Linux: How to find out whether NIC is causing a bottleneck?

Question

There are quite a lot of resource of how to check a bottleneck causes by local harddisk IO (iostat), CPU (top), latency (ping) or even network bandwidth (the switch may tell).

How to tell if a NIC does matter?

Answer 1

I recommend dstat -nf and dstat -i or dstat -if. Things to look for:

• Network data rates (in MB/s) equal or close to ~ 1/8th the bandwidth (in Mbits/s). E.g. 12MBytes/s for a 100Mbit card.
• Very high interrupt rates that correspond with high packet rate / network traffic.

The first is a fundamental bandwidth rate that you can't get around. The second could be a sign that you would benefit from jumbo frames or TCP offloading / that your system is being overloaded by interrupts due to the network traffic (making it difficult for the system to respond to other events.

Answer 2

iftop -i <interface> -B may helpful. It displays bandwidth rates in bytes/sec.

Answer 3

Some kind of graphing application like cacti/mrtg/munin might help you gather networking statistics and see if you have any bottlenecks.

Answer 4

the problem here is there are a bunch of tools listed and not way to get an integrated look. a tool like collectl monitors everything and lets you see cpu,disk,network,memory and even interrupts side by side either in real-time or written to a file for later playback.

if you suspect a cpu problem, look at individual cpu data and see how much time is spent in interrupt processing. if you do suspect interrupt overhead collectl will even show you interrupts by cpu and even allow you to break it out by specific interrupt type by cpu. truly a swiss-army knife knife kind of a tool that can easily replace many others.

-mark