Not “New Data” But “New Ways of Getting Data”

Model-driven telemetry has been one of the most fun projects I’ve worked on in a long time. Right from the start, we worked hand-in-hand with customers to identify their biggest pain points when it came to network operational data. To my surprise, the primary ask was not for new types of data or fancy new counters. Quite the opposite, in fact. The most useful data already exists in the network, it’s just too hard to get. The most frequent culprit? SNMP polling. Almost 30 years since it was first standardized, SNMP has done many good things but it hasn’t kept up with the speed and scale that modern networks require.

If you need an SNMP refresher first, read this blog by my colleague, Frederic.

Push Not Pull

We talk a lot about “push” being a better mechanism than “pull”, but what exactly does that mean? To retrieve large amounts of data, SNMP polling relies on the GetBulk operation. Introduced in SNMPv2, GetBulk performs a continuous GetNext operation that retrieves all the columns of a given table (e.g. statistics for all interfaces). The router will return as many columns as can fit into a packet (subject to the max-repetitions parameter). If the polling client detects that the end of the table has not yet been reached, it will do another GetBulk and so on.

zTake a look at a packet trace for a GetBulk for interface statistics (the IF-Table of the IF-MIB). The router ( responds with 70 OIDs in the first packet and the poller ( does a second request for the remaining OIDs. On receiving the second request, the router continues its lexicographical walk and fills up another packet. When the poller detects that the router has “walked off the table” (by sending OIDs that belong to the next table), it stops sending GetBulks. For very large tables, this could take a lot of requests and the router has to process each one as it arrives.

This situation is bad enough if you have a single SNMP poller, but what if you have two or more polling stations (which most people do if only for redundancy)? The router has to process each request independently, finding the requested place in the lexical tree and doing the walk, even if both pollers requested the same MIB at more-or-less the same time. Many network operators know this empirically: the more pollers you have, the slower the SNMP response.

Telemetry gains efficiency over SNMP by eliminating the polling process altogether. Instead of sending requests with specific instructions that the router has to process each and every time, telemetry uses a configured policy to know what data to collect, how often and to whom it should be sent. zz

Take a look at a packet capture of interface statistics sent with telemetry using the compact Google Protocol Buffer encoding instead of SNMP. (This represents a superset of SNMP interface statistics since the router stores 36 statistics for every interface and the IF-Table has less.) At any rate, as you can see, all of the statistics fit into a single UDP packet. But the biggest gains have nothing to do with packet size. The really important thing is that the router does a single dip into the internal data structures to acquire the data and, if there are multiple interested receivers, the router can just duplicate the packet for different destinations (a simple and efficient operation for a router). So sending telemetry to one receiver has the same latency and overhead as sending it to five.

Push Bags, Not Columns

Polling is not the only computational burden that SNMP imposes on a router. An equally significant performance factor is how data is ordered and exported. I know that sounds incredibly boring and abstract, but hey, nobody said performance optimization was exciting!

SNMP imposes a very tight model when it comes to indexing and exporting. Take the example of the SNMP IF-Table. Each column of the table represents a different parameter for a given interface, indexed by the ifIndex. I’ve shown the first five columns below:


The strict semantics of the GetNext/GetBulk operations force the router to traverse the table column by column (returning the list of ifIndex, followed by a list of ifDescr, etc) from lowest index value to highest. From a router’s perspective, that’s just not natural.

Not surprisingly, routers store internal data in a way that’s most efficient for routers. In IOS XR, for example, the internal data structure for interface statistics is indexed by interface name and is stored in a structure called a bag (basically, an unordered superset of the data in a row of the IF-Table above).  The router’s most efficient internal bulk data retrieval is to grab a whole bag (or, even better, bags) of data at once. But the router can’t just send the bag in SNMP. Instead, it has to re-order the data into a table and walk the columns to fulfill the GetBulk request. Now, the router can do all kinds of internal optimizations to make this process better (like auxiliary indices and caching techniques) but that all adds up to extra processing work and may also result in stale data.

How much better it would be if we could just free the router to present its data in the natural order! Well, that’s exactly what telemetry does. Telemetry collects data using the internal bulk data collection mechanisms, does some minimal processing to filter and translate the internal structure to a Google Protocol Buffer, and then pushes the whole thing to the collector at the configured intervals. We’ve worked hard to minimize processing overhead at every step, so you get the fastest, freshest data possible with the least amount of work.

Bringing The Network Into Focus

After nearly 30 years, SNMP is nearly ubiquitous. Almost all modern network monitoring platforms rely on it to some degree or another. But for large-scale networks with real-time monitoring requirements, the fundamental operational characteristics of the SNMP protocol create bottlenecks that prevent valuable operational data from getting off the router. Model-driven telemetry frees the data from the constraints of SNMP and delivers it at a velocity that is often orders of magnitude better. Earlier this year, our first customers began turning off SNMP in favor of telemetry. Now it feels a bit like getting a new telescope for my tenth birthday — I can’t wait to see what new insights will emerge from the volume of network data we can access now.

Coming to NANOG 67 in Chicago next week? Stop by my presentation “Ten Lessons from Telemetry” on Wednesday June 27th, from 10:30am to 11:00am. There will be lots to talk about!


Shelly Cadora

Technical Marketing Engineer