A Macrotask-level Unlimited Speculative Execution on Multiprocessors
(18/19)
The following description is based on the paper presented at
1995 ACM International Conference on Supercomputing head
in Spain on July 3rd to 7th,1995.
If you want to have the copy of the paper, please
let me know.
The left hand side graph shows the speedup against the TASK_SIZE.
The distributed control always archives good performance in comparison
with the centralized control. Moreover, the proposed scheme can perform
nearly the maximum of the theoretical speedup when the TASK_SIZE is enough
to be large.
Since the average execution time of macrotasks at second hierarchy
level from the bottom is between 150us to 300us, the software
implementation of the proposed scheme is able to be adopted at the
second hierarchy level.
The anothergraph shows the speedup against the number of running MT.
The proposed scheme always achieve good performance in comparison with
the centralized control. The speedup ratio increases when the number
of running MT is less than 40, because the speculation depth increases
in proportion to the number of iterations. After 40, the speculation
depth is balanced with the overhead of the macrotask control. Then it
holds same speedup.
As for the centralized control, the speedup decreases because the
overhead to control the MT increases in proportion to the number of
macrotasks.
These results show that the proposed scheme has small macrotask
control overhead in comparison with the centralized control.