Effect of overhead on energy consumption

To evaluate the effect of RightSpeed overhead on energy consumption, we ran some workloads on the Transmeta machine both with and without RightSpeed. To equalize performance, we instructed RightSpeed to not use the PACE calculator but instead use an algorithm identical to Transmeta's LongRun $^{\text {TM}}$ strategy. Table 5 shows the results for five short workloads derived from VTrace traces. We observe that the performance characteristics (deadlines missed and total delay) of RightSpeed mimicking LongRun $^{\text {TM}}$ are very close to that of LongRun $^{\text {TM}}$ by itself, so it is meaningful to directly compare the energy consumption of the two. We see that simulating LongRun $^{\text {TM}}$ with RightSpeed has little effect on the total energy consumption. In other words, the overhead of signaling the beginnings and ends of tasks, and of implementing the speed schedule in software instead of hardware is insignificant.

Table 5: Comparison of using built-in LongRun $^{\text {TM}}$ scheduling versus doing this scheduling with RightSpeed

							Relative
	Without RightSpeed			With RightSpeed mimicking LongRun $^{\text {TM}}$			energy
Workload	Deadlines missed	Delay	Energy	Deadlines missed	Delay	Energy	increase
A	21 out of 30,635	0.947 s	172.8 J	21 out of 30,635	0.940 s	174.0 J	+0.7%
B	8 out of 19,310	1.172 s	94.58 J	8 out of 19,310	1.172 s	94.86 J	+0.2%
C	1,792 out of 32,288	118.833 s	1007 J	1,796 out of 32,288	118.860 s	1006 J	-0.1%
D	61 out of 19,770	2.364 s	126.8 J	61 out of 19,770	2.366 s	126.8 J	0.0%
E	99 out of 20,641	4.070 s	379.1 J	99 out of 20,641	4.061 s	376.1 J	-0.8%