Chip Implementations - Heterogeneous Multicore Processor Technologies for Embedded Systems - page 141

1.0

0.8

1 Thread
2 Threads
4 Threads
8 Threads
16 Threads

0.6

0.4

0.2

0.0

Water FFT LU Radix Barnes Ocean

Fig. 4.19 RP-2 execution time according to number of POSIX threads

2,000,000 200,000

Core0
Core1
Core2
Core3

-31%

-57%

-7%

1,500,000

150,000

1,000,000

100,000

500,000 50,000

0

0

Conventional Auto-rotating Conventional Auto-rotating Conventional Auto-rotating

Water Radix Barnes

Fig. 4.20 Number of acknowledged interrupts during SPLASH-2 execution

51-63% with two threads and to 41-27% with four or eight threads running on one
cluster. Since there were fewer cores than threads, the eight-thread case showed
similar performance to four-thread one. Furthermore, in some cases, the increase in
the number of threads resulted in an increase in the processing time due to the syn-
chronization overhead.
The autorotating dynamic interrupt distribution mode was evaluated and com-
pared to a conventional one by SPLASH-2 with four threads on SMP Linux using
one cluster of a real chip. Figure 4.20 shows the number of interrupts acknowl-
edged by the CPU cores during the SPLASH-2 execution. The total acknowledged
interrupts by all the cores in the autorotating mode decreased by 7% for Water, 31%
for Radix, and 57% for Barnes from the conventional mode. As a result, it avoided
the redundant interrupt handling. This improvement leads to a reduced processing
time in Linux kernel mode. Figure 4.21 shows the processing time reduction in