I get 11100 keys/sec with Linux/gcc 2.5.8 on a 486/66, optimized with -O2 and -fomit-frame-pointer. -fforce-mem and -funroll-loops made the program run more slowly. I also tried -ffast-math -finline-functions -fforce-addr -fcaller-saves which seemed to make no difference. Now maybe when they get the server working again I can actually try running this. :(