A Network Architecture

Providing Host Migration Transparency

Fumio Teraoka, Yasuhiko Yokote, and Mario Tokoro

Sony Computer Science Laboratory Inc.

3-14-13 Higashigotanda, Shinagawa-ku,
Tokyo 141, Japan.

{tera,ykt ,mario}@csl. sony. co. jp

Abstract puters will increase the desire to carry one’s own com-
puter and experience the same computational environ-
The continued expansion of computer networks and
ment regardless of their location within the large inter-
the miniaturization of computers increase the desire to
connected network.
use one’s computer in a consistent computational envi-
In such an environment, when a user moves, he un-
ronment, regardless of location and time. In this sit-
plugs the computer from the local area network, trans-
uation, host migration transparency is very important.
ports it, and plugs it in to the local area network at
This paper introduces a new network architecture which
the destination. If radio networks, such as the digital
provides host migration transparency in large intercon-
cellular phone system, become widely available as data
nected networks and proposes a protocol based on the
links, it will be possible to communicate with other com-
propagating cache method. We introduce the concept
puters even while moving. One of the most important
virtual network and divide the conventional network
layer into two sublayers to realize host migration trans- features of such an environment is that it provides users
with the same computational environment no matter
parency. Virtual Internet Protocol, or VIP for short,
is derived from DARPA-IP as an example of the ar- where a user plugs in his computer. Even if a computer
is small enough to carry, a user will not carry it if it
chitecture. We estimate the overhead of VIP. Host mi-
requires reconfiguration or significant recompilation to
gration transparency can be realized without significant
achieve the same computational environment after con-
processing and traffic overhead. This architecture also
nection to a different network. From the user’s point of
preserves host migration transparency across virtual cir-
cuits in radio networks. view, host migration transparency allows the same com-
putational environment on the computer regardless of
where he migrates with his computer.
1 Introduction
There are telephone systems for moving entities, such
In recent years, computers have become smaller and as automobiles, vessels, trains, and planes. These sys-
more powerful. Although current small computers have tems all rely on a similar mechanism, i.e. the radio
limited functionality and poor communication facilities, cellular mechanism, For example, in an automobile cel-
in the near future, workstations functionally equiva- lular telephone system, the range of the radio cell is 3-5
lent to current desk top machines will be small enough km in an urban area or 10–15 km in a suburb. The mo-
to carry. Also, computer interconnection will be so bile station registers its location with the home memory
widespread that the various networks will constitute a station so that other stations can call the mobile sta-
world-wide distributed computing environment. Users tion by accessing the home memory station, Channel
will be able to use computers and networks from any lo- switching during a call is supported in all systems ex-
cation: campus, office, or home. The continued growth cept that of the vessel. The services provided in these
of computer networks and the miniaturization of com- systems correspond to those of the data-link layer in the
Permission to copy without fee all or part of this material is 0S1 seven-layer model. In order to use these systems as
granted provided that tho copies are not made or distributed for data links in large interconnected computer networks, a
direct commercial advantage, the ACM copyright notice and the
new network architecture, especially the network layer
title of the publication and its dete appaar, and notice is given
that copying is by permission of the Association for Computing architecture, is necessary.
Machinery. To copy otherwise, or to republieh, requires a fee Neither the 0S1 network architecture [1S0 84] nor the
and/or specific permission. existing protocols such as those in the DARPA protocol
@ 1991 ACM 0-89791-444-9191 /0008 /0209 . ..$1 .50

