Network Working Group
Request for Comments: 1270
F. Kastenholz, Editor
Clearpoint Research Corporation
October 1991

SNMP Communications Services

Status of This Memo

This memo provides information for the Internet community. It does not specify an Internet standard. Distribution of this memo is unlimited.

Table of Contents

   1. Abstract ..............................................    1
   2. Introduction ..........................................    1
   3. Standardization .......................................    3
   4. Interoperability ......................................    3
   5. To Transport or Not To Transport ......................    3
   6. Connection Oriented vs. Connectionless ................    6
   7. Which Protocol ........................................    8
   8. Security Considerations ...............................    9
   9. Appendix ..............................................    9
   10. References ...........................................   10
   11. Acknowledgements .....................................   11
   12. Author's Address .....................................   11

1. Abstract

This memo is being distributed to members of the Internet community as an Informational RFC. The intent is to present a discussion on the issues relating to the communications services for SNMP. While the issues discussed may not be directly relevant to the research problems of the Internet, they may be interesting to a number of researchers and implementors.

2. Introduction

This document discusses various issues to be considered when determining the underlying communications services to be used by an SNMP implementation.

As reported in RFC 1052, IAB Recommendations for the Development of Internet Network Management Standards [8], a two-prong strategy for network management of TCP/IP-based internets was undertaken. In the short-term, the Simple Network Management Protocol (SNMP), defined in RFC 1067, was to be used to manage nodes in the Internet community.

In the long-term, the use of the OSI network management framework was to be examined. Two documents were produced to define the management information: RFC 1065, which defined the Structure of Management Information (SMI), and RFC 1066, which defined the Management Information Base (MIB). Both of these documents were designed so as to be compatible with both the SNMP and the OSI network management framework.

This strategy was quite successful in the short-term: Internet-based network management technology was fielded, by both the research and commercial communities, within a few months. As a result of this, portions of the Internet community became network manageable in a timely fashion.

In May of 1990, the core documents were elevated to "Standard Protocols" with "Recommended" status. As such, the Internet-standard network management framework consists of: Structure and Identification of Management Information for TCP/IP-based internets, RFC 1155 [9], which describes how managed objects contained in the MIB are defined; Management Information Base for Network Management of TCP/IP-based internets, which describes the managed objects contained in the MIB, RFC 1156 [10]; and, the Simple Network Management Protocol, RFC 1157 [1], which defines the protocol used to manage these objects.

In parallel with this activity, documents specifying how to transport SNMP messages over protocols other than UDP/IP have been developed and published: SNMP Over Ethernet [3], SNMP Over OSI [2], and SNMP Over IPX [6] and it would be suprising if more were not developed. These memos have caused a degree of confusion in the community. This document is intended to disperse that confusion by discussing the issues of relevance to an implementor when choosing how to encapsulate SNMP packets.

None of these documents have been made full Internet Standards. SNMP Over ISO and SNMP Over Ethernet are both Experimental protocols. SNMP Over IPX [6] is an Internet Draft. Only the SNMP Specification [1] is an Internet Standard.

No single transport scheme can be considered the absolute best solution for all circumstances. This note will argue that, except for a very small set of special circumstances, operating SNMP over UDP/IP is the optimal scheme.

This document does not present a standard or a protocol for the Internet Community. For production use in the Internet the SNMP and its required communication services are specified in [1].

3. Standardization

Currently, the SNMP Specification [1] only specifies that the UDP protocol be used to exchange SNMP messages. While the IAB may standardize other protocols for use in exchanging SNMP messages in the future, only UDP is currently standardized for this purpose.

In order to claim full compliance with the SNMP Specification, an implementation would have to use UDP for SNMP message exchange.

4. Interoperability

Interoperability is the degree to which the equipment produced by one vendor can can operate with equipment produced by another vendor.

Related to Interoperability is compliance with a standard. Everything else being equal, a device that complies with some standard is more likely to be interoperable than a device that does not.

For commercial product development, the pros and cons of developing an interoperable product must be weighed and a choice made. Both engineering and marketing organizations participate in this process.

The Internet is the single largest market for SNMP systems. A large portion of SNMP systems will be developed with the Internet as a target environment. Therefore, it may be expected that the Internet's needs and requirements will be the driving force for SNMP. SNMP over UDP/IP is specified as the "Internet Standard" protocol. Therefore, in order to operate in the Internet and be managed in that environment on a production basis, a device must support SNMP over UDP/IP. This situation will lead to SNMP over UDP/IP being the most common method of operating SNMP. Therefore, the widest degree of interoperability and widest acceptance of a commercial product will be attained by operating SNMP over UDP/IP.

