Working Group: ARMD Himanshu Shah
Intended Status: Proposed Standard Ciena Corp
Internet Draft
Anoop Ghanwani
Expiration Date: May, 2011 Brocade
Nabil Bitar
Verizon
October 25, 2010
ARP Broadcast Reduction for Large Data Centers
draft-shah-armd-arp-reduction-01.txt
Status of this Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
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/1id-abstracts.html
The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html
This Internet-Draft will expire on May 25, 2011
Copyright Notice
Copyright (c) 2010 IETF Trust and the persons identified as the
document authors. All rights reserved.
Shah, et al. Expires May 2011 1
Internet Draft draft-shah-arp-reduction-01.txt
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with
respect to this document. Code Components extracted from this
document must include Simplified BSD License text as described in
Section 4.e of the Trust Legal Provisions and are provided without
warranty as described in the Simplified BSD License.
Abstract
With the emergence server virtualization technologies, a host is
able to support multiple Virtual Machines (VMs) in a single physical
machine. Data centers can leverage these capabilities to instantiate
on the order of 10s to 100s of VMs in a server. Each VM operates as
an independent IP host with a set of Virtual Network Interface Cards
(vNICs), each having its own MAC address and mapping to a physical
Ethernet interface. These physical servers are typically installed
in a rack with their Ethernet interfaces connected to a top-of-rack
(ToR) switch. The ToR switches are interconnected through End-of-
the-Row (EoR) or aggregation switches which are in turn connected to
core switches.
As discussed in [ARP-Problem] the host VMs use ARP broadcasts to
find other host VMs and use periodic (broadcast) Gratuitous ARPs to
refresh their IP to MAC address binding in other VM hosts. Such
broadcasts in a large data center with potentially thousands of VM
hosts in a Layer 2 based topology can overwhelm the network.
This memo proposes mechanisms to reduce the number of broadcasts
that are sent throughout the network. This is done by having the ToR
switches intelligently process ARP packets, rather than simply
broadcasting them throughout the broadcast domain.
While this document specifically addresses ARP, the Neighbor
Discovery mechanisms used by IPv6 hosts that make use of multicast
rather than broadcast also pose similar issues for the data center.
The solutions defined herein should be equally applicable to hosts
running IPv6. The details will be specified in a subsequent
revision.
Conventions
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].
Shah, et al. Expires May 2011 2
Internet Draft draft-shah-arp-reduction-01.txt
Table of Contents
Copyright Notice .................................................... 1
Abstract.............................................................. 2
1.0 Overview.......................................................... 3
1.1 Terminology ..................................................... 5
2.0 Configuration..................................................... 6
3.0 Building the ARP Tables........................................... 6
3.1 ARP Request ..................................................... 6
3.2 ARP Reply ....................................................... 7
3.3 Gratuitous ARP .................................................. 7
3.4 Uplink Versus Downlink Processing ............................... 8
3.5 Host Mobility ................................................... 8
4.0 Concluding Remarks................................................ 9
5.0 Security Considerations ......................................... 10
6.0 Acknowledgments ................................................. 10
7.0 References....................................................... 10
7.1 Normative References ........................................... 10
7.2 Informative References ......................................... 10
8.0 Author's Address................................................. 10
1.0 Overview
The traditional topology in a data center consists of racks of
servers connected to top-of-rack (ToR) switches, which connect to
aggregation switches, which in turn connect to core switches. The
network architecture is typically a combination Layer 2 and Layer 3
functionality. In some architectures, Layer 2 is terminated at the
ToR, with Layer 3 being run in the aggregation and core devices. In
other architectures, Layer 2 may be extended all the way to the
aggregation switch. The primary concerns that have influenced
network architectures in the data center have been keeping broadcast
domains manageable and the spanning tree diameter contained.
Moving forward, these traditional network architectures are being
challenged due to emerging technologies such as server
virtualization.
Shah, et al. Expires May 2011 3
Internet Draft draft-shah-arp-reduction-01.txt
The effect of server virtualization in the data center brings some
challenges. Because of virtualization, the number of hosts seen by
the network increases dramatically - 10 to 100 times the number of
physical servers. These virtual hosts are referred to as Virtual
machines (VMs). In addition, virtualized environments offer a
feature referred to as "VM mobility" wherein a VM can be relocated
to run on a different physical server. In order for the VM mobility
to be non-disruptive to other hosts that have communication in
progress with the VM being moved, the VM must retain its MAC address
and IP address. Because of the requirement to retain the MAC and IP
address, it is desirable to develop network architectures that would
offer the least restrictions in terms of VM mobility.
