Gus Wirth wrote:
> James G. Sack (jim) wrote:
>> Gus Wirth wrote:
>>> Andrew Lentvorski wrote:
>>>> Gus Wirth wrote:
>>>>> If I do an NFS mount and copy a large file from the server to
>>>>> /dev/null on the client or if I scp a file between the server and
>>>>> client I get pretty much the same results, about 11MB (that's bytes)
>>>>> per second. This is only about 10% better than 100baseT ethernet. I
>>>>> know from previous experiments that my hard drives and general system
>>>>> throughput can handle about 25MB/sec.
>>>> If you are seeing 10 MibiBytes per second on a 100baseT, you
>>>> effectively have no network traffic on it.  Must be nice.  That's one
>>>> reason why you are seeing practically no increase.
>>> This is my home network, with currently only two machines running. But I
>>> would have expected to about double my transfer speed based on my
>>> current system capabilities.
>>>
>>>> Second, I think you need to bump the default TCP packet size.  On a
>>>> network that small, the TCP window is probably filling and throttling
>>>> the transfer rate.
>>>>
>>>> Make sure your systems are sending 9000+ byte packets rather than just
>>>> 1500+ byte packets.
>>> OK, off to the man pages to figure out how to do that.
>>
>> maybe:
>>  ifconfig eth0 mtu 9000
> 
> That's it. Unfortunately, it breaks networking. Or more specifically, it
> breaks when network packets exceed 1500 bytes. Setting the MTU of either
> machine causes something to choke and the connection dies.
> 
> For example, I set MTU=9000 on both machines. I monitor the connection
> on the server with Ethereal. I start a scp operation from the server to
> the client. The initial exchange and negotiation of the ssh keys goes
> fine. However, as soon as a data packet gets transmitted that's bigger
> than 1500 bytes, the transmit queue starts to back up and I stop getting
> any replies from the client. It's really weird and I'm not sure what I
> should be looking at. I'm thinking I should get a gigabit crossover
> cable (is there such a thing?) and try without the switch to eliminate
> that as a potential problem.
> 
> On the other hand, I now know about the ip command, sysctl, ethtool, and
> a bunch of stuff in /proc/sys/net, none of which fixes my problem.
> 

It was my understanding that GigE automatically adjusted itself so that
crossover cables have no impact, and are never needed.

..jim


-- 
[email protected]
http://www.kernel-panic.org/cgi-bin/mailman/listinfo/kplug-list

Reply via email to