Network Monitoring


As with other subsystems which contain instance data, you can monitor both summary (in brief and verbose modes) and detail data. Like disk data, the key brief mode values are bytes and packets (rather than iops). The actual data comes from /proc/net/dev.

The one key thing to keep in mind with network data is that not all networks are the same. Just like there are device mapper disks that shouldn't be included in the summary data the same is true for network devices. Those that are not included in the summary are the loopback, sit, bond and vmnet devices,

Since most lan networks run fairly cleanly and errors are rare, one is usually not interested in seeing long columns of zeros that never change and so by default brief mode does not include any error information. Adding --netopts e will add an additional column with a total error count. To see specific errors one would have to run in summary more and do identify the specific networks on which those errors were occurring you would have to run in detail mode.

What about IB over IP?
Good question. When using Infiniband networking you typically get an IB network device created. So does this mean IB traffice gets counted twice when you monitor both it and network data? As they say, it depends. Some Infiniband data will indeed go over the native IB interface and never show up as network data. This includes MPI traffic or lustre which uses the native IB transport. However, other uses of Infiniband may in fact be counted as network traffic. BUT this is actually a good thing because if you're a heavy user of IB/IP and want to be able to differentiate the native IB traffic from it, simply look at the network detail data and subtract any IB network numbers from the native values.

Tips and Tricks
Ever try looking for a needle in a haystack, in this case maybe it's network errors? --network E works just like its lowercase cousin except it tells collectl to only report intervals that have network errors in them. While this can be extremely boring in real-time mode, consider what happens during playback. During the course of a day you'll have 8640 samples but this switch will allow you to see the one that recorded the network error!

Before describing how filtering works, let's quickly review the difference between summary and detail data. When looking at summary data collectl reports the totals across all network interfaces except for those it knows have already been accounted for. For example, if your machine has eth0 and eth1 bonded together, the bond shows up as a third interface and if collectl were to report the totals across all three, the numbers would be twice that expected. For this reason, collectl only looks at specific types of interfaces, which at time of this writing are limited to eth, ib, em and p1p, though no doubt there will be others in the future.

You can change the set of interfaces collectl includes in its summary totals with --netfilt, the target of which is actually one or more comma separated perl expressions, but if you don't know perl all you really need to know is these are strings that are compared to each network name and only those that match will be included in the summary. This not only makes it possible to reduce those network you wish to summarize (say you want to summarize ethernet networks but not infiniband ones) OR include new ones that are not currently known by collectl.

A second form of the switch, in which the first character is a ^ cause all names that match the expression to be excluded from the summary data. Whether to use the first or second form may be more a matter of which is less complex for the particular form of summarization you're looking for, but if you're trying to extend the list of those interfaces recognized as being summarizable, your only choice is the first form.

In addition to controlling how network data is summarized, this switch also controls which interfaces are reported as detail data. However, be careful when using this switch during playback because you may not wish to see detailed and summary data filtered the same way. If this is the case, you will need to play back the data twice, once with -sn using the --netfilt to control the summary data and a second time with -sN and a different value for --netfilt to control the detail listing. It would certainly be possible to introduce separate filtering switches for summary and detail data, but it is felt that this situation is uncommon enough to not make things more confising than they already are.

Dynamic Network Discovery
Network devices typically don't change unless you install a new network card and reboot the machine, in which case collectl when the list of known networks is created it's simply correct. However, especially when running on a host that creates virtual machines, networks can come and go. The way collectl does this is to simply maintain a dynamic list of the active networks and make sure the statistics get associated with the correct network. As a way to preserve network details in plot format, collectl further retains a list of all the networks ever seen since system boot and uses that list to record the statistics, thereby keeping the columnar data consistent. In many cases where hypervisors reuse the virtual network names, the list of unique names remains relatively low.

However it was recently discovered that the OpenStack Quantum service for handling dynamic networks actually uses a new network name every time a VM is created. If a particular host has dozens of VMs and they come and go many times, the result can be a very long list of network names which not only show up in all the headers of all the files collectl generates, they will also show up as unique sets of data in the network detail file if one is generated. On one long running host, over 500 unique network names were generated!

The way collectl has chosen to deal with this is with --netopts o, which tells collectl to not preserve the ordered list of all the networks that have been seen and will in fact cause collectl to prune that list whenever a network is found to have been removed. While this solves the immediate need to keep only the active networks in the headers of the output files it creates a different problem with the detail files in that the data will no longer be consistent. In fact, it is probably best to not even try to generate detail files for dynamic network when using this switch. As of this writing I'm still trying to think of ways to improve the situation.
updated April 21, 2014