Right now proper_unit_for_byte_size() and byte_size_in_proper_unit() are switching to higher prefix, once we have two significant digits, like this:

 inline const char* proper_unit_for_byte_size(size_t s) {
 #ifdef _LP64
   if (s >= 10*G) {
     return "G";
   }
 #endif

This is inconvenient for logs, because it abruptly makes values too coarse. For example, this is the usual thing to see in GC logs:

[25.186s][info][gc] Heap: 100G reserved, 15G (15.09%) committed, 15G (15.00%) used

This rounding basically masks the difference within the gigabyte, which makes the rounding error close to 10% in worst case, on 10G threshold It would be preferable instead to print more significant digits, e.g. at least three. With three significant digits, the rounding error would reach 1% at max, and checking against e.g. 100*G would yield five significant digits in worst case.

Prototype webrev:
 http://cr.openjdk.java.net/~shade/8217315/webrev.01/

This is the example GC log after the patch:

[23.315s][info][gc] Heap: 100G reserved, 15449M (15.09%) committed, 15361M (15.00%) used