The preponderance of UDP/IP based network management stations also strongly suggests that an agent should operate SNMP over UDP/IP.

The results of the interoperability decision drive a number of technical decisions. If interoperability is desired, then SNMP must be operated over UDP/IP.

5. To Transport or Not To Transport

A major issue is whether SNMP should run on top of a transport-layer protocol (such as UDP) or not. Typically, the choice is to run over a transport/network/data link protocol or just run over the datalink. In fact, several protocols have been published for operating SNMP over several different datalink and transport protocols.

Operation of SNMP over a Transport and Network protocol stack is preferred. These protocols provide at least five functions that are of vital importance to the movement of SNMP packets through a network:

o Routing

The network layer provides routing functions, which improves the overall utility of network management. The network has the ability to re-route packets around failed areas. This allows network management to continue operating during localized losses of service (It should be noted that these losses of service occur not only because of failures, but also for non-failure reasons such as preventive maintenance).

o Media Independence

The network layer provides a high degree of media independence. By using this capability, many different types of network elements may be managed. Tying SNMP to a particular data link protocol limits the management scope of those SNMP entities to just those devices that use that datalink protocol.

o End-to-End Checksum

The end-to-end checksum provided by transport protocols improves the reliability of the data transfer.

o Multiplexing/Demultiplexing

Transport protocols provide multiplexing and demultiplexing services. These services facilitate the many-to-many management relationships possible with SNMP.

o Fragmentation and Reassembly

This is related to media independence. IP allows SNMP packets to transit media with differing MTU sizes. This capability is not available for datalink specific transmission schemes.

Fragmentation and Reassembly does reduce the overall robustness of network management since, if any single fragment is lost along the way, the operation will fail. The worse the network operates, the higher the probability that a fragment will get lost or delayed. For monitoring and data gathering while the network is operating normally, Fragmentation and Reassembly is not a problem. When the network is operating poorly (and the network operators are typically trying to diagnose and repair a failure), small packets should to be used, preventing the packet from being fragmented.

There are other services and functions that are provided by a connection oriented transport. These services and functions are not desired for SNMP. A later section will explore this issue in more detail.

The main drawbacks that are cited with respect to using Transport and Network layers in the managed object are: a) Increased development time and b) Increased resource requirements. These arguments are less than compelling.

There are several excellent public domain or freely redistributable UDP/IP stacks that provide enough support for SNMP. The effort required to port the essential components of one of these stacks is small compared to the overall effort of installing the SNMP software.

The additional resources required in the managed object to support UDP/IP are minimal. CPU resources are required only when actually transmitting or receiving a packet. The largest single resource requirement of a UDP/IP is calculating the UDP checksum, which is very small compared to the cost of doing the ASN.1 encoding/decoding, Object Identifier lookup, and so on.

The author has personal knowledge of a UDP/IP stack that was developed expressly for the purpose of supporting SNMP. This stack requires less than 4Kb of code space. It is a minimalist implementation of UDP/IP in that it is "just enough" so support SNMP. This stack supports UDP, IP, ARP, and handles ICMP redirect and echo request messages. Furthermore, this stack was developed by a single person in approximately two months. Obviously, neither the development effort nor the memory requirements are large.

The network overhead of using UDP/IP is relatively small. A UDP/IP header requires 28 octets (assuming no IP options). Since the UDP is connectionless, it will generate no overhead traffic of its own (such as TCP SYNs, FINs, and ACKs).

The growing popularity of internetworking outside of The Internet mandates that SNMP operate over, at least, a network layer protocol. These internetworks consist of a number of networks all connected together with routers. In order to traverse a router, a packet must be one of the network layer protocols that the router understands. Therefore, for SNMP management to be deployed in an internetwork, the SNMP entities in that internetwork must use a network layer protocol. SNMP over a datalink can not traverse a router.

There are some circumstances where running SNMP over some datalink is appropriate.

There are schemes are under development to provide Out-Of-Band (OOB) management access to network devices. This OOB access is typically provided over point-to-point or dial-up connections. Since these connections are dedicated to OOB network management and go directly from the network management station to the managed device, a Transport/Network protocol may not be necessary.

Using a Transport/Network protocol on these links may be easier from a development point of view though. It is probably a simple configuration operation to have the management station's IP use a serial port rather than the "normal" (e.g., Ethernet) port for traffic destined for a particular node.

If the Out-Of-Band link is also used as a "primary" route to some nodes, then the functions of a network-layer are required. These functions are readily supplied by using UDP/IP.

For a datalink interface and driver (e.g., a PC Ethernet interface card) that must be manageable independent of the higher level protocol suite (which might NOT be manageable), operating SNMP directly over the datalink is reasonable. It is not known, a priori, what higher-level protocol services may be available, so those services can not be used. If an arbitrary choice is made for example, to put in an elementary UDP/IP stack, then there may be two independent UDP/IPs in the system (which is undesireable as this would require two IP addresses per managed node), or a new protocol stack will be introduced into the environment.

6. Connection Oriented vs. Connectionless

While this section primarily addresses itself to transport layer issues, its basic discussion of connection oriented vs connectionless applies to any layer which provides communication services for SNMP.

For SNMP, connectionless transport service (UDP) is specified in the Protocol Specification [1]. This choice was made after careful study and consideration by the original SNMP developers.

The prime motivation of this choice is that SNMP must continue to operate (if at all possible) when the network is operating at its worst. For other applications, such as Telnet or FTP, the user can always "try again later" if the network is operating poorly. On the other hand, the major purpose of a network management protocol is to fix the network when it is operating poorly so the "try again later" strategy is useless.

By using a connectionless transport protocol, SNMP takes on the responsibility of reliable data transmission (A SNMP application may time out outstanding requests and either retransmit them or abort them as appropriate). However, the SNMP requires these functions only of the sender of a Request PDU (get, getNext, or Set), which typically is a network management station. Since the Agent only generates responses, it need not perform any of these functions. This vastly reduces the resource and functional requirements on the Agent.

If a connection oriented transport is used, then a fundamental design choice must be made with respect to connection maintenance:

(1) Keep a connection open to each managed object on the

network,

(2) Establish and tear down connections on a per-operation

               basis, or

(3) Keep a fixed number of connections open and, when another

connection is needed, use some algorithm (e.g., LRU) to select one for closing and opening to the new agent.

All of these alternatives pose severe problems, and because of them, each is undesirable.

The first option reduces the amount of resources required to perform a single operation in that the connection establishment and termination cost is "amortized" over many operations. On the other hand, keeping a connection open implies that the management station needs to maintain a large number of connection records (in the hundreds or even thousands). Furthermore, if either side of the connection engages in "keep-alives" (even though such behavior is frowned upon), a large amount of traffic will be generated, consuming a large amount of network resources, all for no gain.

The second option reduces the amount of idle resources such as connection records, but vastly increases the amount of resources required to perform an operation. A connection must be established, the request Message sent and the response returned, and then the connection closed for each operation. For a TCP, this would typically require 10 separate packet transfers plus the TCP Time-Wait (see the Appendix for details).

In the face of pathological network problems, a connection oriented transport protocol may simply cease to operate because the probability of getting all of these packets through reduces to a very small number.

The third option requires that the management station maintain connection usage information in order to implement the LRU algorithm. This excessively complicates the management station. Furthermore, this option tends to reduce to the second option when doing health check polling for a number of agents that is large compared to the number of supported connections.

A connection oriented transport protocol may provide services that are undesirable or unneeded by SNMP.

For example, one application of network management is to poll nodes to determine if they are up or not. When a node is up, it makes little difference whether SNMP operates over TCP or UDP. However, if the node goes down then TCP will eventually close the connection. Every poll request must then be translated into a TCP Open request while the managed node is down. Once the node comes up, the send must then be done.

For connection oriented transports, the transport ACK does not necessarily indicate delivery of data to the destination application process (for TCP, see section 2.6 of [4]). The SNMP would still need its own timeout/retry procedure to ensure that the SNMP software actually got the packet.

A connection oriented transport such as TCP provides flow control for the data stream. Because of the lock-step nature of SNMP protocol exchanges, this is not a service that SNMP requires.

Architectural purists may argue that an "Application" layer entity (SNMP) must not perform operations that are properly the realm of the Transport layer (timeouts, retransmissions, and so on). While architecturally pure, this line of reasoning is not relevant. The network management applications and protocols are monitoring the "health" of the network and, as a result, have the best information and are in the best position to adapt their own behavior to the state of the network, and thereby, continuing operations in the face of network adversity.

7. Which Protocol

The final point of discussion is the actual choice of a protocol to support SNMP.

If a device is destined for use in the Internet then it must operate SNMP over UDP/IP.

If the device is operating in some other protocol environment, then SNMP ought to use the transport services that are native to that environment. It may make very little sense to introduce a new protocol stack into a network in order to provide just one service. For example, it could require that the network operations staff understand and be able to administer and operate two protocol stacks, that hosts and routers understand both protocols, and so on. It may also be bureaucratically impossible to introduce UDP/IP into the environment (the "We are only a FOONET shop - if it doesn't speak FOONET, we don't want it" argument).

References [2] and [6] are experimental standards for operating SNMP over IPX and OSI respectively. In these environments, those standards ought to be adhered to.

8. Security Considerations

Security issues are not discussed in this memo.

9. Appendix

This appendix details the TCP packet transfers required to perform a single SNMP operation assuming that the connection is established only for that operation and that a single SNMP operation (e.g., get request) is performed. We also assume that all operations are "normal" i.e., that there are no lost packets, no simultaneous opens, no half opens, and no simultaneous closes. We also ignore the possibility of TCP segmentation and IP fragmentation.

The nomenclature used to illustrate the packet transactions is the same as that used in the TCP Specification [4].

              Packet  Management                         Managed
              Number  Station                            Object
                               Connection Open...
               1         >--<CTL=SYN>----------------------->
               2         <--<CTL=SYN,ACK>-------------------<
               3         >--<CTL=ACK>----------------------->
                           Connection now open,
                           SNMP Request is sent.
               4         >--<DATA=SNMP Request>------------->
                           Response comes back
               5         <--<DATA=SNMP Response, CTL=ACK>---<
               6         >--<CTL=ACK>----------------------->
                           Operation is complete,
                           Management station initiates the
                           close.
               7         >--<CTL=FIN,ACK>------------------->
               8         <--<CTL=ACK>-----------------------<
               9         <--<CTL=FIN,ACK>-------------------<
              10         >--<CTL=ACK>----------------------->
                          Wait 2 MSL
                          Connection now closed.

Some optimizations are possible IF the TCP has knowledge of the type of operation. However, a general purpose TCP would not be tuned to SNMP operations so those optimizations would not be done.

10. References

   [1] Case, J., Fedor, M., Schoffstall, M., and J. Davin, "Simple
   
       Network Management Protocol", RFC 1157, SNMP Research,
       Performance Systems International, Performance Systems
       International, MIT Laboratory for Computer Science, May 1990.
   
   [2] Rose, M., Editor, "SNMP over OSI", RFC 1161, Performance Systems
       International, Inc., June 1990.
   
   [3] Schoffstall, M., Davin, C., Fedor, M., and J. Case, "SNMP over
       Ethernet", RFC 1089, Rensselaer Polytechnic Institute, MIT
       Laboratory for Computer Science, NYSERNet, Inc., University of
       Tennessee at Knoxville, February 1989.
   
   [4] Postel, J., "Transmission Control Protocol - DARPA Internet
       Program Protocol Specification", RFC 793, DARPA, September 1981.
   
   [5] Postel, J., "User Datagram Protocol", RFC 768, USC/Information
       Sciences Institute, August 1980.

[6] Wormley, R., "SNMP Over IPX", draft in process, August 1990.

   [7] Postel, J., Editor, "IAB Official Protocol Standards", RFC 1250,
       IAB, August 1991.
   
   [8] Cerf, V., "IAB Recommendations for the Development of Internet
       Network Management Standards", RFC 1052, NRI, April 1988.
   
   [9] Rose M., and K. McCloghrie, "Structure and Identification of
       Management Information for TCP/IP-based internets", RFC 1155,
       Performance Systems International, Hughes LAN Systems, May 1990.
  
  [10] McCloghrie K., and M. Rose, "Management Information Base for
       Network Management of TCP/IP-based internets", RFC 1156, Hughes
       LAN Systems, Performance Systems International, May 1990.

11. Acknowledgements

The author wishes to thank Jeff Case, Chuck Davin and Keith McCloghrie for their technical and editorial contributions to this document.

12. Author's Address

Frank Kastenholz
Clearpoint Research Corporation
35 Parkwood Drive
Hopkinton, Mass. 01748

Phone:

          (508) 435-2000
   
   Email: kasten@europa.clearpoint.com