NetBSD/usr.sbin/named/doc/rfc/rfc1794
1997-04-13 09:06:10 +00:00

396 lines
15 KiB
Plaintext
Raw Blame History

This file contains invisible Unicode characters

This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

Network Working Group T. Brisco
Request for Comments: 1794 Rutgers University
Category: Informational April 1995
DNS Support for Load Balancing
Status of this Memo
This memo provides information for the Internet community. This memo
does not specify an Internet standard of any kind. Distribution of
this memo is unlimited.
1. Introduction
This RFC is meant to first chronicle a foray into the IETF DNS
Working Group, discuss other possible alternatives to
provide/simulate load balancing support for DNS, and to provide an
ultimate, flexible solution for providing DNS support for balancing
loads of many types.
2. History
The history of this probably dates back well before my own time - so
undoubtedly some holes are here. Hopefully they can be filled in by
other authors.
Initially; "load balancing" was intended to permit the Domain Name
System (DNS) [1] agents to support the concept of "clusters" (derived
from the VMS usage) of machines - where all machines were
functionally similar or the same, and it didn't particularly matter
which machine was picked - as long as the load of the processing was
reasonably well distributed across a series of actual different
hosts. Around 1986 a number of different schemes started surfacing
as hacks to the Berkeley Internet Name Domain server (BIND)
distribution. Probably the most widely distributed of these were the
"Shuffle Address" (SA) modifications by Bryan Beecher, or possibly
Marshall Rose's "Round Robin" code.
The SA records, however, did a round-robin ordering of the Address
resource records, and didn't do much with regard to the particular
loads on the target machines. Matt Madison (of TGV) implemented some
changes that used VMS facilities to review the system loads, and
return A RRs in the order of least-loaded to most loaded.
The problem was with SAs was that load was not actually a factor, and
TGV's relied on VMS specific facilities to order the records. The SA
RRs required changes to the DNS specification (in file syntax and in
Brisco [Page 1]
RFC 1794 DNS Support for Load Balancing April 1995
record processing). These were both viewed as drawbacks and not as
general solutions.
Most of the Internet waited in anticipation of an IETF approved
method for simulating "clusters".
Through a few IETF DNS Working Group sessions (Chaired by Rob Austein
of Epilogue), it was collectively agreed upon that a number of
criteria must be met:
A) Backwards compatibility with the existing DNS RFC.
B) Information changes frequently.
C) Multiple addresses should be sent out.
D) Must interact with other RRs appropriately.
E) Must be able to represent many types of "loads"
F) Must be fast.
(A) would ensure that the installed base of BIND and other DNS
implementations would continue to operate and interoperate properly.
(B) would permit very fast update times - to enable modeling of
real-time data. Five minutes was thought as a normal interval,
though changes as fast as every sixty seconds could be imagined.
(C) would cover the possibility of a host's address being advertised
as optimal, yet the machine crashed during the period within the TTL
of the RR. The second-most preferable address would be advertised
second, the third-most preferable third, and so on. This would allow
a reasonable stab at recovery during machine failures.
(D) would ensure correct handling of all ancillary information - such
as MX, RP, and TXT information, as well as reverse lookup
information. It needed to be ensured that such processes as mail
handling continued to work in an unsurprising and predictable manner.
(E) would ensure the flexibility that everyone wished. A breadth of
"loads" were wished to be represented by various members of the DNS
Working Group. Some "loads" were fairly eclectic - such as the
address ordering by the RTT to the host, some were pragmatic - such
as balancing the CPU load evenly across a series of hosts. All
represented valid concerns within their own context, and the idea of
having separate RR types for each was unthinkable (primarily; it
would violate goal A).
Brisco [Page 2]
RFC 1794 DNS Support for Load Balancing April 1995
(F) needed to ensure a few things. Primarily that the time to
calculate the information to order the addressing information did not
exceed the TTL of the information distributed - i.e., that elements
with a TTL of five minutes didn't take six minutes to calculate.
