Next: Asymmetric compression
Up: Results
Previous: Understanding cache behavior
It has been noted that when a platform has a low-power idle state, it
may be sensible to sacrifice energy in the short-term in order to
complete an application quickly and enter the low-power idle state
[26]. Figure 8 shows the effect
of this analysis for compression and sending of text.
Receive/decompression exhibits similar, but less-pronounced variation
for different idle powers. It is interesting to note that, assuming a
low-power idle mode can be entered once compression is complete, one's
choice of compression strategies will vary. With its 1Watt
of idle power, the Skiff would benefit most from zlib
compression. A device which used negligible power when idle would
choose the LZO compressor. While LZO does not
compress data the most, it allows the system to drop into low-power
mode as quickly as possible, using less energy when long idle times
exist. For web data (not shown due to space constraints) the
compression choice is LZO when idle power is low. When idle
power is one Watt, bzip2 energy is over 25% more energy
efficient than the next best compressor.
Figure 8:
Compression + Send energy consumption with varying sleep
power (Text data)
|
|
Next: Asymmetric compression
Up: Results
Previous: Understanding cache behavior
Kenneth Barr
2003-03-04