As an example, in a network architecture where TOR switches
terminate the L2 domain, the range of VM mobility would be
restricted to a single ToR switch. It would be more preferable to
allow the flexibility of moving the VM anywhere within the data
center, or perhaps even a different data center.
Technologies such as TRILL [TRILL] overcome some of the issues of
spanning trees that forced traditional Layer 2 topologies to be
severely constrained. However, because of virtualization there are
2 specific problems that are introduced with respect to broadcast
traffic.
1. A larger number of hosts. A single physical server now hosts
multiple VMs taking the scale factor to a different level. If
each VM issues the same number of broadcasts as a physical
server, the amount of broadcast traffic will increased 10 to
greater than 100 times.
2. If the Layer 2 domains are extended to go across data centers,
then broadcast traffic will now go across the backbone. If
Layer 2 was terminated at the ToR switch, the increase in
broadcast traffic would be been restricted to a single ToR
switch, but as discussed earlier, this restriction is not
desirable.
Excessive broadcast traffic in Layer 2 networks results in wastage
of network bandwidth, as well as in the wastage of CPU resources due
to all of the VMs processing superfluous ARP broadcasts (IPv6 gets
rid of the latter by running ND as a multicast service rather than a
broadcast service).
The solution presented here attempts to minimize the negative
effects of ARP broadcast packets. The solution requires the first
hop Ethernet switches, typically the ToR switch, to maintain an ARP
table that is learned from the ARP packets received by the switch.
The switch then selectively propagates the ARP packet to, or proxy-
responds on behalf of, the remote peer. These types of ARP
processing principles are well-known and are described in L2VPN
Working Group documents such as [ARP-Mediation] and [IPLS].
Shah, et al. Expires May 2011 4
Internet Draft draft-shah-arp-reduction-01.txt
The following sections describe the details of ARP snooping, the
learning and maintenance of ARP tables, the use of learned
information to limit broadcast propagation, and proxy (the response)
on behalf of the remote peers.
1.1 Terminology
ToR Top-of-Rack. An Ethernet switch present on top
of a rack which provides network connectivity to
the servers present on the rack.
Downlink The Ethernet link between the ToR switch and a
directly connected host (server in the rack).
Uplink The network- facing Ethernet connection in the
ToR switch. Typically, the uplinks from ToRs
connect to end-of-row or aggregation switches.
EoR End-of-Row. An Ethernet switch to which the
ToR switches connect, also referred to as an
aggregation switch. Uplinks from ToR switches
connect to an EoR switch and uplinks from EoR
switches connect to a core switch.
Host/Server A host or server running the IP protocol. This
could be a physical entity or a logical entity
(such as a Virtual Machine) in a physical host.
The term server refers to its role in the data
center. Both terms are used interchangeably to
refer to an IP host.
Local hosts Used in the context of a ToR switch to denote
the VM hosts connected to a ToR on the
downlink, i.e. directly attached hosts.
Remote hosts Used in the context of a ToR switch to denote
the hosts that are accessible through uplink of
the ToR.
VM Virtual Machine. This is a logical instance of
a host that operates independently in a
physical host and has its own IP and MAC
addresses. VMs allow efficient use of physical
host resources (such as multiple CPU cores).
Shah, et al. Expires May 2011 5
Internet Draft draft-shah-arp-reduction-01.txt
2.0 Configuration
It is assumed that ARP reduction mechanisms that are defined in this
document will be limited to ToR switches. The maximum benefit of
restraining ARP broadcasts in the network is achieved by the first
hop switches (the ones directly connected to the hosts) without
placing additional burden on second or third tier switches.
First, the ToR switches would need to be configured in order to
enable the ARP reduction feature. Every Ethernet interface needs to
be identified as either a downlink or uplink within the context of
this feature.
In addition the operator may optionally configure various ARP
reduction related parameters such as:
. ARP aging timer.
. Size of the ARP table.
. Static entries of IP to MAC address.
3.0 Building the ARP Tables
When ARP reduction is enabled, the ToR switch will monitor all ARP
traffic transiting the switch (regardless of uplink port or downlink
port) and will process any ARP packets in the following manner:
. ARP Request packets must be redirected to control plane CPU.
. Gratuitous ARP packets (ARP Reply packet with a broadcast MAC
DA) must be redirected to control plane CPU.
. Other ARP Reply packets (ARP Reply packet with a unicast MAC
DA) should be bi-casted; one copy sent to control plane CPU and
other copy forwarded out normally.
3.1 ARP Request
The ToR examines the source IP and the source hardware address (MAC
address) in the ARP Request . The source IP and MAC address
association is learned, or is updated/refreshed if already learned.
