I finally get back to the numbers that I had couple weeks ago. I will
summarize it and see what seems to be contradicted.

You can find hbench-OS from

http://www.eecs.harvard.edu/~vino/perf/hbench/#download

it's a patched a version of lmbench. hbench-OS did fix some bugs that
found in lmbench.

For memory read, I get

  100MB/s 2K <= size <= 8K
   27MB/s 16K <= size <= 4M

For memory write, I get

  155MB/s 2K <= size <= 8K
   20MB/s 16K <= size <= 4M

So, a reasonable prediction for memory copy performance would be

   50MB/s 2K <= size <= 8K
   10MB/s 16K <= size < 4M

Because a copy takes one read and one write. Some architectures with
special instructions can do better than this.

Now, turn to the number of copy I get.

             libc      libc        unrolled   unrolled
             aligned   unaligned   aligned    unaligned

all size     3.5MB/s   3.5MB/s     3.1MB/s    9MB/s

I did not have the number of bzero right at the last message about this,
it should be 20MB/s.

I would like to investigate into the problems when I have more time.

// Shawn