Comprehensive network management systems have a very difficult job. They have to collect vast amounts of data from network interfaces and efficiently store it for rapid access. The job is actually a combination of three tasks:
- Gather the data
- Store the data
- Access the data
I’ll look at each of these tasks in the next few blogs.
The first task is efficiently gathering the data. The main driver for NMS data collection is the number of interfaces from which data must be collected. While most organizations will know how many devices they have, few will be able to estimate the number of interfaces that should be monitored. However, there are some ways to create an estimate.
Enterprise networks tend to built with switches that have 24 or 48 ports per chassis or per blade. A rough approximation of the number of interfaces in the network is a multiple of one of these numbers. Routers will have some number of physical interfaces that depends on device model, so knowing the rough number of devices allows us to estimate the number of physical interfaces. Routers and Layer 3 switches tend to have a number virtual interfaces (e.g., VLAN interfaces) that can be estimated. The total of each of these estimates gives us the approximate number of interfaces in the network. Add to that the number of devices, because the NMS needs to also monitor the cpu utilization, memory, and buffer stats for each device. Service provider networks may have fewer switches and more routers, causing their estimates to be driven more by virtual interfaces than physical interfaces.
An example network comprised of twenty-five 7200s, fifty 6500s, fifty 4500s, and two hundred 3750 switches might have an average of 4 blades of 48 ports each per 6500 or 4500. The port count is then
6500 ports: 50 * 4 * 48 = 9,600
4500 ports: 50 * 4 * 48 = 9,600
3750 ports: 200 * 48 = 9,600
7200 ports: 25 * 3 = 75
Five VLANs per device: 325 * 5 = 1625
300 devices: 1 * 300 = 300
Total number of monitored elements: 30,800
The NMS must monitor about ten values for each of these entities, so 308,000 SNMP variables must be polled per polling cycle. If we assume that each query returns 10 bytes of data, and the polling interval is five minutes, we have a rough data rate of 100Kbps. That sounds like a low volume.
If the NMS does not efficiently combine multiple variable requests in each packet, there could be significant packet over head just due to the ethernet protocol. (See the Network Arcanum Blog by James Ventre.) Even with poor request packing and ethernet protocol overhead, we’re not talking about significant bandwidth.
Netcordia measured a sustained data rate of about 100Kbps in each direction on NetMRI’s interface to a campus network that consisted primarily of switches and a few routers at the edge. Overall, this level of network utilization is easily handled, and unlike many other network services, the utilization tends to be nearly constant.
Another way to look at the data rate is the rate at which data must be written to disk. I’ll discuss the data storage requirements in the next blog post Handling NMS Performance Data, Part 2.
-Terry
_____________________________________________________________________________________________
Re-posted with Permission
NetCraftsmen would like to acknowledge Infoblox for their permission to re-post this article which originally appeared in the Applied Infrastructure blog under http://www.infoblox.com/en/communities/blogs.html