Similarly; it seems a fairly clear goal in the DNS RFC that clients
should not be kept waiting - that request processing should continue
regardless of the state of any other processing occurring.
3. Possible Alternatives
During various discussions with the DNS Working Group and with the
Load Balancing Committee, it was noted that no existing solution
dealt with all wishes appropriately. One of the major successes of
the DNS is its flexibility - and it was felt that this needed to be
retained in all aspects. It was conceived that perhaps not only
address information would need to be changed rapidly, but other
records may also need to change rapidly (at least this could not be
ruled out - who knows what technologies lurk in the future).
Of primary concern to many was the ability to interact with older
implementations of DNS. The DNS is implemented widely now, and
changes to critical portions of the protocol could cause havoc for
years. It became rapidly apparent through conversations with Jon
Postel and Dave Crocker (Area Director) that modifications to the
protocol would be viewed dimly.
4. A Flexible Model
During many hours of discussions, it arose upon suggestion from Rob
Austein that the changes could be implemented without changes to the
protocol; if zone transfer behavior could be subtly changed, then the
zone transfer process could accommodate the changing of various RR
information. What was needed was a smarter program to do the zone
transfers. Pursuant to this, changes were made to BIND that would
permit the specification of the program to do the zone transfers for
particular zones.
There is no specification that a secondary has to receive updates
from its primary server in any specific manner - only that it needs
to check periodically, and obtain new zone copies when changes have
been made. Conceivably the zone transfer agent could obtain the
information from any number of sources (e.g., a load average daemon,
a round-robin sorter) and present the information back to the
nameserver for distribution.
A number of questions arose from this concept, and all seem to have
been dealt with accordingly. Primarily, the DNS protocol doesn't
guarantee ordering. While the DNS protocol doesn't guarantee
Brisco [Page 3]
RFC 1794 DNS Support for Load Balancing April 1995
ordering, it is clear that the ordering is predictive - that
information read in twice in the same order will be presented twice
in the same order to clients. Clients, of course, may reorder this
information, but that is deemed as a "local issue" as it is
configurable by the remote systems administrators (e.g., sortlists,
etc). The zone transfer agent would have to account for any "mis-
ordering" that may occur locally, but remote reordering (e.g., client
side sortlists) of RRs is is impossible to predict. Since local
mis-ordering is consistent, the zone transfer agents could easily
account for this.
Secondarily, but perhaps more subtly, the problem arises that zone
transfers aren't used by primary nameservers, only by secondary
nameservers. To clarify this, the idea of "fast" or "volatile"
subzones must be dealt with. In a volatile environment (where
address or other RR ordering changes rapidly), the refresh rate of a
zone must be set very low, and the TTL of the RRs handed out must
similarly be very low. There is no use in handing out information
with TTLs of an hour, when the conditions for ordering the RRs
changes minutely. There must be a relatively close relationship
between the refresh rates and TTLs of the information. Of course,
with very low refresh rates, zone transfers between the primary and
secondary would have to occur frequently. Given that primary and
secondary nameservers should be topologically and geographically far
apart, moving that much data that frequently is seen as prohibitive.
Also; the longer the propagation time between the primary and
secondary, the larger the window in which circumstances can change -
thus invalidating the secondary's information. It is generally
thought that passing volatile information on to a secondary is fairly
useless - if secondaries want accurate information, then they should
calculate it themselves and not obtain it via zone transfers. This
avoids the problem with secondaries losing contact with the primaries
(but access to the targets of the volatile domain are still
reachable), but the secondary has information that is growing stale.
What is essentially necessary is a secondary (with no primary) which
can calculate the necessary ordering of the RR data for itself (which
also avoids the problem of different versions of domain servers
predictively ordering RR information in different predictive
fashions). For a volatile zone, there is no primary DNS agent, but
rather a series of autonomous secondary agents. Each autonomous
secondary agent is, of course, capable of calculating the ordering or
content of the volatile RRs itself.
