Network Working Group B. Volz
Request for Comments: 3074 Ericsson
Category: Standards Track S. Gonczi
Network Engines, Inc.
T. Lemon
Internet Engines, Inc.
R. Stevens
Join Systems, Inc.
February 2001
DHC Load Balancing Algorithm
Status of this Memo
This document specifies an Internet standards track protocol for the
Internet community, and requests discussion and suggestions for
improvements. Please refer to the current edition of the "Internet
Official Protocol Standards" (STD 1) for the standardization state
and status of this protocol. Distribution of this memo is unlimited.
Copyright Notice
Copyright (C) The Internet Society (2001). All Rights Reserved.
Abstract
This document proposes a method of algorithmic load balancing. It
enables multiple, cooperating servers to decide which one should
service a client, without exchanging any information beyond initial
configuration.
The server selection is based on the servers hashing client Media
Access Control (MAC) addresses when multiple Dynamic Host
Configuration Protocol (DHCP) servers are available to service DHCP
clients. The proposed technique provides for efficient server
selection when multiple DHCP servers offer services on a network
without requiring any changes to existing DHCP clients. The same
method is proposed to select the target server of a forwarding agent
such as a Bootstrap Protocol (BOOTP) relay.
1. Introduction
This protocol was originally devised to support a specific load
balancing optimization of the DHCP Failover Protocol [FAILOVR]. The
authors later realized that it could be used to optimize the behavior
of cooperating DHCP servers and the BOOTP relay agents that forward
packets to them. The proposal makes it possible to set up each
participating server to accept a preconfigured (approximate)
percentage of the client load. This is done using a deterministic
hashing algorithm, that could easily be applied to other protocols
having similar characteristics.
2. Terminology
This section discusses both the generic requirements terminology
common to many IETF protocol specifications, and also terminology
introduced by this document.
2.1. Requirements Terminology
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in RFC 2119 [RFC 2119].
2.2. Load Balancing Terminology
This document introduces the following terms:
Service Delay, SD
A load balancing parameter, allowing delayed service of a client
by a server participating in the load-balancing scheme, instead of
ignoring the client.
Hash Bucket Assignments, HBA
A configuration directive that assigns a set of hash bucket values
to a server participating in the load-balancing scheme.
Server ID, SID
An identifier that can be used to designate one of the
participating Servers. In the context of DHCP, the SID is the IP
address or DNS name of the server.
Service Transaction, ST
A set of client-server exchanges that lead to a server providing
or denying some service to a client. Example: the DISCOVER/OFFER/
REQUEST/ACK message exchange between a DHCP server and client is a
service transaction.
Service Transaction ID, STID
An attribute of the individual client requests used for load-
balancing.
3. Background and External Requirements
Because DHCP clients use UDP broadcasts to contact DHCP servers, a
client DHCPDISCOVER message may be received by more than one server.
All servers receiving such a broadcast may respond to the client,
letting the client choose which server it will use.
When a BOOTP relay agent is used, it typically forwards or
rebroadcasts client broadcasts to all configured servers, so a
similar inefficiency is present.
The optimization described allows a server to be chosen for each such
transaction by performing a "serve" / "do not serve" computation. A
forwarding agent can perform the same computation to choose a
forwarding destination.
In either case, the choice of server can be computed, without the
participants having to negotiate who is to respond.
The approach is probabilistic in nature, because it is nearly
impossible to foresee which client will request service next. For
short periods of time, the actual percentage of clients served by a
given server will likely deviate from the desired percentage. As the
number of requests grows, the actual percentage of the load being
handled by each server will approximate the configured percentage.
4. Overview
DHCP servers MUST use the Client Identifier option as the STID if it
is present. If no Client Identifier option is present, the hlen
field of the DHCP packet MUST be used as the length of the data to be
hashed, and the contents of the chaddr MUST be the data to be hashed.
At most the first sixteen bytes of the Client Identifier or chaddr
are used.
The proposal maps the STID into a hash value using the function in
section 6. The resulting hash value can then be used to decide who
should respond to the request, or who the forwarding target should
be.
The provided hash function generates hash values 0 to 255, and yields
a fairly even hash bucket distribution for random STID-s, and also
for STID sequences that have some pattern. Resource allocation is
accomplished by assigning a set of specific hash values to each
participating server.
A server will only service a request if the STID hash of the request
matches one of its assigned hash values.
Any hash buckets not assigned to servers will result in some client
ST-s being entirely ignored. (In some scenarios, this may be a
desirable outcome.) STID-s need not be unique, but should have
sufficient variety to distribute load to each server.
HBA-s MAY be transmitted as messages, encapsulated in messages of
some other protocol, e.g., e-mail, or DHCP Failover Protocol option.
DHCP server implementations may optionally be configurable to handle
a case where load balancing is being done but the server that is
supposed to respond is not available, or is out of suitable
addresses.
DHCP server implementations that provide this capability SHOULD set
the DS (Delayed Service) configuration parameter to the number of
seconds to wait after the client's first request has been sent before
responding to a client, where the hash would not normally permit the
client to be served.
A DHCP server providing this capability SHOULD use the value in the
secs field of the client request if its value is not zero. Because
some clients may not correctly implement the secs field, a DHCP
server MAY keep track of the first instance of a client transaction
to which it would not normally respond. If the server receives a
request from a client that has the same transaction ID as a
previously recorded request, and if the secs field in the second
packet is zero, the DHCP server MAY use the elapsed time (seconds)
between the first and subsequent client request, instead of the secs
field.
5. Operation
5.1 Configuration
The configuration step consists of assigning hash values to available
servers. This is accomplished by providing one or more Hash Bucket
Assignments (HBA-s). These may come from a configuration file, the
Windows NT registry, EEPROM, etc. Alternatively, the hash bucket
values could be assigned using some agreed upon algorithm. E.g.,
"Every odd value is serviced by server A and every even value is
serviced by server B".
5.2 HBA Intended for a Server
When configuring one specific server, an HBA in the form of a simple
bit map of 32 octet values SHOULD be used.
The first octet in the HBA bitmap represents HBA values 0-7, the next
byte values 8-15, and so on, with the thirty-second octet
representing values 248-255. In each octet, the least significant
bit in that octet represents the smallest HBA value in that octet.
Each bit of the HBA is associated with one possible hash value. If a
bit is set in the map, it means the recipient server MUST service
each client request, where the STID yields the corresponding hash
value.
For example, if a server is configured with an HBA of the following
32 octets:
FF FF FF FF FF FF 00 00 ( 0 - 63 )
FF FF FF FF FF FF FF FF ( 64 - 127 )
00 00 00 00 00 00 00 00 (128 - 191 )
00 00 00 00 00 00 00 00 (192 - 255 )
then it MUST service any client requests where the STID hashes into
the bucket values of 0 through 47 and 64 through 127.
5.3 Delayed Service Parameter
The Delayed Service parameter is optional.
If the parameter is not configured, the HBA sets up a strict Serve/Do
not serve policy.
If the parameter is configured, the server that is not supposed to
serve a specific request (based on the HBA and the STID hash), is
allowed to respond, after S seconds have elapsed since the client
first attempted to get service. A server MAY use the secs field in
the BOOTP header for determining the time since the client has been
trying to get service, or it MAY track repeated requests some other
way.
5.4 HBA Intended for a Forwarder
When configuring a forwarding agent, (e.g., BOOTP relay) HBA-s
consisting of pairs of Server-ID / Hash Bucket values MAY be used.
Here, the Server ID (SID) designates the server responsible for the
specified Hash Bucket. The forwarding agent forwards each client
request, where the STID yields the specified hash value, to the
server designated by the SID.
The Server ID may be any unique server attribute, (e.g., IP address,
DNS name, etc.) that is meaningful in the context of the relay agent
operation.
A forwarder may be configured to forward a given packet to more than
one server. For example, a BOOTP relay could be set up to split the
load between 2 primary-backup server pairs, each pair running the
DHCP Failover Protocol [FAILOVR]. In this case, a packet that is
intended for a server pair Will have to be forwarded to both the
primary, and the secondary server of the pair.
A possible configuration file for a forwarding agent (e.g., BOOTP
relay) may look like this:
192.33.43.11 192.33.43.12: 0..24;
192.33.43.13: 25..55;
192.33.43.15: 56..128;
192.33.43.16: 129 130 131 200..202;
The above configuration consists of 4 HBA-s. The first HBA example
reads: "Any Client request, where the STID yields a hash value 0 to
24, will be forwarded to both server 192.33.43.11 and 192.33.43.12".
The 4th HBA example states: "Any Client request, where the STID
yields a hash value 129,139,131,200,201 or 202, will be forwarded to
server 192.33.43.16.
6. Hash Function for Load Balancing
The following hash function is a C language implementation of the
algorithm known as "Pearson's hash". The Pearson's hash algorithm
was originally published in [PEARSON].
