Re: [UPDATE] Fresh zerocopy patch on kernel.org

[Malcolm Beattie]

David S. Miller writes:
> 
> At the usual place:
> 
> ftp://ftp.kernel.org/pub/linux/kernel/people/davem/zerocopy-2.4.1-1.diff.gz

Hmm, disappointing results here; maybe I've missed something.

Setup is a Pentium II 350MHz (tusk) connected to a Pentium III
733MHz (heffalump) (both 512MB RAM) with SX fibre, each with a
3Com 3C985 NIC. Kernels compared are 2.4.1 and 2.4.1+zc
(the 2.4.1-1 diff above) using acenic driver with MTU set to 9000.
Sysctls set are
    # Raise socket buffer limits
    net.core.rmem_max = 262144
    net.core.wmem_max = 262144
    # Increase TCP write memory
    net.ipv4.tcp_wmem = 100000 100000 100000
on both sides.

Comparison tests done were
    gensink4: 10485760 (10MB) buffer size, 262144 (256K) socket buffer
    ftp: server does sendfile() from a 300MB file in page cache,
         client does read from socket/write to /dev/null in 4K chunks.

                       2.4.1                    2.4.1+zc
                     KByte/s tusk%CPU heff%CPU  KByte/s tusk%CPU heff%CPU
gensink4
  tusk->heffalump    94000   58-100   93        54000   98-102   11-45
  heffalump->tusk    72000   86-100   46-59     70000   71-93    53-71

                      2.4.1                     2.4.1+zc         
                      KByte/s                   KByte/s
ftp heffalump->tusk   86000                     62000

I was impressed with the raw 2.4.1 figures and hoped to be even more
impressed with the 2.4.1+zc numbers. Is there something I'm missing or
can change or do to help to improve matters or track down potential
problems?

--Malcolm

-- 
Malcolm Beattie <[e-mail]>
Unix Systems Programmer
Oxford University Computing Services