209
suite, take host migration into consideration. Although not thought of as host migration. We classify host mi-
several distributed systems, such as Mach [Accetta 86], gration into two types: ofl-line and on-line. In off-line
V-system [Cheriton 88], and Chorus [Rozier 88], pro- migration, the host is unplugged from the network be-
vide network transparency, these systems also do not fore moving and can be used as a standalone machine
support host migration.l In these distributed systems, during transport. On-line migration, in contrast, allows
users must consider the locations of roaming hosts when a host access to the network communication channels
communicating with them. Host migration can be de- and reliable connection-oriented mode communication
fined as changing the network to which a host is con- even during actual movement. In conventional wired
nected. Host migration results in a change in the host networks, only off-line migration is possible due to phys-
address. From the system’s viewpoint, host migration ical restrictions. However, if radio networks, such as the
transparency can be defined as the service that provides digital cellular phone system, become widely available
migration independent host identifiers. There are sev- as data links, on-line migration will also be possible. In
eral possible methods to trace migrating hosts, e.g. the this paper, hosts which can migrate on-line are called
name server method and the broadcast method; how- mobile hosts and hosts which can only migrate off-line
ever, these methods have some problems. According to are called portable hosts.
the network architecture based on the 0S1 seven-layer For the user, host migration transparency means that
model, the network layer is responsible for being aware his machine offers the same computational environment
of host locations within the interconnected networks; regardless of where he chooses to connect his machine
thus it is this layer which should provide host migration to the networks. For example, whether a user plugs his
transparency. computer in to a network in Tokyo or in New York,
This paper introduces a new network architecture he should be able to receive his e-mail on his com-
providing host migration transparency and gives an puter. Such services are offered to the user by appli-
overview of a protocol based on this architecture. Sec- cation entities, which must communicate with other ap-
tion 2 discusses what host migration transparency is and plication entities to continue their computation. If a
which layer should provide it. In Section 3, we intro- host migrates, application entities on the host also mi-
duce the concept of a virtual network to implement host grate.2 If the network can conceal this migration, users
migration transparency and divide the network layer can communicate with each other without any modifi-
into two sublayers to efficiently realize the virtual net- cations. From a program’s viewpoint, host migration
work. In Section 4, we propose a virtual network pro- transparency means that application entities are not
tocol based on the propagating cache method. Section conscious of host migration, and can communicate with-
5 gives an overview of VIP derived from DARPA-IP, as out modifications.
an example virtual network protocol. In Section 6, we
compare methods for providing host migration trans- 2.2 Who Should Provide Host Migra-
parency and estimate the overhead of VIP. Section 7 tion Transparency ?
describes the transport layer protocol which preserves
There are several methods to keep track of migrat-
virtual circuits for moving hosts. Section 8 concludes
ing hosts. For users, name servers provide a kind of
this paper.
migration transparency: an application entity can get
the new address of a migrating peer entity through a
2 Host Migration Transparency query to the name server, after which it can commu-
nicate with that peer. However, this method has sev-
2.1 What is Host Migration Trans-
eral problems. First, it is impossible to provide on-line
parency ? migration transparency without modifying application
It can be assumed that a large interconnected net- programs. If a host with a radio data link roams into
work consists of many connected subnetworks, or local another radio network, the address of the host changes.
area networks (LANs) and that a host is connected to The peer application entity needs to know that the host,
one subnetwork at a time. In such an environment, host has moved and must get the new address of the roam-
migration can be defined as the migration of a host from ing host to preserve communication channels. Second,
one subnetwork to another. Host migration results in this method increases network traffic. It requires com-
a change of the network address of the host. Because munication both between name servers and application
unplugging a host from a subnetwork (or an area in entities for queries and possibly among name servers
the 0S1 IS-IS routing model) and plugging it in to an- themselves to update the local data caches. Third, it is
other point in the same subnetwork does not change not clear when to update local caches at each host.
the network address of the host, this kind of move is
2proce~s/object migration, i.e. the migration of a pro-
1Amoeba supports host migration only within subnetworks cess/object from one machine to another, is beyond the scope
reachable by a broadcast packet [Tanenbaum 91]. of this paper.

