Achieved ~Gigabit on tests - now what?

Archived from groups: comp.dcom.lans.ethernet (More info?)

Hi,

Per a post that I just made, I was finally able to configure several
network test tools that that I've been using to produce results near
Gigabit/sec performance on tests between two machines with Intel GigE
NICs (4 x Xeon 3.0GHz CPUs, 16 GB memory, Windows 2003). To reiterate
what I finally had to do, I had to set command line parameters for:

MSS: 100000
Window Size: 64K
Buffer Size: 24K


Based on these tests, it appears that these machines, with these NICs,
and under Windows 2003, are able to achieve near Gigabit performance.


What I'm wondering about now (and the reason for this post) is that my
impression is that by using these command line parameters, these test
tools are programmatically setting certain parameters in the Windows
TCP/IP Registry entries and or the tool itself.

From what I can tell:

- The "MSS" is controled by the registry item "MTU"
- The "Window Size" is equivalent to the registry item "TcpWindowSize"
- The "Buffer Size" is an application-level (not driver or stack)
parameter.

Now that I know what parameter values are needed (and where they are) to
achieve the Gigabit speeds:

- Can I, and should I modify the parameter settings in the Windows
registry (e.g., TcpWindowSize and MTU) to 'hard code' these values into
the Windows TCP/IP stack?

- The last parameter above, "Buffer Size", appears to be an
application-level parameter, i.e., how large a buffer the
application/tool writes or reads at-a-time. Does this mean that any
application that doesn't use such a larger buffer size will not be able
to achieve the kinds of speeds that I'm seeing with these test tools?


Re. the latter question above, I did some testing with an application
that I have that transfers a large amount of data over the GigE
interface, and it doesn't appear that I would get the kind of throughput
that I'm seeing with these tools (ignoring disk driver performance for
now) unless that application was modified to use larger (e.g., 24Kbytes)
buffers. Is this true?


Again, thanks for all of your comments to the previous thread, and I
hope that you'll help me understand the "implications" of those tests.

Yours,
Jim Lum