before:

==8812== Total:     2,374,977,159 bytes in 6,840,026 blocks
==8812== At t-gmax: 8,090,486 bytes in 3,389 blocks
==8812== At t-end:  3,185,454 bytes in 757 blocks
==8812== Reads:     1,873,472,286 bytes
==8812== Writes:    1,249,411,589 bytes

==11212== I   refs:      6,370,244,180

after:

==18725== Total:     429,769,246 bytes in 957,259 blocks
==18725== At t-gmax: 8,058,316 bytes in 3,502 blocks
==18725== At t-end:  3,045,261 bytes in 1,097 blocks
==18725== Reads:     431,872,599 bytes
==18725== Writes:    214,738,653 bytes

==20839== I   refs:      1,873,010,089

Headers WAY less than 1kb anyway, so this can be improved more?

before

==18725== Total:     429,769,246 bytes in 957,259 blocks
==18725== At t-gmax: 8,058,316 bytes in 3,502 blocks
==18725== At t-end:  3,045,261 bytes in 1,097 blocks
==18725== Reads:     431,872,599 bytes
==18725== Writes:    214,738,653 bytes

after

==49344== Total:     201,418,575 bytes in 957,174 blocks
==49344== At t-gmax: 7,937,250 bytes in 3,310 blocks
==49344== At t-end:  3,035,637 bytes in 1,076 blocks
==49344== Reads:     431,607,448 bytes
==49344== Writes:    210,731,540 bytes

keeping order of inserted Paths having high cost on hot path, collect into HashSet instead and sort afterward.

from 1,858,963,938 to 1,448,975,825 I refs.