Brisco [Page 4]
RFC 1794 DNS Support for Load Balancing April 1995
5. Implementation
With some help from Masataka Ohta (Tokyo Institute of Technology), I
implemented modifications to BIND to permit the specification of the
zone transfer program (zone transfer agent) for particular domains:
transfer <domain-name> <program-name>
Currently I define a separate subdomain that has a few hosts defined
in it - all volatile information. The zone has a refresh rate of
300, and a minimum TTL of 300 indicated. The configuration file is
indicated as "volatile.hosts". Every 300 seconds a program "doAxfer"
is run to do the zone transfer. The program "doAxfer" reads the file
"volatile.hosts.template" and the file "volatile.hosts.list". The
addresses specified in volatile.hosts.list are rotated a random
number of times, and then substituted (in order) into
volatile.hosts.template to generate the file volatile.hosts. The
program "doAxfer" then exits with a value of 1 - to indicate to the
nameserver that the zone transfer was successful, and that the file
should be read in, and the information distributed. This results in
a host having multiple addresses, and the addresses are randomized
every five minutes (300 seconds).
Two bugs continue to plague us in this endeavor. BIND currently
considers any TTL under 300 seconds as "irrational", and substitutes
in the value of 300 instead. This greatly hampers the functionality
of volatile zones. In the fastest of all cases - a 0 TTL -
information would be used once, and then thrown away. Presumably the
new RR information could be calculated every 5 seconds, and the RRs
handed out with a TTL of 0. It must be considered that one
limitation of the speed of a zone is going to be the ability of a
machine to calculate new information fast enough.
The other bug that also effects this is that, as with TTLs, BIND
considers any zone refresh rate under 15 minutes to be similarly
irrational. Obviously zone refresh rates of 15 minutes is
unacceptable for this sort of applications.
For a work-around, the current code sets these same hard-coded values
to 60 seconds. Sixty seconds is still large enough to avoid any
residual bugs associated with small timer values, but is also short
enough to allow fast subzones to be of use.
This version of BIND is currently in release within Rutgers
University, operating in both "fast" and normal zones.
Brisco [Page 5]
RFC 1794 DNS Support for Load Balancing April 1995
6. Performance
While the performance of fast zones isn't exactly stellar, it is not
much more than the normal CPU loads induced by BIND. Testing was
performed on a Sun Sparc-2 being used as a normal workstation, but no
resolvers were using the name server - essentially the nameserver was
idle. For a configuration with no fast subzones, BIND accrued 11 CPU
seconds in 24 hours. For a configuration with one fast zone, six
address records, and being refreshed every 300 seconds (5 minutes),
BIND accrued 1 minute 4 seconds CPU time. For the same previous
configuration, but being refreshed every sixty seconds, BIND accrued
5 minutes and 38 seconds of CPU time.
As is no great surprise, the CPU load on the serving machine was
linear to the frequency of the refresh time. The sixty second
refresh configuration used approximately five times as much CPU time
as did the 300 second refresh configuration. One can easily
extrapolate that the overall CPU utilization would be linear to the
number of zones and the frequency of the refresh period. All of this
is based on a shell script that always indicated that a zone update
was necessary, a more intelligent program should realize when the
reordering of the RRs was unnecessary and avoid such periodic zone
reloads.
7. Acknowledgments
Most of the ideas in this document are the results of conversations
and proposals from many, many people - including, but not limited to,
Robert Austein, Stuart Vance, Masataka Ohta, Marshall Rose, and the
members of the IETF DNS Working Group.
8. References
[1] Mockapetris, P., "Domain Names - Implementation and
Specification", STD 13, RFC 1035, USC/Information Sciences
Institute, November 1987.
Brisco [Page 6]
RFC 1794 DNS Support for Load Balancing April 1995
9. Security Considerations
Security issues are not discussed in this memo.
10. Author's Address
Thomas P. Brisco
Associate Director for Network Operations
Rutgers University
Computing Services, Telecommunications Division
Hill Center for the Mathematical Sciences
Busch Campus
Piscataway, New Jersey 08855-0879
USA
Phone: +1-908-445-2351
EMail: brisco@rutgers.edu
Brisco [Page 7]