Improve accuracy of bandwidth measurements

[ptaylor874]

I've been using Cacti for about a month to monitor just north of a hundred routers (with plans to have over 500 within 3 months), specifically for latency and bandwidth utilization. Today, I performed a trace of Cacti's polling against one of these routers. There were a total of three SNMP polls:

1. Request for ccitt.1 (as interpreted by Wireshark), to which the router replied with the sysDescr.. (Cisco IOS, blah, blah)
2. Request for sysUpTime - Router replied as expected
3. Request for ifInOctets, ifOutOctets, etc. Router replied as expected.

Another NMS that we use performs polls for any bandwidth related task along with sysUpTime in the same SNMP GET request. The delta between the sysUpTimes is used to calculate the bandwidth utilization for that polling interval. The theory is that the sysUpTime request is always synchronized with the request for the data, so the calculated delta values are always in the right time interval. By not performing the sysUpTime request simultaneously with the data request, some small portion of the delta is always assigned by Cacti to the time interval before it actually occurred.

In addition, this would remove one round-trip to each site every polling interval. If the sysDescr request could be rolled in as well, all the better.

This may seem like a relatively minor detail, but attention to detail like this help push a product into the "Enterprise class" arena.

[rony]

It would be an issue if sysUptime was actually used for any calculations, which it is not, because it's not consistent among all the SNMP agent in the world.

[ptaylor874]

So, the time that the reply is received by Cacti is used in the calculation?

I didn't realize that sysUpTime wasn't consistent between devices... Given that, perhaps this is the best compromise.

[rony]

Local system time is used for the update into the RRDTool file.