]. SA is an
adaptive heuristic and belongs to the class of non-deterministic algorithms. Kirkpatrich,
Gelatt and Vecchi first introduced the heuristic in 1983 [6]. The name of Simulated Annealing
is inspirited from the process of carefully cooling molten metals in order to obtain a good
crystal structure. During annealing, a metal is heated to a very high temperature, where the
molecules will have a greater degree of freedom to move, and then slowly cooled. By cooling
the metal at a proper rate, atoms will have an increased chance to regain proper crystal
structure. The movement of molecules is analogous to the generation of new neighborhood in
an optimization process. The objective function is non-convex; this means that it has multiple
maxima. A greedy iterative improvement algorithm starts off with an initial solution
S
0
and
climbs along the curve to find a local maximum L. The only way to find the global maximum
G, is to slide the hill along the local maximum L, which is a characteristic of simulated
annealing algorithm.
In this paper a new non-deterministic algorithm based on simulated annealing is presented.
Experimental results are discussed in the paper.
2. THE ALGORITHM
In this section a simulated annealing algorithm for register allocation is presented. The
algorithm is based on the Metropolis procedure. The procedure accepts a new solution with
less profit based on a probabilistic function. The objective of the heuristic is to color a graph
representing the register allocation problem to maximize a profit function. The profit function
is defined using the profit gained from coloring a subset of the nodes. The algorithm
presented in the paper has a linear time asymptotic complexity.
The algorithm is shown in Figure 1. The core of the algorithm is the Metropolis procedure
given in Figure 2, which uses a procedure (NEWSOLUTION) to generate a local neighbor
new_solution for any given solution. If the profit of the new_solution is greater than the profit
of the current solution, then certainly the new_solution is acceptable. If the new_solution has
less profit, Metropolis will accept the new_solution on a probabilistic basis. A random
number is generated in the range 0 to 1. If this random number is smaller than
e
delta temperature/
, where delta is the difference in profits, then the inferior solution is
accepted.
The objective of the heuristic is to color a graph representing the allocation problem to
maximize a profit function. The profit function is defined as follows:
Profit
)()(
11
nodesuncoloredall
i
i
nodescoloredall
i
i
nodenetsavenodenetsave
(1)
netsave node
i
( )
: A function representing the profit gained from allocating a variable to a
register.
The heuristic proceeds as follows:
1- Find the set of largest profitable nodes. The number of nodes in this set is equal to the
number of available colors.
2- Color a new node from the uncolored set if it is possible. If there is no more nodes that
can be colored or time >= maxtime then stop.
3- Calculate the change in profit after adding the colored node from step 2.
4- If the change is positive then add the node of step 2 to the colored set, otherwise the node
is colored based on a probabilistic comparison.
5- Go to step 2.
Example
Apply the algorithm to the graph of Figure 3 using the following parameters:
No of nodes = 10 Alpha = 0.8
No of colors = 4 Beta = 1
Cooling Schedule = 1 Temperature = 100
Procedure Simulated Annealing (initial_solution, initial_temperature, alpha, beta, cooling_schedule,
Maxtime);
{alpha is the cooling rate < 1};
{beta is a constant > 1};
begin
temperature = initial_temperature;
solution = initial_solution;
time = 0;
repeat
call Metropolis (solution, temperature, cooling_schedule);
time = time + cooling_schedule;
temperature = alpha * temperature;
cooling_schedule = beta * cooling_schedule;
until (time >= Maxtime);
end {of Simulated Annealing}.
Figure 1: Procedure for Simulated Annealing Algorithm.
Procedure Metropolis (solution, temperature, cooling_schedule);
begin
repeat
call NEWSOLUTION;
delta = new_profit - old_profit
If ((delta > 0) or (random <
e
delta temperature/
)) then
solution = new_solution;
cooling_schedule = cooling_schedule - 1;
until (cooling_schedule = 0);
end {of Metropolis}.
Figure 2: The Metropolis Procedure.
The netsave for different nodes is as follows:
Node #
1
2
3
4
5
6
7
8
9
Netsave
10
-30
-10
40
20
-5
50
-15
20
The algorithm proceeds as follows:
Step 1 : The algorithm assigns the four available colors to the nodes which have maximum
netsave as follows :
v
10
is assigned color1
v
4
is assigned color3
v
7
is assigned color2
v
5
is assigned color4.
1
2
3
4
5
6
7
8
9
10
Figure 3: A Graph for Testing Coloring by Simulated Annealing.
Step2 : The least uncolored node is
v
1
, which has positive netsave, the first colored node is
node
v
4
with color3, this color can be assigned to
v
1
, since
v
1
is not connected to
v
4
.
Step3 : Although
v
2
has negative netsave, it is assigned color4 (it is not assigned color3, since
it is connected to
v
1
(color3)). This means that metropolis accepts the inferior solution (i.e.
the random number generated is less than
e
delta temperature/
).
Step4 :
v
3
can be colored by color2, but
v
3
has negative netsave. Metropolis rejects the
solution (i.e. the random number generated is greater than
e
delta temperature/
), that is
why
v
3
is not colored.
Step5 :
v
6
is assigned color2, although it has negative netsave. This means that metropolis
procedure accepts the inferior solution.
Step6 : No possible color is available for
v
8
, since it is connected to nodes colored by all
different available colors.
v
9
has positive netsave, it is assigned color1, since it is
connected to
v
4
(color3),
v
5
(color4), and
v
6
(color2).
Step7 : The best solution is the one obtained in step2 (assignment #1).
3. EXPERIMENTAL RESULTS
In this section we apply the simulated annealing algorithm for register allocation on a graph
of size 128. The graph used for this study is of random nature and random values for the
nodes' netsave are generated. The effect of different parameters on the proposed algorithm
performance is evaluated. The effect of the following parameters is considered: number of
colors, cooling-schedule, beta, temperature, and alpha.
3.1 Number of Colors
To evaluate the effect of number of colors on the profit gained from coloring the graph and
the number of trials to get the best solution, the number of colors has been changed from 10 to
20. The values for the parameters cooling-schedule, beta, temperature, and alpha are 1.0, 1.0,
100, and 0.8 respectively. Figure 4 shows the profit versus number of colors and Figure 5
shows the number of trials versus the number of colors.
Figure 4 The Profit Versus Number of
Colors.
Figure 5. The Number of Trials versus
the Number of Colors
The following observations are concluded:
1- The best solution achieved is 2965. This value is constant for colors greater than or equal
to 14.
2- While increasing the number of colors (i.e. greater than 14), the profit remains constant
but the number of trials decreases. The effect of increasing the number of colors is
reflected on decreasing the number of trials.
3- The graph can be colored using 14 colors only while achieving a netsave of 2965.
3.2 Cooling Schedule (m)
To evaluate the effect of cooling schedule on the profit gained from coloring the graph and
the number of trials to get the best solution, the cooling schedule has been changed from 1 to
20. The values for the parameters beta, temperature, and alpha are 1.0, 100, and 0.8
respectively. Figure 6 shows the profit versus cooling schedule and Figure 7 shows the
number of trials versus cooling schedule. The following is noticed:
1- The best solution is obtained at m = 1, and for m > 1 the solution may be less than or
equal to the solution at m = 1.
2- If the best solution remains constant while m is increased, the number of trials remains
constant as well.
3- If the solution decreases as m increases, the number of trials decreases as well.
4- The profit does not change with m if the number of colors is greater than a certain
threshold value (14 in this example).
5- As m increases there is a better chance for a negative solution to be accepted. This implies
that the solution may be worse or better than the previous solution. For example if the
number of colors is 10 and m is increased from 4 to 5, the profit drops from 2565 to 2415
and the number of trials drops from 107 to 99, and if the number of colors is 13 and m
increases from 9 to 10, the profit increases from 2685 to 2865.
2500
2550
2600
2650
2700
2750
2800
2850
2900
2950
3000
8 13 18 23
Profit
colors
105
107
109
111
113
115
117
8 13 18 23
Trials
Colors
Figure 6: The Profit versus Cooling Schedule.
Figure 7: The Number of Trials Versus Cooling Schedule.
3. 3 Beta
To evaluate the effect of beta on the profit gained from coloring the graph and the number
of trials to get the best solution, beta has been changed from 1 to 20. The values for the
parameters cooling-schedule, temperature, and alpha are 1.0, 100, and 0.8 respectively.
Figure 8 shows the profit versus Beta and Figure 9 shows the number of trials versus Beta.
The following is noticed:
2300
2400
2500
2600
2700
2800
2900
3000
0 5 10 15 20 25
Profit
M
10
11
12
13
14
15
20
96
98
100
102
104
106
108
110
112
114
116
0 5 10 15 20 25
Trials
M
10
11
12
13
14
15
20
No of
Colors
No of
Colors
1- The best solution is obtained at low values for beta. This value is equal to 1 except for
number of colors equal to 12 and 13, the best solution is obtained at beta equal to 3.
2- If the solution remains constant while beta is increased, the number of trials may decrease
slightly ( number of colors is 15 ), or remains constant ( number of colors is 20 ).
3- If the solution decreases as beta increases, the number of trials decreases as well.
4- The profit does not change with beta if the number of colors is greater than a certain
threshold value (14 in this example).