The destination IP address is searched in the ARP table. If an entry
exists, the associated MAC address from the table is used to prepare
a unicast ARP Reply packet. The same MAC address is used as the
source MAC address in the MAC header, as well as for the target
hardware address, in the unicast ARP Reply packet.
If the destination IP address in the ARP Request is not present in
the ARP table, then the original ARP Request packet is broadcast to
all the switch ports that are members of the same VLAN except the
source port that the ARP Request was received from. However, if the
requested (destination) IP address is present in the ARP table, a
unicast ARP Reply packet is prepared as described above and sent to
the switch port from which the ARP Request was received and original
ARP Request packet is dropped.
Shah, et al. Expires May 2011 6
Internet Draft draft-shah-arp-reduction-01.txt
The intent is to prevent propagation of ARP Request broadcasts as
much as possible using the information present in the ARP table. The
following observations can be made from such behavior.
. Most of the ARP Request packets from the local hosts of a ToR
switch for the local hosts of that ToR switch can be prevented
from being broadcast on uplinks or downlinks.
. Most of the ARP Request packets from remote hosts of a ToR
switch for local hosts of that ToR switch can be prevented
from being broadcast on downlinks or other uplinks of the ToR
switch.
. Many of the ARP Request packets from local hosts of a ToR
switch for remote hosts of that ToR switch can be prevented
from being forwarded on uplinks if the remote host IP to MAC
association is known to the ToR switch.
3.2 ARP Reply
The unicast ARP Reply is examined to learn/update the ARP table for
source and destination IP/MAC address association, but is also
forwarded out as a normal frame.
3.3 Gratuitous ARP
Gratuitous ARP is a broadcast ARP Reply packet with the destination
IP address set to the IP address of the sender and target hardware
address set to the MAC address of the sender. It is typically used
by IP hosts (including VMs) to keep its IP-to-MAC address
association fresh in its peers' ARP cache.
The ToR switch should process Gratuitous ARP in the following
manner.
. Learn/update/refresh the ARP table entry.
. If the IP address is new, or exists but with a different
hardware address, then the Gratuitous ARP packet is forwarded
out; otherwise the packet is discarded.
The goal for handling of Gratuitous ARP packets received from the
downlinks (i.e. local hosts) is to avoid propagating it into the
'network' (i.e. to the uplinks), unless there is a new association.
By suppressing the propagation of Gratuitous ARP packets, the peer
IP hosts will end up aging out the corresponding ARP table entries.
This will result in generation of the broadcast ARP Requests by
those IP hosts if they need to continue to communicate with the IP
host whose Gratuitous ARPs were obstructed. The handling of the ARP
Request by the first-hop ToR switch, as described above, will be
able to respond to this request based on the ARP cache maintained in
the ToR switch. In essence, the presence of large ARP tables with
longer aging times compensates for the smaller ARP table present in
Shah, et al. Expires May 2011 7
Internet Draft draft-shah-arp-reduction-01.txt
the IP hosts and eliminates the need for periodic use of Gratuitous
ARPs in order to refresh the ARP table in the IP hosts.
3.4 Uplink Versus Downlink Processing
With respect to processing of the ARP packets as described above,
the behavior is different depending on whether the packet was
received from an uplink or downlink in the following ways.
. The aging timer will typically be higher for entries learned
from an uplink versus those learned from a downlink. The
reason for this is to avoid flooding ARP broadcast packets on
uplinks since they have a much larger negative impact.
. If ARP table fills up, then entries learned from downlinks
(i.e. directly attached hosts) will take precedence over those
learned from an uplink (i.e. remote hosts). This will trade
off sending broadcasts on host links versus sending them into
the core of the network. The reason for this is that access
links are typically lower bandwidth, and also this will
conserve CPU resources involved in processing unnecessary ARP
traffic.
3.5 Host Mobility
As mentioned earlier, server virtualization technology allows
mobility of VMs to different physical servers. The flexibility to
move VMs is one of the key benefits of server virtualization. VM
mobility could be manual (operator initiated) or may be done
automatically in reaction to demands placed by the application
users. The important point is that in either case, VM movement is
not transparent and is made known to the network.
There is ongoing work in IEEE 802.1 standards organization (IEEE
802.1Qbg) to coordinate/communicate the presence and capabilities of
the VMs to the directly connected network switch.
VMs typically retain their MAC and IP address across a VM mobility
event, and as such, there would be little impact to the ARP table
maintained by the ARP reduction mechanism described herein.
