running the following:
tiobench -b 65536 -t 8 -d /mnt/test -s 8192
results in 8GB worth of sequential I/O, followed by only 500 * 31MB * 8 threads of random I/O. This makes the numbers appear skewed as there aren't enough random samples to be meaningful (especially considering that some of the data from the sequential read test may still be in cache if we are lucky enough). tiobench.pl should really do as much random I/O as sequential.