| Doug Freyburger 2005-10-29, 5:52 pm |
| blackdog wrote:
> Doug Freyburger wrote:
>
BTW, there is no such thing as NIS mounting. You
must mean NFS mounting by machines that are NIS
clients.
[vbcol=seagreen]
>
>
>
>
> how to check switching MTU router?
Login to routers. Ask it the MTU on all interfaces
from one end to the other. See if they are all the
same. While you're on the routers check the latency
at each step and end those together.
Login to each host. Use ifconfig to view the MTU on
various interfaces. See if they are all the same.
For example the default MTU on ethernet (10, 100 and
crzily enough 100) are all 1500. The MTU on FDDI is
4500 IIRC. Various types of WAN line have all sorts
of MTU sizes. A carefully designed gigabit ethernet
infastruction will use jumbo frames with a higher MTU.
The lower the CPU performance of the routers on the
end to end path, the more the impact a change in
MTU has. Getting 1500 byte frames broken into 750
byte frames doubles the number of interupts on the
receiving end and interupt time domain is one of the
most common limiting factors.
An amusing story about network bottlenecks. I once
adminstered a network of workstations and some file
servers. After a migration to a new building the
performance fell through the floor. I got the Visio
of the design and flipped out. All file servers and
web servers were on one large switch. All of the
CAD workstations were on another large switch. There
was one ISL between the switches. The netheads said
that it stayed at 80% of saturation and was therefore
not overloaded. I pointed out that TCP and UDP-NFS2
both used a back-off formula to deal with slow lines
and I showed nfsstat and iostat outputs that showed
that all of the workstations thought the line was slow.
It held at 80% because the back-off formula stopped
backing off at that point. Moral of the story - your
network architecture can effect your performance and
both bandwidth and latency matter more than average
load across a channel.
|