[Gllug] What's that network congestion setting called?

[Ian Northeast]

John Winters wrote:
>>John Winters wrote:
>>
>>>DNS issues would affect everything (not just web browsing) and wouldn't
>>>show this 100% load thing.  I find it very odd that I can have 100% CPU
>>>usage but nothing apparently using it.
>>
>>I missed the 100%-cpu-but-nothing-using-it thing.  what CPU state does
>>top show the usage is in?
> 
> 
> 90% user, 10% sys.  No process in the list is clocking up a significant
> amount of CPU time.
> 
> This happens with two different web browsers, so I've eliminated the browser.

It may be totally unrelated but I saw something a bit like this today. 
It was a test (luckily) SLES9 (2.6.5) box I was updating to patch 1 
using YaST (unlike in SLES8, there isn't a nice script on the patch CD, 
seems you *have* to use YaST now:() remotely from an NFS mounted CD, 
with both the NFS and X servers on the same PC. It took ages to do 
anything, responded very poorly and at one point locked my X server up 
for a while. But it finished eventually. Top didn't show anything 
munching CPU.

It turned out to be a NIC with the wrong duplex settting (bloody Cisco 
switches). I'd noticed that with the e100 driver for the EtherExpress 
Pro it remembers the NIC settings across a reboot - but not, it seems, 
after a power down. So I hadn't got around to putting the settings into 
the config, just set it with ethtool. I'd powered it down this morning 
to put in a new disk. TBH I'm amazed it worked at all.

So - do you use NFS? Is your X server remote? Have you verified that all 
your NICs are set correctly? Incorrect NIC settings are a common cause 
of network torpitude, especially when you mix in NFS. And web browsers 
tend to stress X servers quite heavily. I can imagine seeing the 
symptoms you describe running Mozilla in an NFS'ed $HOME on a remote X 
server with the NIC settings wrong. The network would be flooded with 
retransmissions. But other tasks which stressed the network less heavily 
would hardly be affected. My machine was apparantly fine until I started 
this task, but I hadn't used either NFS or X before this.

PS I don't actually know if a Cisco switch would autonegotiate correctly 
with Linux. Our network admins always disable autonegotiation on them 
because they fail to do it properly with a number of other platforms, so 
I've never been able to check. I expect other network admins routinely 
do the same thing. Plugging a device defaulting to autonegotiate into a 
Cisco switch which has its settings fixed at either 10 or 100 full 
duplex is guaranteed to get the speed correct and the duplex wrong, 
resulting in a connection which works, and is fine for low density 
traffic like command line over ssh, but slows to a crawl with high traffic.

Regards, Ian


-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug