https://bugzilla.novell.com/show_bug.cgi?id=381873
Forwarding this because it is a GNU issue, not specifically a Novell one.
I reproduced this myself with the latest coreutils from git
(BTW: You might want to repack that repo, "counting objects" during the
clone was rather slow in the initial counting.)
Could it be a libiconv problem?
Glibc used: 2.8 (i586)
Coreutils: v6.11-62-g7c9c6c7
21:15 yaguchi:../rt/bin > LC_ALL=en_US time ./wc /tmp/ARC
3323886 30096741 537371570 /tmp/ARC
10.41user 1.86system 0:14.94elapsed 82%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (13major+177minor)pagefaults 0swaps
21:19 yaguchi:../rt/bin > LC_ALL=en_US.UTF-8 time ./wc /tmp/ARC
3323886 30096738 537371570 /tmp/ARC
87.21user 1.88system 1:38.89elapsed 90%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (2major+213minor)pagefaults 0swaps
So while ARC is definitely hot (cached) after the first run,
counting with the utf8 locale still takes up considerately more time.