5- As beta increases there is a better chance for a negative solution to be accepted. This
implies that the solution may be worse or better than the previous solution. For example if
the number of colors is 11 and beta is increased from 1 to 2, the profit drops from 2705 to
2555 and the number of trials drops from 109 to 101, but if the number of colors is 12 and
beta increases from 2 to 3, the profit increases from 2695 to 2785.
Figure 8: The Profit versus Beta.
Figure 9: The Number of Trials Versus Beta.
3.4 Temperature
2300
2400
2500
2600
2700
2800
2900
3000
0 5 10 15 20 25
Profit
Beta
10
11
12
13
14
15
20
90
95
100
105
110
115
120
1
3
5
7
9
11
13
15
17
19
Trials
Beta
10
11
12
13
14
15
20
No of
Colors
No of
Colors
To evaluate the effect of temperature on the profit gained from coloring the graph and the
number of trials to get the best solution, the temperature has been changed from 1 to 100. The
values for the parameters cooling-schedule, beta, and alpha are 1.0, 1.0, and 0.8 respectively
Figure 10 shows the profit versus temperature and Figure 11 shows the number of trials
versus temperature The following was noticed :
1- The best solution is obtained at low values of temperature.
2- If the solution remains constant while temperature is increased, the number of trials
remains constant as well.
3- If the solution decreases as temperature increases, the number of trials decreases as well.
4- The profit does not change with temperature if the number of colors is greater than or
equal to a certain threshold value (14 in this example).
5- As temperature increases there is a better chance for a negative solution to be accepted.
This implies that the solution may be worse or better than the previous solution. If the
number of colors is equal to 11 and temperature is increased from 10 to 20 the profit
increases from 2885 to 2895, but if the temperature is increased from 20 to 30 the profit
drops from 2895 to 2835.
3.5 Alpha
To evaluate the effect of alpha on the profit gained from coloring the graph and the number of
trials to get the best solution, alpha has been changed from 0.1 to 0.99 The values for the
parameters cooling-schedule, beta, and temperature are 1.0, 1.0, and 100 respectively. Figure
12 and 13 show the profit versus Alpha and Figure 14 shows the number of trials versus
Alpha. The following is noticed :
1- The best solution is obtained at low values for alpha. This value is less than or equal to
0.6, except for number of colors equal to 11 the best solution is obtained at alpha in the
range 0.2 to 0.6.
2- If the solution remains constant while alpha is increased, the number of trials remains
constant as well.
3- If the solution decreases as alpha increases, the number of trials decreases as well.
4- The profit does not change with alpha if the number of colors is greater than or equal to
14, except for the values of alpha in the range .95 to .99 where a significant change is
noticed.
5- As alpha decreases there is no chance that a negative solution is accepted. That explains
why the solution remains constant for alpha less than or equal to 0.6.
6- The profit achieved is very sensitive for values of alpha greater than 0.95.
Figure 10: The Profit Versus Temperature.
2500
2600
2700
2800
2900
3000
0 10 20 30 40 50 60 70 80 90 100 110
Profit
Temperature
10
11
12
13
14
15
20
No of
Colors
Figure 11: The Number of Trials versus Temperature
Figure 12: The Profit versus Alpha.
Figure 13: The Profit Versus Alpha (Alpha > 0.9 )
100
102
104
106
108
110
112
114
116
118
0 10 20
Trials
Temperature
10
11
12
13
14
15
20
2000
2200
2400
2600
2800
3000
3200
0 0.2 0.4 0.6 0.8 1 1.2
Profit
Alpha
10
11
12
13
14
15
20
2200
2300
2400
2500
2600
2700
2800
2900
3000
3100
0.89 0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 1
Profit
Alpha
10
11
12
13
14
15
20
No of
Colors
No of
Colors
No of
Colors
Figure 14: The Number of Trials Versus Alpha.
CONCLUSIONS
In this paper a simulated annealing algorithm for register allocation was presented. The
algorithm uses the Metropolis procedure to accept a new solution with less profit based on a
probabilistic function. Random generated graphs were used to thoroughly study the behavior
of the algorithm parameters. The experimental results show optimal solutions in many of the
graphs used for testing. The algorithm presented in the paper has a linear time asymptotic
complexity
ACKNOWLEDGEMENTS
The first author of this paper would like to acknowledge the support of King Fahd University
of Petroleum and Minerals in preparation and presentation of this paper.
REFERENCES
[1] Peng, M., Gupta, N. K. and Armitage, A., "An investigation into the improvement of local minima of the
Hopefield network", Neural networks, Vol. 9, 1241-1253, Oct. 1996.
[2] Taniguchi, M., Kurokawa, K. and Itoh, K., "Sidelobeless multiple-object discriminant filters recorded as
discrete-type computer generated holograms", Applied optics, Vol. 36, 9138-9145, Dec. 1997.
[3] Wilson, J., "A simulated annealing algorithm for optimizing RF power efficiency in coupled-cavity traveling-
wave tubes", IEEE Transactions on Electron Devices, Vol. 44, 2295-2299, Dec. 1997.
[4] Athier, G., Floquet, P. and Pibouleau, L., "Synthesis of heat-exchanger network by simulated annealing and
NLP procedures", AIChE Journal, Vol. 43, 3007-3020, Nov. 1997.
[5] Duncan, A. A. and Hendry, D. C., "High-Level synthesis of DSP datapaths by global optimization of variable
lifetimes", IEE proceedings: Computers and Digital Techniques, Vol. 142, 215-224, May 1995.
[6] Kirkpatrich, S., Gelatt, J. C., and Vecchi, M., "Optimization by Simulated Annealing" Science, Vol. 220,
498-516, May 1983.
90
95
100
105
110
115
120
0 0.2 0.4 0.6 0.8 1 1.2
Trials
Alpha
10
11
12
13
14
15
20
No of
Colors

** 







ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
There is a deep and useful connection between statistical mechanics (the behavior of systems with many degrees of freedom in thermal equilibrium at a finite temperature) and multivariate or combinatorial optimization (finding the minimum of a given function depending on many parameters). A detailed analogy with annealing in solids provides a framework for optimization of the properties of very large and complex systems. This connection to statistical mechanics exposes new information and provides an unfamiliar perspective on traditional optimization problems and methods.
Article
Full-text available
Decreasing the radio frequency (RF) phase velocity in the output section of a traveling-wave tube (TWT) is a technique commonly used to increase RF power efficiency. In order to optimize the profile of the phase velocity, a simulated annealing algorithm has been developed and implemented into the NASA multidimensional large-signal coupled-cavity TWT computer model. This algorithm allows the determination of the lengths of the individual cavities at the end of the output section necessary to provide the optimized phase velocity profile. The resulting nonlinear computer-generated phase velocity profile provides a design with optimized RF efficiency. In this paper, the optimization algorithm is described and computational results are shown. These results indicate an increase in center-frequency RF efficiency from 7.1 to 13.5% for a V-band coupled-cavity TWT
Article
A two-level procedure for heat-exchanger network synthesis problems, which is related to structural optimization, is carried out by a simulated-annealing procedure. The main control parameters, that is, the initial annealing temperature, the temperature length, and the temperature-decreasing scheme, are extensively studied on the well-known 10SP1 problem, and their values are used for solving a 5SP1 problem with forbidden matches and a larger scale problem, the 20SP1. In all cases, the structural search is initialized with the network involving only utilities. For each generated network, the operating conditions (temperatures and split rates) are optimized by an NLP procedure to minimize the sum of investment and operating costs. Particular attention is given to the size reduction of the NLP problem. The proposed procedure leads to a reduction by a factor of 3 of the problem size in terms of number of variables and constraints. This slave problem is solved by an SQP procedure from the IMSL Library.
Article
The paper investigates the improvement of local minima of the Hopfield network. A local minima escape algorithm (LME algorithm), is proposed for improving local minima of small-scale networks. Experiments on travelling salesman problems (TSP) show that the LME algorithm is an efficient algorithm in improving the local minima, and the comparison with the simulated annealing algorithm (SA) shows that the LME algorithm can produce better results in less time. The paper then investigates the improvement of local minima of large-scale networks. By combining the LME algorithm with a network partitioning technique, a network partitioning algorithm (NPA) is proposed. Experiments on 51 and 101-city TSP problems show that the local minima of large-scale networks can be greatly improved by the NPA algorithm, however, the global minima are still difficult to achieve. Copyright 1996 Elsevier Science Ltd
Article
The computer generation of sidelobeless multiple-object discriminant correlation filters has been stressed. We propose to synthesize the filter functions by use of the simulated-annealing algorithm. By this method the filters can be obtained as discrete-type computer-generated holograms. The filters can suppress the sidelobes and provide sharp correlation peaks. A computer simulation and an optical experiment were performed, and the expected correlation responses were obtained. (C) 1997 Optical Society of America.
COBRA (Column-Oriented Butted Regular Architecture) is a behavioural high-level synthesis tool for datapath-dominated applications. It globally optimises the synthesised datapath by performing the scheduling and allocation tasks simultaneously. COBRA uses a bit-sliced target architecture and layout style which, when compared with conventional approaches, significantly reduces the area of the final datapaths. The synthesis problem is formulated as an optimisation problem on the configuration of variable lifetimes when mapped into a 3D `datapath space'. The configuration of the data in the datapath space implies the structure required to achieve the data configuration and hence the datapath. Simulated annealing is used to optimise the solution. A description is given of the target architecture, the mapping of the input description into the datapath space, the optimisation of the data configuration in the datapath space, and the post-processing operations. Results for a number of examples are presented