Optimizing Parallel Processes
Parallel Algorithms
David Mayerich
Scalable Tissue Imaging and Modeling (STIM) Laboratory
Department of Electrical and Computer Engineering
Cullen College of Engineering
University of Houston
David Mayerich
STIM Laboratory, University of Houston
Static Assignment
-
Allocate work before assigning it
-
Specify rules at compile-time
-
Assignment remains constant
-
Useful when:
-
Work is predictable
-
or statistically predicatable
-
David Mayerich
STIM Laboratory, University of Houston
t0
t1
t2
t3
\sum
t0
t1
t2
t3
t4
t8
t5
t9
t6
t10
t7
t11
t12
t13
t14
t15
Semi-Static Assignment
-
Determine the short-term cost of work
-
Assign workers based on that cost
-
Barnes-Hutt Algorithm
David Mayerich
STIM Laboratory, University of Houston
Dynamic Assignment
-
Prime test using trial division:
David Mayerich
STIM Laboratory, University of Houston
for(size_t i = 0; i < N; i++) {
p[i] = test_prime(x[i]);
}
// test prime using trial division
bool test_prime(int y) {
int m = sqrt(y);
for(size_t j = 2; j < m; j++) {
if(y % j == 0) return false;
}
return true;
}30
31
\cdots
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
\cdots
9
11
13
10
12
test_prime
duration
15
17
19
29
14
18
20
16
\cdots
queue
t0
t1
t2
t3
Assess Assignment Strategies
-
Profile the entire program
-
Calculate the overhead:
-
\(T_o\) is the overhead (time spent on non-useful work)
-
\(p\) is the number of processors
-
\(T_p\) is the parallel runtime
-
\(T_s\) is the serial runtime
-
-
Successful strategies
David Mayerich
STIM Laboratory, University of Houston
T_o = p T_p - T_s
p_0
p_1
p_2
p_3
p_4
p_5
s_1
s_0
sync()
sync()
sync()
p_0
p_1
p_2
s_0
p_3
p_4
p_5
s_1
serial
parallel
\(T_p\)
\(T_s\)
D.3 Parallel Optimization
By STIM Laboratory