210
 A broadcast mechanism offers users another kind of by T-1 through T-3. On-line migration transparency,
host migration transparency. For example, a host which required by mobile hosts, requires all four services.
wants to send a packet to a migrating host broadcasts a To support these services in the transport layer, the
query packet. The migrating host receives the query and network layer must offer the following six services in ad-
returns its current address to the sender. The sender dition to the normal services. We assume connectionless
can then send the packet to the migrating host. One mode network services.
of the most serious defects of this method is that it
cannot be used in large interconnected networks, where
@ (N-1) Migration transparent host addressing pro-
vides the upper layer with a migration independent
broadcasting causes a prohibitive increase in traffic.
identifier of the host. The transport layer can spec-
In both methods, application entities must be aware
ify the target host by the same host identifier re-
of the host location. Even if the library routines can
gardless of the host’s location.
conceal the host location, the application layer must still
trace the migrating hosts. That is, application entities e (N-2) Migration transparent connectionless mode
must be aware of changes of target location for every communication: offers the network level connec-
transmission and get the new address of the target if it tionless mode communication regardless of migra-
haa moved. tion.
According to the network architecture of the 0S1
seven-layer model, host-to-host communication is a net- ● (N-3) Disconnect request: prepares to leave a sub-
work layer service, while end-to-end communication is network when a host is to be unplugged.
the responsibility of the transport layer. In other words,
● (N-4) Disconnect indication: called by the lower
the transport layer and upper layers do not perceive the
layer to prepare to leave the current radio network
notion of hosts and it is up to the network layer to be
and enter another in on-line migration.
aware of host location in the interconnected networks.
In conformance with that architecture, the concealment ● (N-5) Connect request: prepares to join a network
of host location should occur in the network layer. If the when a host is plugged in.
network layer provides the upper layers with host mi-
gration transparency, application entities do not need ● (N-6) Connect indication: called by the lower layer
to be modified and on-line migration is possible. to prepare to join a new radio network in on-line
migration.
2.3 What are the Key Services ?
T-1 is supported by N-1. T-2 and T-3 are supported by
End-to-end services are provided by the transport and
N-1, N-2, N-3, and N-5. T-4 is supported by N-2, N-4,
lower layers. To offer host migration transparency in
and, N-6.
end-to-end services, the transport layer must provide
In the network layer, N-3, N-4, N-5, and N-6 are
the following four services in addition to the normal
supported by the lower layer. Therefore, N-1 and N-
services.
2, migration transparent host addressing and migration
● (T-1) Migration transparent transport entity ad- transparent connectionless mode communication, are pe-
dressing: provides the upper layer with the migra- culiar to the network layer and can be thought of as key
tion independent identifier of the transport entity. services. The next section introduces a new network
architecture to provide these two services.
● (T-2) Migration transparent connectionless mode
communication: offers the transport level connec- 3 Virtual Network
tionless mode communication regardless of migra-
tion. 3.1 Concept
In the existing network architectures, host addresses
● (T-3) Migration transparent connection establish- have a hierarchical structure, for example:
ment: establishes a transport level connection re-
gardless of migration. (a) area address + ID + SEL
(b) network number + host number
● (T-4) Migration transparent connection-oriented
mode communication: preserves the transport level In 0S1 [1S0 90], the network layer address has format
connection during actual movement and offers the (a), where area address indicates the routing domain, ID
transport level connection-oriented mode commu- is a unique number in the area, and SEL is a selector

nication. for the transport entity which is to receive the packet.

In XNS [Xerox 81] and DARPA-IP [SRI 81a], the host
Off-1ine migration transparency, which must be pro- address has format (b). In XNS, the host number is
vided to both portable and mobile hosts, is provided unique across the system and the network number is

211
added to facilitate the routing. In DARPA-IP, the net- cal networks. A host is assigned a virtual network
work number is unique across the system and the host address (VN-address), which this layer converts to
number is unique in the network. In general, it can be the corresponding physical network address (PN-
assumed that a host address has the two level hierarchy address).
as in (b). In networks employing such host addresses,
if a host migrates to another network, the host address ● The physical network subiayer (PN-sublayer) is the
changes. Since application entities must be conscious conventional network layer. According to the 0S1
of changes in the host address, host migration trans- model, this sublayer consists of the subnetwork
parency cannot be offered under present conditions. independent sublayer, the subnetwork dependent
We introduce the concept of a virtual network to en- sublayer, and the subnetwork access sublayer. This
able host migration transparency in the network layer. sublayer is responsible for packet relaying, routing
Virtual networks are logical networks which exist virtu- information exchange, and so on. A host is assigned
ally on the actual, or physical, networks. Virtual net- a physical network address (PN-address).
works logically construct interconnected networks above
the physical networks. Only the virtual networks are According to this separation, the VN-sublayer header
visible from the transport layer. Each host is connected is added between the (physical) network layer and trans-
to a virtual network just as it is connected to a physical port layer headers. If the (physical) network layer
network. A host never migrates in the virtual networks header of a incoming packet specifies the VN-sublayer
even if it migrates in the physical networks. That is, as the next upper layer, the packet is passed to the
each host is always connected to the same virtual net- VN-sublayer. Otherwise, the packet bypasses the VN-
work, called the native network of the host. A host has sublayer and is passed to the transport layer. Thus, the
a virtual network address ( VN-address) in the virtual VN-sublayer can be bypassed if a incoming packet does
network and a network address, which can be called the not have the VN-sublayer header or if the receiving host
physical network address (PN-address), in the physical does not have the VN-sublayer.
network. Since a host never migrates in the virtual net-
works, its VN-address never changes. The PN-address
of a host indicates the location of the host in the physi-
cal networks and is used for packet routing. The trans-
port layer specifies the target host by its VN-address
Transport Layer
I Transport Layer

no matter where it migrates in the physical networks.