However, the ARP reduction mechanism would benefit from knowing if a
VM is completely decommissioned so that the ToR switch can remove
the ARP entry that it has for that VM in a timely fashion, rather
than waiting for it to age out.
3.6 Scaling Considerations
Depending on the number of hosts in the network, the ARP table in a
ToR switch needed for the ARP reduction mechanisms described above
can be quite large. Although it is possible to implement some of the
mechanisms for ARP reduction in hardware in the forwarding plane,
Shah, et al. Expires May 2011 8
Internet Draft draft-shah-arp-reduction-01.txt
the number of ARP entries favors maintaining the ARP table in the
control plane memory.
3.7 Miscellaneous Issues
Because of the distributed nature of the mechanisms described
herein, there are a few additional issues that warrant consideration
from the network operator.
Earlier in the document, we had mentioned the configuration of a
aging timer for ARP entries. A longer timer for holding onto ARP
entries helps with reduction of broadcasts. However, having a "too
large timer" can lead to problems in certain situations.
Consider the following scenario. Host A is attached to ToR switch
#1, and host B is attached to ToR switch #2. If host B issues an
ARP Request for host A, and if the entry is available at switch #2,
then switch #2 would send the ARP Reply on behalf of host A. It is
possible that host A is no longer available, but there is no way for
switch #2 to know this, and it would continue to respond on behalf
of host A, until its entry for host A has aged out. In this case,
it is easy to see that a smaller aging timer would be beneficial.
Additionally, since host B has an ARP aging timer, it means that
host B would find out about host A's unavailability only after its
entry has aged out, which would be some time after it the entry has
aged out of switch #2.
Another issue that can be somewhat problematic could be the
inconsistency of tables in switches. Once again, consider a
scenario similar to the one described above with two hosts each
connected to its respect ToR switch. Let the ARP entries at both A
and B be learned by both switches. Now assume that the IP address
on host A changes. This change is signaled to switch #1 which in
turn broadcasts the message on its uplink. Now, if this message is
discarded due to network congestion or signal integrity issues, then
switch #2 will not learn about the change and will continue to
respond to host B's ARP Requests for host A's old IP address with
stale information. This lasts until the ARP entry for A ages out at
Switch #2.
4.0 Concluding Remarks
Based on the procedures described in this document, it is possible
for ToR switches in the data center to contain ARP broadcasts
significantly. The solution is based on well known, non-intrusive
procedures and strives to curtail ARP broadcasts that are
increasingly becoming a cause for concern in the data centers. In
essence, ToR switches offload some of the ARP table management from
the IP hosts to themselves. The ARP table aging timer can be tuned
higher by the operator based on the available switch resources and
network traffic behavior. The larger capacity of the ARP table
Shah, et al. Expires May 2011 9
Internet Draft draft-shah-arp-reduction-01.txt
coupled with a long aging time for entries in the table directly
translates to more effective subduing of the ARP broadcasts.
5.0 Security Considerations
Security aspects will be addressed in a subsequent revision.
6.0 Acknowledgments
This document resulted from discussions with Linda Dunbar (Huawei),
Sue Hares (Huawei), and T Sridhar (Force10). We would like to
acknowledge their contribution to this work.
7.0 References
7.1 Normative References
[ARP] D. Plummer, "An Ethernet Address Resolution Protocol: Or
Converting Network Protocol Addresses to 48.bit Ethernet
Addresses for Transmission on Ethernet Hardware, " RFC 826 (also
STD 37), November 1982.
[ARP-Problem] L.Dunbar et al., "Scalable Address Resolution for
Large Data Center Problem Statements," <draft-dunbar-arp-for-
large-dc-problem-statement-00>, July 2010.
7.2 Informative References
[ARP-Mediation] H. Shah et al., "ARP Mediation for IP interworking
in Layer 2 VPN," <draft-ietf-l2vpn-arp-mediation-14>, July 2010.
[IPLS] H.Shah et al., "IP-only LAN service,"
<draft-ietf-l2vpn-ipls-09>, February 2010.
[PROXY-ARP] J. Postel, "Multi-LAN Address Resolution," RFC 925,
October 1984.
[TRILL] R. Perlman et al., "RBridges: Base Protocol Specification",
<draft-ietf-trill-rbridge-protocol-16>, March 2010.
8.0 Author's Address
Himanshu Shah
Ciena Corp
Email: hshah@ciena.com
Anoop Ghanwani
Shah, et al. Expires May 2011 10
Internet Draft draft-shah-arp-reduction-01.txt
Brocade
Email: anoop@brocade.com
Nabil Bitar
Verizon
Email: nabil.n.bitar@verizon.com
Shah, et al. Expires May 2011 11