On Mon, Aug 28, 2000 at 09:03:06AM -0500, Mark_H_Johnson.RTS@raytheon.com wrote:
>
>
> Hmm. I must be particularly dense today - I didn't notice what you describe
> after a short review of the code in drivers/net/bonding.c. That code appears to
> use each bonded channel in order [round robin as described in
> Documentation/network/bonding.txt]. Is that the right place to look or is there
> some other information [online or otherwise] that describes the algorithm used
> in channel bonding? I am also curious why it is "better" to send packets out of
> order [and may cause congestion]. What problems does that algorithm solve that
> are worse than channel congestion and 1/10th the expected performance?
There must be some misunderstanding. My point was that round robin inherently
leads to packet reordering (easy to see when the packets have different lengths,
but even with equal packet lengths small timing differences can cause it).
Packet reordering is bad for performance.
>
> Re: setting the TCP window size.
> I'm using the defaults [unless the application happens to set it - I will
> check]. For the "how to set the TCP window size", I assume you are referring to
> using "route" to do so (e.g., route ... window W). I noticed the route (8) man
> page says this is "typically used on AX.25 networks and drivers unable to handle
> back to back frames". Is there something inherent in the Ethernet interface that
> would require this?
The default 32K window is too small for fast networks (fast ethernet and up,
or anything with longer latency like round robin setups). You can change it
using the /proc/sys/net/core/[rw]mem_{default,max} sysctls (see socket(7))
-Andi
-
: send the line "unsubscribe linux-net" in
the body of a message to majordomo@vger.kernel.org