Basically, the PN-address is invisible from the transport
layer. The VN-address and PN-address have the same
format.
The relationship between the virtual network address
and the physical network address of a host is analogous Physical Layer I I Physical Layer
to that between the virtual address and the physical
address in the virtual memory system. In the virtual
memory system, the virtual address used in a program Figure 1: Protocol Layers
never changes no matter where the program is located
in physical memory. Similarly, a host can be indicated
by its virtual network address no matter where the host
migrates in the physical networks. 4 The General Virtual Network
Protocol
3.2 Protocol Layering
Since the transport layer specifies the target host by This section gives an overview of the general pro-
its VN-address, the network layer must convert the VN- tocols of the virtual and physical network sublayers

address of a host to the corresponding PN-address in which do not depend on specific network layer proto-
order to deliver packets. The network layer is respon- cols. In [Teraoka 91a], the general protocol is defined
sible for host-to-host communication, including packet using Pascal-like language and a state transition dia-
relaying, routing control, and so on. Since this address gram. One of the main services in the VN-sublayer is
conversion is not in the traditional category of basic net- address conversion aa described in Section 3.2. We make
work layer functions, we divide the network layer into the following assumptions about the PN-sublayer: it
two sublayers as follows (see Figure 1): is the conventional connectionless mode network layer,
routing control is executed in the PN-sublayer using the
● The virtual network sublayer ( VN-sublayer) con- PN-address of a host, and conventional error control oc-
teals the physical location of hosts in the physi- curs in the PN-sublayer,

212
4.1 The Propagating Cache Method ● Source address timestamp (AMT,SrCTS): used to
There are several possible methods for address con- determine whether thk entry is obsolete.
version, such as conversion server method and broad-
● Idle time (AMTidze): used to delete unused entries.
cast method. In the conversion server method, the
VN-sublayer of the sender host converts the target VN- Since the network layer is divided into two sublayers,
address before transmission using a query to the conver- the network layer header includes both the VN-sublayer
sion server. In the broadcast method, the VN-sublayer header and the PN-sublayer header. The mandatory
of the sender broadcasts the target VN-address. The fields of the VN-sublayer header are as follows:
target host receives this broadcast packet and replies
with its PN-address. These methods, however, have the ● Source and destination virtual network address
same defects described in Section 2.2. (VNs,A~~, and VNLWAW,).
To make the overhead of address conversion as small
as possible, we introduce the propagating cache method.
● Packet type (VNtYPe). This field specifies whether
the packet is a data packet or a control packet.
This method falls under the category of lazy evalua-
tion. In this method, each gateway/host (or IS/ES in ● Source address timestamp (VN.S,CTS). The source
0S1 terms) holds a cache for address conversion. If the host sets the transmission start time in this field.
sender host does not know the destination PN-address,
The value of this field is copied into the A.MTSTCTS
it assumes that the target host is physically connected
field of the AMT entry for the source host.
to its native network, i.e. the sender assumes that the
target’s PN-address is equal to its VN-address. The ● Destination address timestamp (VNDesmS). This
packet transmitted then travels to the native network field is filled or modified when the destination VN-
of the target. If a gateway on the path holds a cache for address is converted to its PN-address. If the value
the target host, the address conversion is done at that of this field is older than that of the AMT.sVCTs
gateway. Then, the packet is forwarded to the actual field of the AMT entry for the destination, the PN-
physical location of the target. address field in the PN-sublayer header is overwrit-
Basically, the relationship between the VN-address ten with the value held in the AMT entry.
and the PN-address of a host is maintained by the hosts
or gateways in its native network. In the worst case, the In order to determine whether the AMT entry for the
packet sent to the migrating host is forwarded by its na- destination is obsolete, the virtual network protocol
tive network. Many packets, however, are forwarded by compares two timestamp fields: the VNDe*tTS field of

gateways before they reach their native network, since the packet and the AMT,S.CTS field of the local cache

