Postcard-based On-Path Flow Data Telemetry

In order to gain detailed data plane visibility to support effective network OAM, it is important to be able to examine the trace of user packets along their forwarding paths. Such on-path flow data reflect the state and status of each user packet's real-time experience and provide valuable information for network monitoring, measurement, and diagnosis. The telemetry data include but not limited to the detailed forwarding path, the timestamp/latency at each network node, and, in case of packet drop, the drop location and reason. The emerging programmable data plane devices allow user-defined data collection or conditional data collection based on trigger events. Such on-path flow data are from and about the live user traffic, which complement the data acquired through other passive and active OAM mechanisms such as IPFIX and ICMP. In-band Network Telemetry (INT) was designed to cater this need (note that although INT has been widely used, the term "in-band" here does not comply with IETF's definition. "on-path" or "in-situ" may be more accurate terms). in-situ OAM (IOAM) represents the related standardization efforts. In essence, INT augments user packets with instructions to tell each network node on their forwarding paths what data to collect. The requested data are inserted into and travel along with the user packets. Some end nodes are responsible to strip off the data trace and export it to a data collector for processing. While the concept is simple and straightforward, INT faces several technical challenges: Issue 1: INT header and data processing needs to be done in data plane fast path. It may interfere with the normal traffic forwarding (e.g., leading to forwarding performance degradation) and lead to inaccurate measurements (e.g., resulting in longer latency measurements than usual). This undesirable "observer effect" is problematic to carrier networks where stringent SLA must be observed. Issue 2: INT may significantly increase the user packet's original size by adding the instruction header and data at each traversed node. The longer the forwarding path and the more the data collected, the larger the packet will become. The size may exceed the path MTU so either INT cannot apply or the packet needs to be fragmented. Limiting the data size or path length reduces the effectiveness of INT. On the other hand, the INT header and data can be deeply embedded in a packet due to various transport protocol and tunnel configurations. The required deep packet header inspection and processing may be infeasible to some data plane fast path where only a limited number of header bytes are accessible. Issue 3: INT requires attaching an instruction header to user packets to inform network nodes what types of data to collect. Due to the header overhead constraint and hardware-friendly consideration, TLV is undesirable for data type encoding. Instead, IOAM use a bitmap where each bit indicates one pre-defined data type . However, new use cases may require new data types. The current allocated 16-bit bitmap limits the data type scalability. The proposed bitmap extension in provides a method to support more data types but it also increases the IOAM header size. Issue 4: INT header needs to be encapsulated into user packets for transport. has discussed several encapsulation approaches for different transport protocols. However, it is difficult to encapsulate extra header in MPLS and IPv4 networks which happens to be the most widely deployed and where the path-associated telemetry data is most wanted by operators. The proposed NVGRE encapsulation for IPv4 in requires a tunnel to be built between each pair of nodes which may be unrealistic for plain IP networks. Issue 5: The INT header and data are vulnerable to eavesdropping and tampering as well as DoS attack. Extra protective measurement is difficult on the fast data path. Issue 6: Since INT only exports the telemetry data at the designated end node, if the packet is dropped in the network, the data will be lost as well. It cannot pinpoint the packet drop location which is required for fault diagnosis. Even worse, the end node may not be aware of the lost of packet at all. The above issues are inherent to the INT-based solutions. Nevertheless, the on-path data acquired by INT are valuable for network operators. Therefore, alternative approaches which can collect the same data but avoid or mitigate the above issues are desired. This document provides a new approach named Postcard-Based Telemetry (PBT) with two different implementation variations, each having its own trade-off and addressing some or all of the above issues. The basic idea of PBT is simple: at each node, instead of inserting the collected data into the user packets, the data are directly exported through dedicated OAM packets. Such "postcard" approach is in contrast to the "passport stamps" approach adopted by INT . The OAM packets or postcards can be generated by the node's slow path and transported in band or out of band, independent of the original user packets.

This section describes the variation of PBT which triggers the postcard export with a mark in user packets. PBT-M aims to address the challenges of INT listed above and introduce some new benefits. We first list all the design requirements of PBT-M.

