Nah - your results are way too close. On slower hardware lexicals are clearly fastest, globals come in second and symbolic refs in last place. This is all easily explained in terms of involved opnodes. Symbolic refs add a handful of nodes, and globals are one opnode heavier than lexicals.

Benchmark: running use_global_hash, use_global_vars, use_lexical_hash,
+ each for at least 10 CPU seconds...
use_global_hash: 11 wallclock secs (10.45 usr +  0.00 sys = 10.45 CPU)
+ @ 112.63/s (n=1177)
use_global_vars: 11 wallclock secs (10.55 usr +  0.00 sys = 10.55 CPU)
+ @ 83.32/s (n=879)
use_lexical_hash: 11 wallclock secs (10.48 usr +  0.00 sys = 10.48 CPU
+) @ 120.71/s (n=1265)
                   Rate  use_global_vars  use_global_hash use_lexical_
+hash
use_global_vars  83.3/s               --             -26%             
+-31%
use_global_hash   113/s              35%               --             
+ -7%
use_lexical_hash  121/s              45%               7%             
+  --
[download]