INTERNET-DRAFT Nick Duffield (Editor) draft-ietf-psamp-framework-01.txt Albert Greenberg November 2002 Matthias Grossglauser Expires: May 2003 Jennifer Rexford AT&T Labs - Research Derek Chiou Avici Systems Peram Marimuthu Ganesh Sadasivan Cisco Systems A Framework for Passive Packet Measurement Copyright (C) The Internet Society (2002). All Rights Reserved. This document is an Internet-Draft and is in full conformance with all provisions of Section 10 of RFC2026. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet- Drafts. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." The list of current Internet-Drafts can be accessed at http://www.ietf.org/ietf/1id-abstracts.txt The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. Abstract A wide range of traffic engineering and troubleshooting tasks rely on reliable, timely, and detailed traffic measurements. We describe a framework for passive packet measurement that is (a) general enough to serve as the basis for a wide range of operational tasks, and (b) needs only a small set of packet selection operations that facilitate ubiquitous deployment in router interfaces or dedicated measurement devices, even at very high speeds. Comments on this document should be addressed to the PSAMP WG mailing list: psamp@ops.ietf.org To subscribe: psamp-request@ops.ietf.org, in body: subscribe Archive: https://ops.ietf.org/lists/psamp/ Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 1] Internet-Draft Passive Packet Measurement November 2002 0 Contents 1 Motivation ................................................. 3 2 Requirements ............................................... 3 3 Elements, Terminology, and Architecture .................... 5 4 Packet Selection Operations ................................ 7 4.1 Filtering .............................................. 7 4.2 Sampling ............................................... 7 4.3 Hashing ................................................. 7 4.4 Selection According to Packet Treatment ................ 8 4.5 Classification and Relation of Selection Operations .... 9 4.6 Proposal on Requirements for Selection Operations ...... 9 5 Reporting .................................................. 9 6 Export and Congestion Avoidance ............................ 10 6.1 Collector Destination .................................. 10 6.2 Local Export ........................................... 10 6.3 Reliable vs. Unreliable Transport ..................... 11 6.4 Limiting Delay in Exporting Measurement Packets ........ 11 6.5 Configurable Export Rate Limit ......................... 11 6.6 Congestion-aware Unreliable Transport .................. 12 6.7 Collector-based Rate Reconfiguration ................... 12 6.7.1 Changing the Export Rate and Other Rates ........... 12 6.7.2 Notions of Fairness ................................ 13 6.7.3 Behavior Under Overload and Failure ................ 13 7 Parallel Measurement Processes ............................. 13 8 Configuration and Management ............................... 14 9 Feasibility and Complexity ................................. 14 9.1 Feasibility ............ ............................... 14 9.1.1 Filtering .......................................... 14 9.1.2 Sampling ........................................... 14 9.1.3 Hashing ............................................ 15 9.1.4 Reporting .......................................... 15 9.1.5 Export ............................................. 15 9.2 Potential Hardware Complexity .......................... 15 10 Applications .............................................. 16 10.1 Baseline Measurement and Drill Down ................... 16 10.2 Customer Performance .................................. 17 10.3 Troubleshooting ....................................... 17 11 References ................................................ 19 12 Authors' Addresses ........................................ 20 13 Intellectual Property Statement ........................... 21 14 Full Copyright Statement .................................. 21 Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 2] Internet-Draft Passive Packet Measurement November 2002 1 Motivation This document describes a framework in which to define a standard set of capabilities for network elements to sample subsets of packets by statistical and other methods. The framework will accommodate future work to (i) specify a set of selection operations by which packets are sampled (ii) specify the information that is to be made available for reporting on sampled packets; (iii) describe protocols by which information on sampled packets is reported to applications; (iv) describe protocols by which packet selection and reporting are configured. The motivation to standardize these capabilities comes from the need for measurement-based support for network management and control across multivendor domains. This requires domain wide consistency in the types of selection schemes available, the manner in which the resulting measurements are presented, and consequently, consistency of the interpretation that can be put on them. The capabilities are positioned as suppliers of packet samples to higher level consumers, including both remote collectors and applications, and on board measurement-based applications. Indeed, development of the standards within the framework described here should be open to influence by the requirements of standards in related IETF WGs, for example, IP Performance Metrics (IPPM) [PAMM98] and Internet Traffic Engineering (TEWG) [LCTV02]. Conversely, we expect that aspects of this framework not specifically concerned with the central issue of packet selection may be able to leverage work in other WGs. Potential examples are the format and export of measurement reports, which may leverage the work in IP Flow Information Export (IPFIX) [QZCZCN02], and work in congestion aware unreliable transport in the Datagram Congestion Control Protocol (DCCP) [FHK02]. 2 Requirements The broad requirements for the measurement capabilities are: * Ubiquity: The capabilities must be simple enough to be implemented ubiquitously at maximal line rate. In particular, they must involve only minimal per-packet processing and require only minimal additional state. Capabilities should not be tightly integrated with other packet control actions such as policing, marking, shaping, and queueing. * Applicability: the set of selection operations must be rich enough to support a range of existing and emerging measurement based applications and protocols. The standard will have to find a workable trade-off between the range of traffic engineering applications and operational tasks it enables, and the complexity of the set of capabilities. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 3] Internet-Draft Passive Packet Measurement November 2002 * Timeliness: reports on selected packets should be made available to the collector quickly enough to support real time applications. * Transparency: allow transparent interpretation of measurements as communicated by PSAMP reporting, without need to obtain additional information from the measuring device. * Robustness: allow robust interpretation of measurements with respect to reports missing due to loss, e.g. in transport, or omission at the measurement device. Inclusion in reporting of information enabling accuracy of measurements to be determined. * Privacy: selection of the content of packet reports will be cognizant of privacy and anonymity issues while being responsive to the needs of measurement applications, and in accordance with RFC 2804. Full packet capture of arbitrary packet streams is explicitly out of scope. * Faithfulness: all reported quantities that relate to the packet treatment must reflect the router state and configuration encountered by the packet in the PSAMP device. * Configuration: ease of configuration of sampling and export parameters, e.g. for automated remote reconfiguration in response to measurements. * Security: the use of secure means of configuration and reporting, and robustness of packet selection w.r.t. attempts to evade measurement. * Extensibility: to allow for additional packet selection operations to support future applications. * Flexibility: to support measurement of packets using different network protocols or encapsulation layers (e.g. IPv4, IPv6, MPLS, etc), and under packet encryption. * Parallel measurements: support multiple independent measurements at the same device, each possibly with different selection, reporting and export configuration. * Congestion Avoidance: export of a report stream across a network must be congestion avoiding in compliance with RFC 2914. Reuse of existing protocols will be encouraged provided the protocol capabilities are compatible with the above requirements. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 4] Internet-Draft Passive Packet Measurement November 2002 3 Elements, Terminology, and Architecture This section defines the basic elements of the PSAMP framework. * PSAMP Device: a device hosting at least an observation point and a measurement process. * Observation Point: The observation point is a location in the network where packets can be observed. Examples are, a line to which a probe is attached, a shared medium, such as an Ethernet-based LAN, a single port of a router, or set of interfaces (physical or logical) of a router, an embedded measurement subsystem within an interface. * Measurement Process: a packet measurement process comprising the following: a selection process, a reporting process, and an export process. * Selection Process: A selection process selects packets for reporting at an observation point. The inputs to the selection process are the packets observed at the observation point (including packet encapsulation headers), information derived from the packets' treatment at the observation point, and selection state that may be maintained by the observation point. Selection is accomplished through operating on these inputs with one or more selection operations. * Selection Operation: A configurable packet selection operation. It takes as input the selection process input for a single packet. The output is a binary outcome of whether or not the packet was sampled. Selection operations may also change the selection state. * Selection State: the observation point may maintain state information for use by the reporting process, and/or by multiple selection operations, either on the same packet, or on different packets. Examples include counters, timestamps, iterators for pseudorandom number generators, calculated hash values, and indicators of whether a packet was selected by a given selection operation. * Composite Selection Operation: a selection operation that is expressed as a combination of other selection operations. A packet deemed selected by the composite operation if it is selected by all its constituent selection operations. * Reporting Process: the creation of a report stream of information on packets selected by a selection process, in preparation for export. The input to a reporting process comprises that information available to a selection process, for the selected packets. The report stream contains two distinguished types of information: packet reports, and report interpretation. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 5] Internet-Draft Passive Packet Measurement November 2002 * Packet Reports: a configurable subset of the per packet input to the reporting process. * Report Interpretation: subsidiary information relating to, and used for the interpretation the reports on, one or more packets. Examples include counters, and configuration parameters of the PSAMP device, and the selection and reporting process. * Export Process: sends the output of the reporting process from the PSAMP device to one or more collectors. * Collector: a collector receives a report stream exported by one or more measurement processes. The collector may or may not be co-located with the PSAMP device. * Measurement packets: report interpretation and/or one or more packet reports are bundled by the export process into a measurement packet for export to a collector. * Parallel Measurement Processes: a given PSAMP device may host multiple independent measurement processes, each with potentially different constituent selection, reporting and export processes, and destination collectors. The parallel processes may or may not use derive their input from the same observation point. The various possibilities for the high level architecture of these elements is as follows. Note in the last case: the PSAMP device may also be a collector. +---------------------+ +------------------+ |PSAMP Device(1) | | Collector(1) | |[Obsv. Point(s)] | | | |[Meas. Process(1)]<--+---------------->| | |[Meas. Process(2)]<--+-----------+---->| | |[Meas. Process(3)]<--+--------+ | | | +---------------------+ | | +------------------+ | | +---------------------+ | | +------------------+ |PSAMP Device(2) | | +---->| Collector(2) | |[Obsv. Point(s)] | +------->| | |[Meas. Process(4)]<--+---------------->| | +---------------------+ +------------------+ +---------------------+ +------------------+ |PSAMP Device(3) | | Collector(3) | |[Obsv. Point(s)] | | | |[Meas. Process(5)]<--+---------------->| | |[Meas. Process(6)]<--+---\ +------------------+ |Collector(4)<--------+---/ +---------------------+ Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 6] Internet-Draft Passive Packet Measurement November 2002 4 Packet Selection Operations The function of packet selection is to select a subset out of the stream of all packets. Selection may be used to select a subset of packets of interest based on their content, and/or to reduce the rate of packets into the measurement flow regardless of content. This section details some candidate operations for standardization. No restriction on the allowed combination of these into composite selection operations is imposed in this document. Packet selection techniques are discussed in [ZMR02]. 4.1 Filtering Filtering can be accomplished by applying deterministic operations, such as match/mask, to any combination of bit positions in the generic selection function input. Higher level interfaces to the match/mask operations may be used to specify mask and matches for particular fields, for example, for IP addresses and/or TCP/UDP port numbers. 4.2 Sampling In current practice, sampling has been performed using particular algorithms, e.g., (i) pseudorandomly independent sampling with probability 1/N; (ii) periodic sampling of every Nth packet. The aim is to select packets representatively in conformance with some desired probabilistic selection law. Examples of selection laws are selecting packets (i) with long term probability 1/N; (ii) independently with probability 1/N; (iii) n out of every m packets independently; (iv) by importance, non-uniformly according to field contents, e.g. sample larger packets or certain protocols more frequently. A given sampling algorithm will reproduce the selection law if the packets under observation conform to a certain probabilistic content law. Examples of content laws are (i) correlations between contents of different packets decay at a specified rate; (ii) the contents of certain complimentary subfields are significantly variable and essentially uncorrelated. It follows that the extent to which sampling algorithms should be realized as distinct selection operations depends on the functional requirements, as expressed by the selection laws that are desired and the content laws that one is prepared to assume. For example, under the content law that correlations between packet contents vanish for packets separated by at least N-1 positions in the packet stream, then sampling every Nth packet periodically yields the same selection law as sampling independently with probability 1/N. With a weaker content law (i.e. admitting stronger Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 7] Internet-Draft Passive Packet Measurement November 2002 correlations between packets), then only the long term selection probabilities may be 1/N. One task for the work under the PSAMP framework will be to decide the appropriate level of functional granularity, e.g. whether to distinguish periodic from pseudorandom sampling as a packet selection operations, or to regard them only as different possible implementations of the same packet selection operation. Sampling at full line rate, i.e. with probability 1, is not excluded in principle, although resource constraints may not support it in practice. 4.3 Hashing A hashing function operates on the selection function input for a packet, and associates the resulting hash with the packet. Bit positions can be excluded from the input to the hashing function by masking. This ability would be used, for example, by applications that require the hash to be independent on packet header fields, such as TTL or header CRC, that are mutable on its passage through the network. Packets may be selected by filtering on the hash value, this regarded as part of the selection state. Although hashing and filtering are deterministic operations, a good choice of the hash function and its inputs should yield a selection law that is almost indistinguishable from independent sampling (for a given subfield of the packet) given an appropriate content law (that the contents of complimentary subfields are sufficiently uncorrelated and variable). At the application level, hash-based sampling is of interest since using the same hash functions at different PSAMP devices satisfies a functional requirement that the same sampling decision be made on a given packet observed at different devices; in [DuGr01] this is called Trajectory Sampling. It enables reconstruction of the network paths followed by individual packets, from packet reports exported from different PSAMP devices. In this application, a second distinct hash, called the label hash, may be calculated for selected packets in order to identify them at the collector. PSAMP devices are expected to be able to use a more complex hash for this second purpose, since it is only applied to the reduced set of selected packets. Calculating and reporting a set of hashes for all packets (i.e. without selecting a subset) would be required in order to support some packet tracing applications; see [SPSJTKS01]. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 8] Internet-Draft Passive Packet Measurement November 2002 4.4 Selection According to Packet Treatment Router architectural considerations may preclude some information concerning the packet treatment, e.g routing state, being available at line rate for selection of packets. However, if selection not based on routing state has reduced down from line rate, subselection based on routing state may be feasible. 4.5 Classification and Relation of Selection Operations From the above examples, it is clear that notions of sampling, filtering and hashing are not distinct. For example: (i) sampling can be accomplished by hashing and filtering; and (ii) since a sampling selection law can depend on packet content, filtering can be regarded as a degenerate case of sampling, although it does not appear useful to do so. For this reason, it is likely to be more fruitful to standardize selection operations according to agreed functional requirements, than to strive to define a non-overlapping classification of selection operations. 4.6 Proposal on Requirements for Selection Operations Section 10 describes potential PSAMP applications. These would be supported by the following set of selection operations: (i) filtering by match/mask to support drill down (ii) hash-based sampling to support Trajectory Sampling (iii) independent sampling method to support widespread baselineing 5 Reporting Information eligible for inclusion in packet reports includes (i) the packet content itself (including encapsulating headers); (ii) information relating to the packet treatment: incoming and outgoing interfaces, subinterfaces and channel identifiers, routing state applied to or derived from the packet e.g. next hop IP address, routing prefixes, source and destination AS numbers; (iii) selection state associated with the packet, e.g. timestamps, counters, hash values. In order to satisfy the requirement of ubiquity, it may be necessary to admit different levels of reporting. Concerning the packet content: some devices may not have the resource capacity or functionality to identify all fields within a packet. Concerning packet treatment: routing state is unlikely to be available on devices that do not route. At a minimum, all PSAMP devices must support simple reporting of packet content, specifically of some number of bytes contiguous packet bytes, as measured from an offset. In this style of reporting, the burden of interpretation is placed at the collector or applications that it supplies. More detailed reporting, by fields of specific protocols, is desirable where feasible. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 9] Internet-Draft Passive Packet Measurement November 2002 Information for use in report interpretation includes (i) configuration parameters of the selectors of the packets reported on; (ii) format of the packet reports (iii) configuration parameters and state information of the network element; (iv) quantities (e.g. sequence numbers) that enable collectors and applications to infer attained packet sampling rates, detect loss during selection, report loss in transmission, and correct for information missing from the packet report stream; (v) indication of the inherent accuracy of the reported quantities, e.g., of timestamps. The requirements for robustness and transparency are motivations for including report interpretation in the report stream. Inclusion makes the report stream self-defining. The PSAMP framework excludes reliance on an alternative model in which interpretation is recovered out of band. This latter approach is not robust with respect to undocumented changes in selection configuration, and leaves an architectural hostage for network management systems to coherently manage both configuration and data collection. It is not envisaged that all report interpretation be included in every report. Many of the quantities listed above are expected to be relatively static; they could be communicated periodically, and upon change. To conserve network bandwidth and resources at the collector, the PSAMP device may compress the measurement records before export. Compression should be quite effective since the sampled packets may share many fields in common, e.g. if a filter focuses on packets with certain values in particular header fields. Using compression, however, could impact the timeliness of reports. Any consequent delay should not violate the timeliness requirement for availability of packet reports at the collector. 6 Export and Congestion Avoidance 6.1 Collector Destination At least when exporting to a remote collector, the export process is configured to transmit to the collector, as identified by IP address and port number. 6.2 Local Export The report stream may be directly exported to on-board measurement based applications, for example those that for composite statistics from more than one packet. Local export may be presented through an interface direct to the higher level applications, i.e., without employing the transport used for off-board export. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 10] Internet-Draft Passive Packet Measurement November 2002 6.3 Reliable vs. Unreliable Transport The export of the report stream does not does require reliable export. On the contrary, retransmission of lost measurement packets consumes additional network resources and require maintenance of state by the export process. The PSAMP device would have to be addressable, and able to receive and process acknowledgments, and to store unacknowledged data. These requirements would be a significant impediment to having ubiquitous support PSAMP. Instead, it is proposed that PSAMP devices support an unreliable export mechanism. Sequence numbers on the measurement packets would indicate when loss has occurred, and the analysis of the collected measurement data can account for this loss. In some sense, packet loss becomes another form of sampling (albeit a less desirable, and less controlled, form of sampling). 