Req. 1: We should avoid augmenting user packets with new headers or introducing new data plane protocols. This helps to alleviate or eliminate the issue 1, 2, 4, and 5. We expect the OAM data collecting signaling remains in data plane. Simple packet marking techniques suffice to serve this purpose. It is also possible to configure the OAM data collecting from the control plane. Req. 2: We should make the scheme extensible for collecting arbitrary new data to support possible future use cases. The data set to be collected is preferred to be configured through management plane or control plane. Since there is no limitation on the types of data, any custom data including those generated by DNPs can be collected. Since there is no size constraints any more, it is free to use the more flexible data set template for data type definition. This addresses the issue 2 and 3. Req. 3: We should avoid interfering the normal forwarding and affecting the forwarding performance when conducting data plane OAM tasks. Hence, the collected data are better to be transported independently by dedicated OAM packets through in-band or out-of-band channels. The data collecting, processing, assembly, encapsulation, and transport are therefore decoupled from the forwarding of the corresponding user packets and can be performed in data plane slow path if necessary. This addresses the issue 1, 4, and 5. Req. 4: The data collected from each node is not necessarily identical, depending on application requirements and node capability. Data for different operation modes can be collected at the same time. These requirements are either impossible or very difficult to be supported by INT in which data types collected per node are supposed to be identical and for a single mode. Req. 5: The flow's path-associated data can be sensitive and the security concerns need to be carefully addressed. Sending OAM data with independent packets also makes it easy to secure the collected data without exposing it to unnecessary entities. For example, the data can be encrypted before being sent to the collector so passive eavesdropping and man-in-the-middle attack can both be deterred. This addresses the issue 5. Req. 6: Even if a user packet under inspection is dropped in network, the OAM data that have been collected should still be exported and help to diagnose the packet drop location and reason. This addresses the issue 6.

In light of the above discussion, the sketch of the proposed solution, PBT-M, is as follows. The user packet, if its path-associated data need to be collected, is marked at the path head node. At each PBT-aware node, if the mark is detected, a postcard (i.e., the dedicated OAM packet triggered by a marked user packet) is generated and sent to a collector. The postcard contains the data requested by the management plane. The requested data are configured by the management plane through data set templates (as in IPFIX). Once the collector receives all the postcards for a single user packet, it can infer the packet's forwarding path and analyze the data set. The path end node is configured to unmark the packets to its original format if necessary. The overall architecture of PBT-M is depict in Figure 1.

+------------+ +-----------+ | Network | | Telemetry | | Management |(-------| Data | | | | Collector | +-----:------+ +-----------+ : ^ :configurations |postcards (OAM pkts) : | ...............:.....................|........ : : : | : : +---------:---+-----------:---+--+-------:---+ : | : | : | : | V | V | V | V | +------+-+ +-----+--+ +------+-+ +------+-+ usr pkts | Head | | Path | | Path | | End | ====>| Node |====>| Node |====>| Node |====>| Node |====> | | | A | | B | | | +--------+ +--------+ +--------+ +--------+ gen postcards gen postcards gen postcards gen postcards mark usr pkts unmark usr pkts

Although PBT-M solves the issues of INT, it introduces a few new challenges. Challenge 1: A user packet needs to be marked in order to trigger the path-associated data collection. Since we do not want to augment user packets with any new header fields (i.e., Req. 1), we must reuse some bit from existing header fields. Challenge 2: Since the packet header will not carry OAM instructions any more, the data plane devices need to be configured to know what data to collect. However, in general, the forwarding path of a flow packet (due to ECMP or dynamic routing) is unknown beforehand. Configuring the data set for each flow at all data plane devices is expensive in terms of configuration load and data plane resources. Challenge 3: Due to the variable transport latency, the dedicated OAM packets for a single packet may arrive at the collector out of order or be dropped in networks for some reason. In order to infer the packet forwarding path, the collector needs some information from the OAM packets to identify the user packet affiliation and the order of path node traversal.

To address the above challenges, we propose several design details of PBT-M.

To trigger the path-associated data collection, usually a single bit from some header field is sufficient. While no such bit is available, other packet marking techniques are needed. we discuss three possible application scenarios. IPv4. IPFPM is an IP flow performance measurement framework which also requires a single bit for packet coloring. The difference is that IPFPM does in-network measurement while PBT only collects and exports data at network nodes (i.e., the data analysis is done at the collector rather than in the network nodes). IPFPM suggests to use some reserved bit of the Flag field or some unused bit of the TOS field. Actually, IPFPM can be considered a subcase of PBT so the same bit can be used for PBT. The management plane is responsible to configure the actual operation mode. SFC NSH. The OAM bit in NSH header can be used to trigger the path-associated data collection . PBT does not add any other metadata to NSH. MPLS. Instead of choosing a header bit, we take advantage of the synonymous flow label approach to mark the packets. A synonymous flow label indicates the path-associated data should be collected and forwarded through a postcard.

By default, all PBT-aware nodes are configured to react to the marked packets by exporting some basic data such as node ID and TTL before a data set template for that flow is configured. This way, the management plane can learn the flow path dynamically. If the management plane wants to collect the path-associated data for some flow, it configures the head node(s) with a probability or time interval for the flow packet marking. When the first marked packet is forwarded in the network, the PBT-aware nodes will export the basic data to the collector. Hence, the flow path is identified. If other types of data need to be collected, the management plane can further configure the data set template to the target nodes on the flow's path. The PBT-aware nodes would collect and export data accordingly if the packet is marked and a data set template is present. If for any reason, the flow path is changed. The new path nodes can be learned immediately by the collector, so the management plane controller can be informed to configure the new path nodes. The outdated configuration can be automatically timed out or explicitly revoked by the management plane controller.

