times have gotten worse since the USENIX paper.  Turns out we were operating
at lower BW than the paper (62/Mb sec vs. 70Mb) due to clock granularity.
The disk was falling idle too much.  Cranked it back up to 72Mb/sec for
"standard" (/usr/testbed) images.  Actually lowered it to about 54Mb/sec
for "user" images that have to be read across NFS (/proj).