An old gripe: Reading via mmap stinks

[Mikhail T.]

01/14/10 21:56, Matthew Dillon написав(ла):
>      This would explain why the performance is not as bad as linux but
>      is not as good as a properly pipelined case.
For what it may be worth, here are the stats for Solaris as well:

    * Solaris 8, native, 32-bit binary (using -lcrypto instead of -lmd):
      mmap: 103.54u 27.18s 2:56.46 74.0%
      read: 99.12u 40.37s 2:53.06 80.6%
    * Solaris 10, native, 32-bit binary (using -lcrypto instead of -lmd):
      mmap: 159.36u 83.23s 5:28.25 73.9%
      read: 173.50u 104.16s 4:48.30 96.3%
    * Solaris 10, using the 32-bit binary built on Solaris-8:
      mmap: 217.74u 101.20s 5:58.89 88.8%

All of the "read" results on Solaris (and earlier on Linux) were 
obtained from using ``openssl md5 < largefile''.

Seems like BSD is not the only OS, where the mmap's theoretical promise 
lags behind the actual offering -- read wins on Solaris overall too, 
despite being quite a bit more expensive in sys-time. Would still be 
nice to be the first to deliver...

    -mi