The hash function is computationally inexpensive, requires an array
lookup and xor operation for each key byte. To make this proposal
work, all interoperable implementations MUST use this hash function,
with the set of mixing table values given below:
/* A "mixing table" of 256 distinct values, in pseudo-random order. */
unsigned char loadb_mx_tbl[256] ={
251, 175, 119, 215, 81, 14, 79, 191, 103, 49, 181, 143, 186, 157, 0,
232, 31, 32, 55, 60, 152, 58, 17, 237, 174, 70, 160, 144, 220, 90, 57,
223, 59, 3, 18, 140, 111, 166, 203, 196, 134, 243, 124, 95, 222, 179,
197, 65, 180, 48, 36, 15, 107, 46, 233, 130, 165, 30, 123, 161, 209, 23,
97, 16, 40, 91, 219, 61, 100, 10, 210, 109, 250, 127, 22, 138, 29, 108,
244, 67, 207, 9, 178, 204, 74, 98, 126, 249, 167, 116, 34, 77, 193,
200, 121, 5, 20, 113, 71, 35, 128, 13, 182, 94, 25, 226, 227, 199, 75,
27, 41, 245, 230, 224, 43, 225, 177, 26, 155, 150, 212, 142, 218, 115,
241, 73, 88, 105, 39, 114, 62, 255, 192, 201, 145, 214, 168, 158, 221,
148, 154, 122, 12, 84, 82, 163, 44, 139, 228, 236, 205, 242, 217, 11,
187, 146, 159, 64, 86, 239, 195, 42, 106, 198, 118, 112, 184, 172, 87,
2, 173, 117, 176, 229, 247, 253, 137, 185, 99, 164, 102, 147, 45, 66,
231, 52, 141, 211, 194, 206, 246, 238, 56, 110, 78, 248, 63, 240, 189,
93, 92, 51, 53, 183, 19, 171, 72, 50, 33, 104, 101, 69, 8, 252, 83, 120,
76, 135, 85, 54, 202, 125, 188, 213, 96, 235, 136, 208, 162, 129, 190,
132, 156, 38, 47, 1, 7, 254, 24, 4, 216, 131, 89, 21, 28, 133, 37, 153,
149, 80, 170, 68, 6, 169, 234, 151
};
unsigned char loadb_p_hash(
const unsigned char *key, /* The key to be hashed */
const int len ) /* Key length in bytes */
{
unsigned char hash = len;
int i;
for (i=len ; i > 0 ; )
hash = loadb_mx_tbl [ hash ^ key[ --i ] ];
return( hash );
}
int accept_service_request(
const unsigned char HBA[32], /* The hash bucket bitmap */
const unsigned char *key, /* The service transaction id
*/
const int len ) /* length of the above */
{
unsigned char hash = loadb_p_hash(key,len);
int index = (hash >> 3) & 31;
int bitmask = 1 << (hash & 7);
/* return 1 if we should service this transaction */
return((HBA[index] & bitmask) != 0);
}
7. Security Considerations
This proposal in and by itself provides no security, nor does it
impact existing security. Servers using this algorithm are
responsible for ensuring that if the contents of the HBA are
transmitted over the network as part of the process of configuring
any server, that message be secured against tampering, since
tampering with the HBA could result in denial of service for some or
all clients.
8. References
[FAILOVR] Kinnear, K,, Droms, R., Rabil, G., Dooley, M., Kapur, A.,
Gonczi, S. and B. Volz, "DHCP Failover Protocol", Work in
Progress.
[PEARSON] The Communications of the ACM Vol.33, No. 6 (June 1990),
pp. 677-680.
[RFC2131] Droms, R., "Dynamic Host Configuration Protocol", RFC
2131, March 1997.
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels," BCP 14, RFC 2119, March 1997.
9. Acknowledgements
Special thanks to Peter K. Pearson, the author of Pearson's hash who
has kindly granted his permission to use his algorithm, free of any
encumbrances.
This proposal stems from the original idea of hashing MAC addresses
to a single bit by Ted Lemon, during a Failover Protocol discussion
held at CISCO Systems in February, 1999. Rob Stevens suggested the
potential use of this algorithm for purposes beyond those of the
Failover Protocol.
Many thanks to Ralph Droms, Kim Kinnear, Mark Stapp, Glenn Waters,
Greg Rabil and Jack Wong for their comments during the ongoing
discussions.
10. Authors' Addresses
Bernie Volz
Ericsson
959 Concord Street
Framingham, MA 01701
Phone: +1-617-513-9060
EMail: bernie.volz@ericsson.com
Steve Gonczi
Network Engines, Inc.
25 Dan Road Canton, MA 02021-2817
Phone: 781-332-1165
EMail: steve.gonczi@networkengines.com
Ted Lemon
950 Charter Street
Redwood City, CA 94043
EMail: ted.lemon@nominum.com
Rob Stevens
Join Systems, Inc.
1032 Elwell Ct Ste 243 Palo Alto CA 94203
Phone: (650)-968-4470
EMail: robs@join.com
11. Full Copyright Statement
Copyright (C) The Internet Society (2001). 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.
Acknowledgement
Funding for the RFC Editor function is currently provided by the
Internet Society.
|
Comment about this RFC, ask questions, or add new information about this topic: