Check out the new USENIX Web site. next up previous
Next: Buffer Allocation Up: Experiments on Disk Access Previous: Experiments on Disk Access

Accesses

In our first experiment we execute ijpeg and gqview concurrently, and each application demand fetches data from the disk. Ijpeg performs image compression on a set of ppm format image files. Each file is a copy of the same SPEC input (command line options: -GO.compress vigo.ppm) that is 2,359,355 bytes. When running unconstrained, ijpeg requires about 2.452 seconds to process each file and start to read the next. Gqview displays the same set of image files using autobrowse mode with a 50 second think time. We set the total power allocation to 1500mW and the two tasks each get an equal share of 750mW. These severe constraints are used to accentuate the disk's impact on performance. The entry price for initiating a disk access is set to 24000mJ (twice the combined cost of spinup and spindown). We use an immediate disk spindown.

Without pricing/bidding, the total average disk power consumption is 911mW, with 403mW and 508mW for ijpeg and gqview, respectively. Our currentcy-based policy formulated in terms of pricing/bidding reduces this value to 655mW (313mW ijpeg and 342 gqview) by engineering more task cooperation in disk spinup sessions. Furthermore, the performance of each application improves, particularly ijpeg which requires only 57 seconds to process the file compared to 74 without pricing/bidding.

The next experiment is designed to show the benefits of bidding for energy efficient disk prefetching. We set the total allocation at 1500mW and execute ijpeg (same input as above) with 300mW concurrently with the MP3 player, x11amp, which receives 1200mW allocation. X11amp reads a 3MB file, and is amenable to prefetching because of its sequential access pattern. We use the combined pricing/bidding approach where there is a high entry price for a disk spinup and x11amp contributes by bidding based on its prefetch buffer consumption. The average total disk power consumption is 357mW compared to 565mW without pricing/bidding. Ijpeg's average disk power consumption reduces from 365mW to 229mW and its performance improves from 90 seconds per file to 66 seconds. X11amp's disk power consumption reduces from 200mW to 128mW, and it does not incur any pauses in either policy.


next up previous
Next: Buffer Allocation Up: Experiments on Disk Access Previous: Experiments on Disk Access
Heng Zeng 2003-04-07