The collector needs to correlate all the OAM packets for a single user packet. Once this is done, the TTL (or the timestamp, if the network time is synchronized) can be used to infer the flow forwarding path. The key issue here is to correlate all the postcards for a same user packet. The first possible approach is to include the flow ID plus the user packet ID in the OAM packets. The flow ID can be the 5-tuple IP header of the user traffic. The user packet ID can be some unique information pertaining to a user packet (e.g., the sequence number of a TCP packet). If the packet marking interval is large enough, then the flow ID itself is enough to identify the user packet. That is, we can assume all the exported OAM packets for the same flow during a short period of time belong to the same user packet. Alternatively, if the network is synchronized, then the flow ID plus the timestamp at each node can also infer the postcard affiliation. However, some errors may occur under some circumstances. For example, if two consecutive user packets from the same flows are both marked but one exported postcard from a node is lost, then it is difficult for the collector to decide which user packet the remaining postcard belongs to. In many cases, such rare error has no catastrophic consequence therefore is tolerable.

It is possible to avoid needing to mark user packets yet still allowing in-band flow data collection. We could simply configure the Access Control List (ACL) to filter out the set of target flows. This approach has two potential issues: (1) Since the packet forwarding path is unknown in advance, one needs to configure all the nodes in a network to filter the flows and capture the complete data set. This wastes the precious ACL resource and is not scalable. (2) If a node cannot collect data for all the filtered packets of a flow, it needs to determine which packets to sample independently, so the collector may not be able to receive the full set of postcards for a same user packet. Nevertheless, since this approach does not require to touch the user packets at all, it has its unique merits: (1) User can freely choose any nodes as vantage points for data collection; (2) No need to worry that any "modified" user packets to leak out of the PBT domain; (3) It has the minimum impact to the forwarding of the user traffic. No data plane standard is required to support this mode, except the postcard format.

Since PBT-M has some challenges as listed in , this section describes another variation of PBT, which essentially compromises some of the design requirements listed in , yet retains most of the benefits of PBT. PBT-I can be seen as a trade-off between INT and PBT-M. PBT-I needs to add a fixed length instruction header to user packets for OAM data collection. However, the collected data will be exported through dedicated postcards. On the one hand, PBT-I violates the Req. 1 in . It also makes it harder to meet the Req. 2. On the other hand, the overhead of the instruction header is fixed and user packets will not inflate with path length or telemetry data quantity. We also introduce an optimization to mitigate the impact on Req. 2. In return, PBT-I addresses all the challenges of PBT-M: There is no need to find an existing header field to mark a user packet. We can implement PBT-I as an option of IOAM which uses the same encapsulation method for IOAM. So far, the IOAM header encapsulation methods have been defined for several protocols, including IPv6, VXLAN-GPE, NSH, SRv6 ,, GENEVE , and GRE . describes the approach to encapsulate the instruction header into MPLS packets. There is no need to configure the nodes about the data to be collected since the data set information is carried in the instruction header. If implemented as an IOAM option, the data set representation can be identical to that of the IOAM trace optioin. The instruction header can be designed to contain enough information to help correlate the postcard packets belonging to a user packets. Even better, new fields can be added to track each flow and each packet, and easily detect any packet drop.

The sketch of the proposed solution, PBT-I, is as follows. If the path-associated data need to be collected for a user packet, an instruction header is inserted into the packet at the path head node. At each PBT-aware node, if the instrution header is detected, a postcard is generated and sent to a collector. Once the collector receives all the postcards for a single user packet, it can combine and analyze the data set. The path end node is configured to remove the instruction header. The overall architecture of PBT-I is depict in Figure 2. Note that in the figure we omit the controller which configures the nodes for necessary functions (e.g., head node encapsulation) and information (e.g., IP address of the data collector).

+-----------+ | Telemetry | | Data | | Collector | +-----------+ ^ |postcards (OAM pkts) | | | +--------------+------+-------+--------------+ | | | | | | | | +---+----+ +---+----+ +---+----+ +---+----+ usr pkts | Head | | Path | | Path | | End | ====>| Node |====>| Node |====>| Node |====>| Node |====> | | | A | | B | | | +--------+ +--------+ +--------+ +--------+ insert Instr Hdr remove Instr Hdr gen postcards gen postcards gen postcards gen postcards

The implementation of PBT-I as a standalone IOAM option named IOAM Direct Export is described in .

Postcard can use the same data export format as that used by IOAM. proposes a raw format that can be interpreted by IPFIX.

Several security issues need to be considered. Eavesdrop and tamper: the postcards can be encrypted and authenticated to avoid such security threats. DoS attack: PBT can be limited to a single administration domain. The mark must be removed at the egress domain edge. The node can rate limit the extra traffic incurred by postcards.

No requirement for IANA is identified.

TBD.