cache information propagates to other gateways aa com- entry. This does not imply that the clocks of the hosts
munication progresses (see Section 4.2). In addition, the in the system must be synchronized because these two
sender learns the PN-address of the migrating host when timestamps originate from the same host. If the clock
it receives a reply from that host. Then, it can send of each host in the system increases monotonously, this

packets directly to the migrating host. The propagation obsoleteness check works correctly.
of cache information is due to two causes: transmission Transmission, reception, and relay procedures are as
of control packets of the VN-sublayer and transmission follows. If the source host has the AMT entry for the
of normal data packets. Control packets are transmit- target, it sets the destination PN-address field of the
ted when a host is connected to or disconnected from packet the value held in the AMT entry. Otherwise,
a network. Since the virtual and physical network ad- the source assumes that the target PN-address is equal
dresses of the sender are included in the headers of both to its VN-address. The transmitted packet travels by
packet types, gateways snoop the header and update lo- hopping from gateway to gateway. At each gateway,
cal caches when relaying packets. the AMT entry for the source host is created/modified

The cache held in each distinct gateway or host is if necessary, the destination PN-address in the packet
called an address mapping table (AMT). When a gate- header is modified, if necessary, according to the AMT
way or host receives a packet whose source PN-address entry for the destination, and the gateway relays the
is different from its source VN-address, the AMT entry packet. The destination host creates/updates the AMT
for the source host is created/updated. An AMT entry entry for the source if necessary.
consists of the following four fields:
4.2 Host Migration
● virtual network address (d&fT~nAdd~ ): acts as the Figure 2 depicts packet flow at the time of host mi-
key to the entry. gration. In the figure, an ellipse is a subnetwork, a
small circle is a host, and a rectangle is a gateway. The
● Physical network address (AMTpnAd&): the re- figure shows Host-X migrating from Net-A to Net-G.
quested value. Assume that the native network of Host-X is Net-A.

213
In this migration, the PN-address of Host-X changes
while its VN-address remains constant. After migra-
tion,3 Host-X sends a ConnectionNotification packet to
its native network, Net-A. This packet includes its VN-
address and its new PN-address. It travels via Net-G,
Net-C, and Net-B. Thus, Gw-CG, Gw-BC, and Gw-AB
learn the new PN-address of Host-X. The packet finally
reaches Net-A and then is broadcast within Net-A to Gw-AH &
+ Host-Y
notify hosts in Net-A of the new PN-address of Host-
X. Gw-AB
Host-X.
represents
returns
In the figure,
a gateway
an acknowledgment
a rectangle
which
marked
(ack) packet to
with
knows the new PN-address
a cross G A
. . . .
et-
Host-Z

Figure 3: Packet forwarding

of Host-x.

connectionNotifica tion packet in Net-G. This packet in-

cludes the VN-address of Host-X. When Gw-CG re-
.:. Host-X ~ Gw-BF & Gw-CG ceives this packet, it deletes the AMT entry for Host-
I I
X and broadcasts the packet in Net-C because Gw-
CG had an AMT entry for Host-X. Similarly, Gw-BC
broadcasts the packet in Net-B. Gw-B.F, however, does
Gw-AH 6 nothing when it receives the packet because it does not
* Host-Y have an AMT entry for Host-X. Thus, most AMT en-

0’ Net-H Host-Z
tries

related
for Host-X
be unplugged.
networks.
This
are deleted
control
The state
when
packet
transition
Host-X is about
does not travel
diagram
to
un-
of the
Figure 2: Plug-in to a network VN-sublayer is shown in Appendix A.

m“ost-x
The packet forwarding procedure is depicted in Fig-
ure 3. In the figure, Host-Y knows the VN-address of
Host-X but does not know its new PN-address. There- -
fore, Host-Y sets the PN-address field of the packet ~:~Host-X d Gw-BF & Gw-CG
equal to the VN-address of Host-X. The pack~t heads
for Net-A since Net-A is the native network of Host-X.
When the packet reaches Gw-BC, Gw-BC converts the
VN-address of Host-X to the corresponding
Then, the packet is forwarded to Host-X.
PN-address.
Similarly, if L Gw-AH b
Host-Y

