Contrary to popular belief, urandom is not free.  In generic testing, a simple dd using /dev/urandom performs at about half the performance compared to /dev/zero for fully synchronous I/O, about 1/20 for buffered I/O (i.e. the difference on the input side is mostly masked when waiting for disks on the output side).  There's probably we can do about read-perf without platform-specific flushing of kernel buffers, but for write-perf we can at least get more accurate and useful numbers by using /dev/zero instead.