6.4 Limiting Delay in Exporting Measurement Packets The device may queue the report stream in order to export multiple records in a single measurement. Any consequent delay should not violate the timeliness requirement availability of packet reports at the collector. 6.5 Configurable Export Rate Limit The export process must be able to limit its export rate; otherwise it could overload the network and/or the collector. (Note this problem would be exacerbated if using reliable transport mode, since the PSAMP device would retransmit any lost packets, thereby imposing an additional load on the network). At times, the device may generate new records faster than the allowed export rate. In this situation, the device should discard the excess records rather than transmitting them to the collection system. The device may record information (such as sequence numbers, or packet and byte counter values accumulated at the inputs and outputs of a packet selector) to aid the collection system in compensating for the missing data in any subsequent analysis. The export rate must be configurable per export process. Note that since congestion loss can occur at any link on the export path, it is not sufficient to limit rate simply as a function of the bandwidth of the interface out of which export takes place. A candidate for implementation of rate limiting is the leaky bucket, with tokens corresponding e.g. to bytes or packets. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 11] Internet-Draft Passive Packet Measurement November 2002 6.6 Congestion-aware Unreliable Transport Exported measurement traffic competes for resources with other Internet transfers. Congestion-aware export is important to ensure that the measurement records do not overwhelm the capacity of the network or unduly degrade the performance of other applications, while making good use of available bandwidth resources. The PSAMP WG will evaluate (at least) the following alternatives for congestion aware unreliable transport: (i) protocols under development, including the Datagram Congestion Control Protocol (DCCP); see [FHK02] (ii) protocols adopted in the future by the IPFIX WG, (iii) collector-based rate reconfiguration, as now described. 6.7 Collector-based Rate Reconfiguration Since collector-based rate reconfiguration is a new proposal, this draft will discuss it in some detail. The collector can detect congestion loss along the path from the PSAMP device through lost packets, manifest as gaps in the sequence numbers, or the absence of packets for a period of time. The server can run an appropriate congestion-control algorithm to compute a new export rate limit, then reconfigure the PSAMP device with the new rate. This is an attractive alternative to requiring the PSAMP device to receive acknowledgment packets. Implementing the congestion control algorithm in the collection server has the added advantages of flexibility in adapting the sending rate and the ability to incorporate new congestion-control algorithms as they become available. 6.7.1 Changing the Export Rate and Other Rates Forcing the PSAMP device to discard excess records is an effective control under short term congestion. Alternatively, the device could be reconfigured to select fewer packets, and/or send smaller reports on each selected packet. This may be a more appropriate reaction to long-term congestion. In some cases, a collection server may receive measurement records from more than one device, and could decide to reduce the export or other rates at one device rather than another, in order to prioritize the measurement data. This type of flexibility is valuable for network operators that collect measurement data from multiple locations to drive multiple applications. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 12] Internet-Draft Passive Packet Measurement November 2002 6.7.2 Notions of Fairness In some cases, it may be reasonable to allow the collection server to have flexibility in deciding how aggressively to respond to congestion. For example, the PSAMP device and the collection server may have a very small round-trip time relative to other traffic. Conventional TCP-friendly congestion control would allocate a very large share of the bandwidth to this traffic. Instead, the collection server could apply an algorithm that reacts more aggressively to congestion to give a larger share of the bandwidth to other traffic (with larger RTTs). In other cases, the measurement records may require a larger share of the bandwidth than other flows. For example, consider a link that carries tens of thousands of flows, including some non TCP-friendly DoS attack traffic. Restricting the PSAMP traffic to a fair share allocation may be too restrictive, and might limit the collection of the data necessary to diagnose the DoS attack. In order to maintain report collection during periods of congestion, PSAMP report streams may claim more than a fair share of link bandwidth, provided the number of report streams in competition with fair sharing traffic is limited. The collection server could also employ policies that allocate bandwidth in certain proportions amongst different measurement processes. 6.7.3 Behavior Under Overload and Failure The congestion control algorithm has to be robust to severe overload or complete loss of connectivity between the device and the collection system, and also to the failure of the device or the collection system. For example, in a scenario where the collection system is unable to reconfigure the export rate because of loss of reverse (collection system to device) connectivity, it is desirable that the device reduce the export rate automatically. Similarly, if no measurement reports reach the collection system because of loss of forward connectivity, the collection system should not react to this by increasing the export rate. This problem may be solved through periodic heartbeat packets in both directions (i.e., measurement reports in the forward direction, configuration refresh messages in the reverse direction). This allows each side to detect a loss in connectivity or outright failure and to react appropriately. 7 Parallel Measurement Processes Because of the increasing number of distinct measurement applications, with varying requirements, it is desirable to set up parallel measurement processes on a stream of packets. Each process should consist of independently-configurable selection, reporting and export processes. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 13] Internet-Draft Passive Packet Measurement November 2002 Each of the parallel measurement processes should be, as far as possible, independent. However, resource constraints may prevent complete reporting on a packet selected by multiple selection processes. In this case, reporting for the packet must be complete for at least one information flow; other information flows need only report that they selected the packet. The priority amongst information flows to report packets must be configurable. It is not proposed to standardize the number of parallel measurement processes available. 8 Configuration and Management A key requirement for PSAMP is the easy reconfiguration of parameters the parameters of the measurement process: those for selection, packet reports and export. Examples are (i) support of measurement based applications that want to drill-down on traffic detail in real-time; (ii) collector-based rate reconfiguration. To facilitate reconfiguration and retrieval of parameters, they are to reside in a Management Information Base (MIB). CLI and SNMP access to these parameters must be available. 9 Feasibility and Complexity In order for PSAMP to be supported across the entire spectrum of networking equipment, it must be simple and inexpensive to implement. One can envision easy-to-implement instances of the mechanisms described within this draft. Thus, for that subset of instances, it should be straightforward for virtually all system vendors to include them within their products. Indeed, sampling and filtering operations are already realized in available equipment. Here we give some specific arguments to demonstrate feasibility and comment on the complexity of hardware implementations. We stress here that the point of these arguments is not to favor or recommend any particular implementation, or to suggest a path for standardization, but rather to demonstrate that the set of possible implementations is not empty. 9.1 Feasibility 9.1.1 Filtering Filtering consists of a small number of mask (bit-wise logical), comparison and range (greater than) operations. Implementation of at least a small number of such operations is straightforward. 9.1.2 Sampling Sampling based on either counters (counter set, decrement, test for equal to zero) or range matching on the hash of a packet (greater than) are also straightforward given a small number of them. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 14] Internet-Draft Passive Packet Measurement November 2002 9.1.3 Hashing Hashing functions vary greatly in complexity. Execution of a small number of sufficient simple hash functions is implementable at line rate. 9.1.4 Reporting The simplest packet report would duplicate the first n bytes of the packet. However, such an uncompressed format may tax the bandwidth capabilities of the PSAMP device for high sampling rates; reporting selected fields would save on bandwidth within the PSAMP device. Thus there is a trade-off between simplicity and bandwidth limitations within the PSAMP device. 9.1.5 Export Ease of exporting measurement packets depends on the system architecture. Most systems should be able to support PSAMP export by insertion of measurement packets, even through the software path. 9.2 Potential Hardware Complexity We now comment on the complexity of possible hardware implementations. Achieving low constants for performance while minimizing hardware resources is, of course, a challenge, especially at very high clock frequencies. Most of these operations, however, are very basic and their implementations very well understood; in fact, the average ASIC designer simply uses canned library instances of these operations rather then design them from scratch. In addition, networking equipment generally does not need to run at the fastest clock rates, further reducing the effort required to get reasonably efficient implementations. Simple bit-wise logical operations are easy to implement in hardware. Such operations (NAND/NOR/XNOR/NOT) directly translate to four-transistor gates. Each bit of a multiple-bit logical operation is completely independent and thus can be performed in parallel incurring no additional performance cost above a single bit operation. Comparisons (EQ/NEQ) take O(lg(M)) stages of logic, where M is the number of bits involved in the comparison. The lg(M) is required to accumulate the result into a single bit. Greater than operations, as used to determine whether a hash falls in a selection range, are a determination of the most significant not-equivalent bit in the two operands. The operand with that most-significant-not-equal bit set to be one is greater than the other. Thus, a greater than operation is also an O(lg(M)) stages of logic operation. Optimized implementations of arithmetic Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 15] Internet-Draft Passive Packet Measurement November 2002 operations are also O(lg(M)) due to propagation of the carry bit. Setting a counter is simply loading a register with a state. Such an