Host-Z does not know the new PN-address of Host-X,

the packet reaches Net-A and then is forwarded to Host- W“os’-z
X. In the figure, a thin dotted line shows the path of a
packet with an incorrect destination PN-address, and a Figure 4: Unplug from a network
thick dotted line shows the path of a packet with a cor-
rect destination PN-address. If Host-X sends a packet to
Host- Y, Gw- CD and Host-Y both learn the PN-address 4.3 Obsolete Cache Entries
of Host-X. Thus, as communication between Host-X When a host is about to be unplugged from a network,
and hosts in other networks progresses, the cache in- the DisconnectionNoti fication packet is sent and most
formation for Host-X propagates more widely. Both the AMT entries for the host are deleted. Some obsolete
AMT entry and the VN-sublayer header have times- cache entries, however, might remain. A gateway or
tamp fields, so obsolete AMT entries do not convert host which finds a packet with an obsolete PN-address
a destination VN-address to an obsolete PN-address. returns an ObsoleteCache control packet to the source.
This prevents routing loops. As the ObsoleteCache packet travels, any gateway or
Figure 4 illustrates packet flow when a host is un- host along its path with a corresponding AMT entry
plugged from a network. In the figure, Host-X is about deletes that entry. Also, if an AMT entry at a gateway
to be unplugged from Net-G. Host-X broadcasts a Dis- or host is not accessed for a certain time, it is deleted
3 procedures ~XeCUt~dat the time a host is about to be Un- from that cache, with the exception of AMT entries in
plugged from a network are described below. the native network of the host, which are never deleted.

214
4.4 Tolerance against Network Parti- can communicate with the target regardless of the tar-
tion get’s migration. The 1P address, as the PN-address, is
invisible from the transport layer in the VIP environ-
In the worst case, a packet sent to a migrating host is
ment. Each host has a network to which it is normally
forwarded by its native network. The failure of a gate-
connected, the native network of the host. The VIP ad-
way which partitions the current location of the migrat-
dress is orignally set to the physical host address in the
ing host from its native network (e.g. Gw-AB in Figure
native network. Thus, orignally the 1P address and the
3), might cause a problem. There are two cases: Gw-AB
VIP address have the same value in the native network.
may either crash before or after Host-X is plugged in to
When a host is plugged in to a network other than its
Net-G. In the first case, since connection has already
native network, it requires a new 1P address. The proto-
completed, information for Host-X has already propa-
col for such 1P address allocation is not described in this
gated to Gw-CG and Gw-BC when the partition occurs.
paper. This allocation can be done with a reverse-ARP
In the second case, the ConnectionNotification packet
like protocol. That is, each local area network has an
transmitted by Host-X cannot reach its native network.
allocation server to which a migrating host broadcasts
However, some gateways, e.g. Gw-CG and Gw-BC, do
an address request. The allocation server then allocates
receive the packet. Therefore, hosts on the same side
an 1P address from the address pool.
of the network partition can communicate with the mi-
grating host.4 However, the native network cannot learn 5.2 Header Format
the new PN-address of the migrating host. This prob-
The VIP packet header proposed in this paper is im-
lem can be solved by having the migrating host continue
plemented as an option of the 1P packet header. There-
to retransmit the ConnectionNotification packet at cer-
fore, even if a gateway does not implement VIP, it can
tain intervals until an ack is received.
relay VIP packets. The VIP header consists of the fol-
lowing eight fields (see figure 6):
5 The Virtual 1P – An Example
● Option type (VIPOPt~vP.) and Option length
This section gives an example of the virtual network
(VIPOPt~~~). These fields are defined in 1P for op-
protocol. Virtual IP (VIP, for short) [Teraoka 91b] is
tions. The VIPoP~~vPe field is set to a value speci-
derived from DARPA-IP. The conventional 1P layer is
fying VIP. The VIP@tLe~ holds the length of the
divided into two sublayers: the VIP sublayer and the
VIP header including these two fields (= 20 bytes).
IP sublayer (see Figure 5). The functionality of the 1P
sublayer is the same as that of the conventional 1P layer. ● Protocol (VIPProto). This field indicates the upper
The VIP sublayer provides host migration transparency. layer protocol, i.e. TCP or UDP.

mM
● VIP type (VIPtvP~). This field indicates the type

3
RPC, ... FTP, . .. RPC, ... FTP, . .. of the VIP packet. There are five types:

— VipData A normal data packet.

