Re: Server slowdown...
In article <[🔎] 200404130009.16053.russell@coker.com.au> you wrote:
> On shared media (such as 10base2) accidentally leave an interface in
> promiscuous mode (there used to be a bug in tcpdump whereby running two
> copies of it at the same time could cause the interface to remain in
> promiscuous mode after both copies had exited). A moderately busy 10base2
> could destroy the performance of a decent 1995 server machine if an interface
> was in promiscuous mode, and as the CPU use occurred in interrupt context
> none of the usual tools would tell you what was happening.
>
> Send lots of minimal size packets to a server or to the media broadcast
> address.
Both can be seen with the software (si) und hardware (hi) interrupt times in
recent top output, as well as the interrupt count for the various IRQs. You
can also see the number of context switches, which are usefull to monitor.
So using "vmstat 1" is a good first help, the "in" column are the
interruopts, it is usually below 1020 on a idle system (with 2.6.4). The
number of context switches (cs) between 200 and 350. On a network loaded
system you can see minimum 20% hi/si cpu time and 5000in/cs in vmstat.
Try to use "cat /proc/interrupts" to dentify the hardware which is causing
the interrupts. On an idle system, the timer interrupts are about 20 times
as many as any other interrupts (on 2.6 kernels)
You normally have to compare it with "normal" workload, otherwise the numbers are not
very easy to decipher.
Greetings
Bernd
--
eckes privat - http://www.eckes.org/
Project Freefire - http://www.freefire.org/
Reply to: