Network Working Group
Request for Comments: 2052
Updates: 1035, 1183
Category: Experimental
A. Gulbrandsen
Troll Technologies
P. Vixie
Vixie Enterprises
October 1996

A DNS RR for specifying the location of services (DNS SRV)

Status of this Memo

This memo defines an Experimental Protocol for the Internet community. This memo does not specify an Internet standard of any kind. Discussion and suggestions for improvement are requested. Distribution of this memo is unlimited.


This document describes a DNS RR which specifies the location of the server(s) for a specific protocol and domain (like a more general form of MX).

Overview and rationale

Currently, one must either know the exact address of a server to contact it, or broadcast a question. This has led to, for example, aliases, the SMTP-specific MX RR, and using MAC- level broadcasts to locate servers.

The SRV RR allows administrators to use several servers for a single domain, to move services from host to host with little fuss, and to designate some hosts as primary servers for a service and others as backups.

Clients ask for a specific service/protocol for a specific domain (the word domain is used here in the strict RFC 1034 sense), and get back the names of any available servers.

Introductory example

When a SRV-cognizant web-browser wants to retrieve
   it does a lookup of

and retrieves the document from one of the servers in the reply. The example zone file near the end of the memo contains answering RRs for this query.

The format of the SRV RR

Here is the format of the SRV RR, whose DNS type code is 33:

Service.Proto.Name TTL Class SRV Priority Weight Port Target

(There is an example near the end of this document.)


The symbolic name of the desired service, as defined in Assigned Numbers or locally.

Some widely used services, notably POP, don't have a single universal name. If Assigned Numbers names the service indicated, that name is the only name which is legal for SRV lookups. Only locally defined services may be named locally. The Service is case insensitive.


TCP and UDP are at present the most useful values
for this field, though any name defined by Assigned Numbers or locally may be used (as for Service). The Proto is case insensitive.


The domain this RR refers to. The SRV RR is unique in that the name one searches for is not this name; the example near the end shows this clearly.


Standard DNS meaning.


Standard DNS meaning.


As for MX, the priority of this target host. A client MUST attempt to contact the target host with the lowest-numbered priority it can reach; target hosts with the same priority SHOULD be tried in pseudorandom order. The range is 0-65535.


Load balancing mechanism. When selecting a target host among the those that have the same priority, the chance of trying this one first SHOULD be proportional to its weight. The range of this number is 1-65535. Domain administrators are urged to use Weight 0 when there isn't any load balancing to do, to make the RR easier to read for humans (less noisy).


The port on this target host of this service. The range is

0-65535. This is often as specified in Assigned Numbers but

need not be.


As for MX, the domain name of the target host. There MUST be one or more A records for this name. Implementors are urged, but not required, to return the A record(s) in the Additional Data section. Name compression is to be used for this field.

A Target of "." means that the service is decidedly not available at this domain.

Domain administrator advice

Asking everyone to update their telnet (for example) clients when the first internet site adds a SRV RR for Telnet/TCP is futile (even if desirable). Therefore SRV will have to coexist with A record lookups for a long time, and DNS administrators should try to provide A records to support old clients:

  • Where the services for a single domain are spread over several hosts, it seems advisable to have a list of A RRs at the same DNS node as the SRV RR, listing reasonable (if perhaps suboptimal) fallback hosts for Telnet, NNTP and other protocols likely to be used with this name. Note that some programs only try the first address they get back from e.g. gethostbyname(), and we don't know how widespread this behaviour is.
  • Where one service is provided by several hosts, one can either provide A records for all the hosts (in which case the round- robin mechanism, where available, will share the load equally) or just for one (presumably the fastest).
  • If a host is intended to provide a service only when the main server(s) is/are down, it probably shouldn't be listed in A records.
  • Hosts that are referenced by backup A records must use the port number specified in Assigned Numbers for the service.

Currently there's a practical limit of 512 bytes for DNS replies. Until all resolvers can handle larger responses, domain administrators are strongly advised to keep their SRV replies below 512 bytes.

All round numbers, wrote Dr. Johnson, are false, and these numbers are very round: A reply packet has a 30-byte overhead plus the name of the service ("" for instance); each SRV RR adds 20 bytes plus the name of the target host; each NS RR in the NS section is 15 bytes plus the name of the name server host; and finally each A RR in the additional data section is 20 bytes or so, and there are A's for each SRV and NS RR mentioned in the answer. This size estimate is extremely crude, but shouldn't underestimate the actual answer size by much. If an answer may be close to the limit, using e.g. "dig" to look at the actual answer is a good idea.

The "Weight" field

Weight, the load balancing field, is not quite satisfactory, but the actual load on typical servers changes much too quickly to be kept around in DNS caches. It seems to the authors that offering administrators a way to say "this machine is three times as fast as that one" is the best that can practically be done.

The only way the authors can see of getting a "better" load figure is asking a separate server when the client selects a server and contacts it. For short-lived services like SMTP an extra step in the connection establishment seems too expensive, and for long-lived services like telnet, the load figure may well be thrown off a minute after the connection is established when someone else starts or finishes a heavy job.

The Port number

Currently, the translation from service name to port number happens at the client, often using a file such as /etc/services.

Moving this information to the DNS makes it less necessary to update these files on every single computer of the net every time a new service is added, and makes it possible to move standard services out of the "root-only" port range on unix.

Usage rules

A SRV-cognizant client SHOULD use this procedure to locate a list of servers and connect to the preferred one:

        Do a lookup for, QCLASS=IN,

If the reply is NOERROR, ANCOUNT>0 and there is at least one SRV RR which specifies the requested Service and Protocol in the reply:

If there is precisely one SRV RR, and its Target is "." (the root domain), abort.

Else, for all such RR's, build a list of (Priority, Weight, Target) tuples

Sort the list by priority (lowest number first)

Create a new empty list

For each distinct priority level

While there are still elements left at this priority level

Select an element randomly, with probability Weight, and move it to the tail of the new list

For each element in the new list

query the DNS for A RR's for the Target or use any RR's found in the Additional Data secion of the earlier SRV query.

for each A RR found, try to connect to the (protocol, address, service).

else if the service desired is SMTP

skip to RFC 974 (MX).


Do a lookup for QNAME=target, QCLASS=IN, QTYPE=A

for each A RR found, try to connect to the (protocol, address, service)


  • Port numbers SHOULD NOT be used in place of the symbolic service or protocol names (for the same reason why variant names cannot be allowed: Applications would have to do two or more lookups).
  • If a truncated response comes back from an SRV query, and the Additional Data section has at least one complete RR in it, the answer MUST be considered complete and the client resolver SHOULD NOT retry the query using TCP, but use normal UDP queries for A RR's missing from the Additional Data section.
  • A client MAY use means other than Weight to choose among target hosts with equal Priority.
  • A client MUST parse all of the RR's in the reply.
  • If the Additional Data section doesn't contain A RR's for all the SRV RR's and the client may want to connect to the target host(s) involved, the client MUST look up the A RR(s). (This happens quite often when the A RR has shorter TTL than the SRV or NS RR's.)
  • A future standard could specify that a SRV RR whose Protocol was TCP and whose Service was SMTP would override RFC 974's rules with regard to the use of an MX RR. This would allow firewalled organizations with several SMTP relays to control the load distribution using the Weight field.
  • Future protocols could be designed to use SRV RR lookups as the means by which clients locate their servers.

Fictional example

This is (part of) the zone file for, a still-unused domain:

        @               SOA (
                            1995032001 3600 3600 604800 86400 )
        ftp.tcp         SRV 0 0 21
        finger.tcp      SRV 0 0 79
        ; telnet - use old-slow-box or new-fast-box if either is
        ; available, make three quarters of the logins go to
        ; new-fast-box.
        telnet.tcp      SRV 0 1 23
                        SRV 0 3 23
        ; if neither old-slow-box or new-fast-box is up, switch to
        ; using the sysdmin's box and the server
                        SRV 1 0 23
                        SRV 1 0 23
        ; HTTP - server is the main server, new-fast-box is the backup
        ; (On new-fast-box, the HTTP daemon runs on port 8000)
        http.tcp        SRV 0 0 80
                        SRV 10 0 8000
        ; since we want to support both and
        ; we need the next two RRs as well
        http.tcp.www    SRV 0 0 80
                        SRV 10 0 8000
        ; SMTP - mail goes to the server, and to the IP provider if
        ; the net is down
        smtp.tcp        SRV 0 0 25
                        SRV 1 0 25
        @               MX  0
                        MX  1
        ; NNTP - use the IP providers's NNTP server
        nntp.tcp        SRV 0 0 119
        ; IDB is an locally defined protocol
        idb.tcp         SRV  0 0 2025
        ; addresses
        server          A
        old-slow-box    A
        sysadmins-box   A
        new-fast-box    A
        ; backup A records - new-fast-box and old-slow-box are
        ; included, naturally, and server is too, but might go
        ; if the load got too bad
        @               A
        ; backup A RR for
        www             A
        ; NO other services are supported
        *.tcp           SRV  0 0 0 .
        *.udp           SRV  0 0 0 .

In this example, a telnet connection to "" needs an SRV lookup of "" and possibly A lookups of "new-" and/or the other hosts named. The size of the SRV reply is approximately 365 bytes:

30 bytes general overhead
20 bytes for the query string, "" 130 bytes for 4 SRV RR's, 20 bytes each plus the lengths of "new-

        fast-box", "old-slow-box", "server" and "sysadmins-box" -
        "" in the query section is quoted here and doesn't
        need to be counted again.
      75 bytes for 3 NS RRs, 15 bytes each plus the lengths of
        "server", "" and "ns2" - again, "ip-" is quoted and only needs to be counted once.
      120 bytes for the 6 A RR's mentioned by the SRV and NS RR's.


   RFC 1918: Rekhter, Y., Moskowitz, R., Karrenberg, D., de Groot, G.,
        and E.  Lear, "Address Allocation for Private Internets",
        RFC 1918, February 1996.
   RFC 1916 Berkowitz, H., Ferguson, P, Leland, W. and P. Nesser,
        "Enterprise Renumbering: Experience and Information
        Solicitation", RFC 1916, February 1996.
   RFC 1912 Barr, D., "Common DNS Operational and Configuration
        Errors", RFC 1912, February 1996.
   RFC 1900: Carpenter, B., and Y. Rekhter, "Renumbering Needs Work",
        RFC 1900, February 1996.
        STD 1, RFC 1920, March 1996.
   RFC 1814: Gerich, E., "Unique Addresses are Good", RFC 1814, June

RFC 1794: Brisco, T., "DNS Support for Load Balancing", April 1995.

RFC 1713: Romao, A., "Tools for DNS debugging", November 1994.

   RFC 1712: Farrell, C., Schulze, M., Pleitner, S., and D. Baldoni,
        "DNS Encoding of Geographical Location", RFC 1712, November
   RFC 1706: Manning, B. and R. Colella, "DNS NSAP Resource Records",
        RFC 1706, October 1994.
   RFC 1700: Reynolds, J., and J. Postel, "ASSIGNED NUMBERS",
        STD 2, RFC 1700, October 1994.
   RFC 1183: Ullmann, R., Mockapetris, P., Mamakos, L., and
        C. Everhart, "New DNS RR Definitions", RFC 1183, November
   RFC 1101: Mockapetris, P., "DNS encoding of network names and other
        types", RFC 1101, April 1989.
   RFC 1035: Mockapetris, P., "Domain names - implementation and
        specification", STD 13, RFC 1035, November 1987.
   RFC 1034: Mockapetris, P., "Domain names - concepts and
        facilities", STD 13, RFC 1034, November 1987.
   RFC 1033: Lottor, M., "Domain administrators operations guide",
        RFC 1033, November 1987.
   RFC 1032: Stahl, M., "Domain administrators guide", RFC 1032,
        November 1987.
   RFC 974: Partridge, C., "Mail routing and the domain system",
        STD 14, RFC 974, January 1986.

Security Considerations

The authors believes this RR to not cause any new security problems. Some problems become more visible, though.

  • The ability to specify ports on a fine-grained basis obviously changes how a router can filter packets. It becomes impossible to block internal clients from accessing specific external services, slightly harder to block internal users from running unautorised services, and more important for the router operations and DNS operations personnel to cooperate.
  • There is no way a site can keep its hosts from being referenced as servers (as, indeed, some sites become unwilling secondary MXes today). This could lead to denial of service.
  • With SRV, DNS spoofers can supply false port numbers, as well as host names and addresses. The authors do not see any practical effect of this.

We assume that as the DNS-security people invent new features, DNS servers will return the relevant RRs in the Additional Data section when answering an SRV query.

Authors' Addresses

Arnt Gulbrandsen
Troll Tech
Postboks 6133 Etterstad
N-0602 Oslo

   Phone: +47 22646966
   Paul Vixie
   Vixie Enterprises
   Star Route 159A
   Woodside, CA  94062
   Phone: (415) 747-0204