operation is simple and fast O(1). Incrementing or decrementing a counter is a read, followed by an arithmetic operation followed by a store. Making the register dual-ported does take additional space, but it is a well-understood technique. Thus, the increment/decrement is also an O(lg(M)) operation. Hashing functions come in a variety of forms. The computation involved in a standard Cyclic Redundancy Code (CRC) for example are essentially a set of XOR operations, where the intermediate result is stored and XORed with the next chunk of data. There are only O(1) operations and no log complexity operations. Thus, a simple hash function, such as CRC or generalizations thereof, can be implemented in hardware very efficiently. At the other end of the range of complexity, the MD5 function uses a large number of bit-wise conditional operations and arithmetic operations. The former are O(1) operations and the latter are O(lg(M)). MD5 specifies 256 32b ADD operations per 16B of input processed. Consider processing 10Gb/sec at 100MHz (this processing rate appears to be currently available). This requires processing 12.5B/cycle, and hence at least 200 adders, a sizeable number. Because of data dependencies within the MD5 algorithm, the adders cannot be simply run in parallel, thus requiring either faster clock rates and/or more advanced architectures. Thus selection hashing functions as complex as MD5 may be precluded from ubiquitous use at full line rate. This motivates exploring the use of selection hash functions with complexity somewhere between that of MD5 and CRC. However, identification hashing with MD5 on only selected packets is feasible at a sufficiently low sampling rate. 10 Applications We first describe several representative operational applications that require traffic measurements at various levels of temporal and spatial granularity enabled by a PSAMP device. 10.1 Baseline Measurement and Drill Down Packet sampling is ideally suited to determine the composition of the traffic across a network. The approach is to enable measurement on a cut-set of the network such that each packet entering the network is seen at least once, for example, on all ingress and egress links. Unfiltered sampling with a relatively low rate establishes baseline measurements of the network traffic. Reports include packet attributes of common interest: source and destination address and port numbers, prefix, protocol number, type of service, etc. Traffic matrices are indicated by reporting source and destination AS matrices. Absolute traffic volumes are estimated by renormalizing the sampled traffic volumes through division by either the target sampling rate, or the attained sampling rate (as derived by interface packet counters included in the report stream) Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 16] Internet-Draft Passive Packet Measurement November 2002 Suppose an operator or a measurement based application detects an interesting subset of traffic identified by a particular packet attribute. Real-time drill-down to that subset is achieved by instantiating a new measurement process at the PSAMP device from which the subset was reported. The selection process of the new measurement process filters according to the attribute of interest, and composes with sampling if necessary to manage the rate of packet selection. 10.2 Customer Performance Hash-based sampling enables the tracking of the performance experience by customer traffic, customers identified by a list of source or destination prefixes, or by ingress or egress interfaces. Operational uses include the verification of SLAs, and troubleshooting following a customer complaint. In this application, Trajectory Sampling is enabled at all ingress and egress interfaces. The label hash is used to match up ingress and egress samples. Rates of loss in transit between ingress and egress are estimated from the proportion of trajectories for which no egress report is received. Note loss of customer packets is distinguishable from loss of packet reports through use of report sequence numbers. Assuming synchronization of clock between PSAMP devices, delay of customer traffic across the network may also be measured. Extending hash-sampling to all interfaces in the network would enable attribution of poor performance to individual network links. 10.3 Troubleshooting PSAMP can also be used to diagnose problems whose occurrence is evident from aggregate statistics, per interface utilization and packet loss statistics. These statistics are typically moving averages over relatively long time windows, e.g., 5 minutes, and serve as a coarse-grain indication of operational health of the network. The most common method of obtaining such measurements are through the appropriate SNMP MIBs (MIB-II and vendor-specific MIBs.) Suppose an operator detects a link that is persistently overloaded and experiences significant packet drop rates. There is a wide range of potential causes: routing parameters (e.g., OSPF link weights) that are poorly adapted to the traffic matrix, e.g., because of a shift in that matrix; a denial of service attack or a flash crowd; a routing problem (link flapping). In most cases, aggregate link statistics are not sufficient to distinguish between such causes, and to decide on an appropriate corrective action. For example, if routing over two links is unstable, and the links flap between being overloaded and inactive, this might be averaged out in a 5 minute window, indicating moderate loads on both links. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 17] Internet-Draft Passive Packet Measurement November 2002 Baseline PSAMP measurement the congested link, as described in Section 10.1, enables measurements that are fine grained in both space and time. The operator has to be able to determine how many bytes/packets are generated for each source/destination address, port number, and prefix, or other attributes, such as protocol number, MPLS forwarding equivalence class (FEC), type of service, etc. This allows to pinpoint precisely the nature of the offending traffic. For example, in the case of a DDoS attack, the operator would see a significant fraction of traffic with an identical destination address. In certain circumstances, precise information about the spatial flow of traffic through the network domain is required to detect and diagnose problems and verify correct network behavior. In the case of the overloaded link, it would be very helpful to know the precise set of paths that packets traversing this link follow. This would readily reveal a routing problem such as a loop, or a link with a misconfigured weight. More generally, complex diagnosis scenarios can benefit from measurement of traffic intensities (and other attributes) over a set of paths that is constrained in some way. For example, if a multihomed customer complains about performance problems on one of the access links from a particular source address prefix, the operator should be able to examine in detail the traffic from that source prefix which also traverses the specified access link towards the customer. While it is in principle possible to obtain the spatial flow of traffic through auxiliary network state information, e.g., by downloading routing and forwarding tables from routers, this information is often unreliable, outdated, voluminous, and contingent on a network model. For operational purposes, a direct observation of traffic flow is more reliable, as it does not depend on any such auxiliary information. For example, if there was a bug in a router's software, direct observation would allow to diagnose the effect of this bug, while an indirect method would not. Trajectory sampling by enabling common hash-based sampling on all routers in a domain supports such diagnoses. In particular, routing loops are revealed as cycles in trajectories. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 18] Internet-Draft Passive Packet Measurement November 2002 11 References [B88] R.T. Braden, A pseudo-machine for packet monitoring and statistics, in Proc ACM SIGCOMM 1988 [DuGr01] N. G. Duffield and M. Grossglauser, Trajectory Sampling for Direct Traffic Observation, IEEE/ACM Trans. on Networking, 9(3), pp. 280-292, June 2001. [FHK02] S. Floyd, M. Handley. E. Kohler, Problem Statement for DCCP, Internet Draft draft-ietf-dccp-problem-00.txt, work in progress, October 2002. [LCTV02] W.S. Lai, B.Christian, R.W. Tibbs, S. Van den Berghe, A Framework for Internet Traffic Engineering Measurement, Internet Draft draft-ietf-tewg-measure-03.txt, work in progress, September 2002. [PAMM98] V. Paxson, G. Almes, J. Mahdavi, M. Mathis, Framework for IP Performance Metrics, RFC 2330, May 1998 [QZCZCN02] J. Quittek, T. Zseby, B. Claise, S. Zander, G. Carle, K.C. Norseth, Requirements for IP Flow Information Export, Internet Draft draft-ietf-ipfix-reqs-06.txt, work in progress, October 2002. [SPSJTKS01] A. C. Snoeren, C. Partridge, L. A. Sanchez, C. E. Jones, F. Tchakountio, S. T. Kent, W. T. Strayer, Hash-Based IP Traceback, Proc. ACM SIGCOMM 2001, San Diego, CA, September 2001. [ZMR02] T. Zseby, M. Molina, F. Raspall, Sampling and Filtering Techniques for IP Packet Selection, Internet Draft draft-ietf-psamp-sample-tech-00.txt, work in progress, October 2003. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 19] Internet-Draft Passive Packet Measurement November 2002 12 Authors' Addresses Nick Duffield AT&T Labs - Research Room B-139 180 Park Ave Florham Park NJ 07932, USA Phone: +1 973-360-8726 Email: duffield@research.att.com Albert Greenberg AT&T Labs - Research Room A-161 180 Park Ave Florham Park NJ 07932, USA Phone: +1 973-360-8730 Email: albert@research.att.com Matthias Grossglauser AT&T Labs - Research Room A-167 180 Park Ave Florham Park NJ 07932, USA Phone: +1 973-360-7172 Email: mgross@research.att.com Jennifer Rexford AT&T Labs - Research Room A-169 180 Park Ave Florham Park NJ 07932, USA Phone: +1 973-360-8728 Email: jrex@research.att.com Derek Chiou Avici Systems 101 Billerica Ave North Billerica, MA 01862 Phone: +1 978-964-2017 Email: dchiou@avici.com Peram Marimuthu Cisco Systems 170, W. Tasman Drive San Jose, CA 95134 Phone: (408) 527-6314 Email: peram@cisco.com Ganesh Sadasivan Cisco Systems 170 W. Tasman Drive San Jose, CA 95134 Phone: (408) 527-0251 Email: gsadasiv@cisco.com Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 20] Internet-Draft Passive Packet Measurement November 2002 13 Intellectual Property Statement AT&T Corp. may own intellectual property applicable to this contribution. AT&T is currently reviewing its licensing intent relative to the Intellectual Property and will notify the IETF when AT&T has made a determination of that intent. 14 Full Copyright Statement Copyright (C) The Internet Society (1999). All Rights Reserved. This document and translations of it may be copied and furnished to others, and derivative works that comment on or otherwise explain it or assist in its implementation may be prepared, copied, published and distributed, in whole or in part, without restriction of any kind, provided that the above copyright notice and this paragraph are included on all such copies and derivative works. However, this document itself may not be modified in any way, such as by removing the copyright notice or references to the Internet Society or other Internet organizations, except as needed for the purpose of developing Internet standards in which case the procedures for copyrights defined in the Internet Standards process must be followed, or as required to translate it into languages other than English. The limited permissions granted above are perpetual and will not be revoked by the Internet Society or its successors or assigns. This document and the information contained herein is provided on an "AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING TASK FORCE DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Duffield et. al. draft-ietf-psamp-framework-01.txt [Page 21]