Some interesting results, me thinks: Writing is exorbitant costly, reading is nearly equivalent (looking at the wallclocks real secs are a somewhat different matter ;-).

System: P4 2.4GHz, SuSE 8.1, perl 5.8.0 (Edit:512MB)

Benchmark: timing 100 iterations of berkeley write, file write...
berkeley write: 23 wallclock secs (12.55 usr +  9.56 sys = 22.11 CPU) 
+@  4.52/s (n=100)
file write: 59 wallclock secs (11.78 usr + 44.22 sys = 56.00 CPU) @  1
+.79/s (n=100)
                 Rate     file write berkeley write
file write     1.79/s             --           -61%
berkeley write 4.52/s           153%             --
Benchmark: timing 100 iterations of berkeley read, file read...
berkeley read:  8 wallclock secs ( 3.70 usr +  4.78 sys =  8.48 CPU) @
+ 11.79/s (n=100)
 file read:  7 wallclock secs ( 3.00 usr +  3.15 sys =  6.15 CPU) @ 16
+.26/s (n=100)
                Rate berkeley read     file read
berkeley read 11.8/s            --          -27%
file read     16.3/s           38%            --
[download]

regards,
tomte