UDP TCP UDP TCP
— VipDisc This packet announces that the
VIP ------ source is about to be unplugged from the net-
1P =)
-------
work.
1P
— VipConn: This packet announces that the
theme FDDI etc. source has just been plugged in to the net-
work.
— VipConnAck: This is an ack packet of the Vip-
Figure 5: VIP layering
Conn packet.
— VipErrObs This is an error notification
5.1 1P address and VIP address packet specifying that an AMT entry is ob-
As described in Section 3.1, a host has two addresses: solete.
a VN-address and a PN-address, The VIP address is
int reduced as the VN-address. The conventional 1P ad-
● Source VIP address (VIPsrc.4~d.) and Destination
dress acts as the PN-address. The format of both is the VIP address (VIPDe~tAddr).

same. If a host migrates, the 1P address changes while b Source address timestamp (VIP.S~.~s). This field
the VIP address is constant. The transport layer spec-
indicates the transmission start time at the source
ifies the target host using the VIP address so that it host. When an AMT entry for the source host
41n the Inter net, routing cent rol protocols such as EGp must is updated, this value is copied to the cache en-
be modified. try’s AMTsVCTs field. The AMT is updated if the

215
 value in this field is newer than the value in the 6 Estimation
AMTsVCTs field of the AMT entry.
This section compares the availability, traffic and pro-
cessing overhead, and tolerance against network par-
tition of the name server, broadcast, and propagating
● Destination address timestamp
cache methods. This section also quantitatively esti-
(VIPDe.t~s). When the destination VIP address
mates the overhead of VIP using the propagating cache
is converted to its 1P address, this field is copied
method. The overheads of VIP compared to 1P is neg-
from the value in the AMT,SVCTS field of the AMT
ligible.
entry for the destination host. Address conversion
occurs only if the value in this field is older than
6.1 Comparison of the Three Methods
the value in the AiWTSvCTS field of an AMT entry
for the destination. Table 1 shows the availability of the three methods.
As described in Section 2.2, the name server and broad-
cast methods cannot offer on-line host migration. The
In 1P, when a packet is sent to a subnetwork, the broadcast method cannot be used in large intercon-
packet might be split into small fragments if its size nected networks. The propagating cache method can
exceeds the maximum transmission length of the sub- offer off-line and on-line host migration transparency
network. These fragments are reassembled at the des- and it can be used in large interconnected networks.
tination host. This processing is called fragmentation
and reassembly. When fragmentation occurs, the VIP
header must be copied aa an option into the 1P header
Table 1: Method availability
for each fragment. Although the source and destination
1P addresses and the ID field are used to identify frag-
ments in 1P, the source and destination VIP addresses
are used instead of 1P addresses in VIP.

1P defines eight options, all of which are consistent

~
with VIP, except for Loose/Strict Source Routing. Since
the notion of source routing is incompatible with that
of migration transparency, these options should not be Table 2 shows the traffic and processing overhead of
used. These two options can be used only for fixed hosts the three methods. The traffic overhead of the name
whose 1P addresses are always equal to their VIP ad- server method is due to the registration/modification of
dresses. the entry and query/reply of the destination address. In
Eleven message types are defined in the ICMP addition, if the name server holds the local cache, the
[SRI 81b] to report control messages to the source host cache consistency control among name servers causes
of an 1P datagram. When a gateway forwards a packet, traffic overhead. The traffic overhead of the broadcast
if the outgoing subnetwork and the incoming subnet- method is due to the query/reply of the destination ad-
work are the same, an ICMP redirect packet is sent to dress at packet transmission time. In these two meth-
the source host. In contrast, a VIP gateway performing ods, it is possible that queries and replies take place at
address conversion might send a packet to the same sub- every transmission. The traffic overhead of the propa-
network from which it was received. An ICMP redirect gating cache method is due to the three types of con-
must not be sent in such case since such forwarding is trol packets sent at disconnection and connection time.
normal. This method does not cause traffic overhead at packet
transmission time. This method does cause processing
overhead at gateways and the destination host while the
other two methods do not, but this overhead is negligi-
ble as described in Section 6.2.
The broadcast method is omitted from our estima-
tion of tolerance against network partition since it can-
not be used in large interconnected networks. In the
destination VIP address name server method, the name server maintains the re-
source address timestamD lationship between the host name and its address. It
can be assumed that the host name has a hierarchical
I destination address timestamD I
structure as in the Domain Name System [SRI 87]. In
such a naming system, the host name and address are
Figure 6: VIP header format
maintained by the name server in its naming domain. If
a network partition occurs between the current location

216
 Table 2: Comparison of overhead
traflic at traflic at traflic at processing processing processing
plug-in time unplug time packet xmit at source at gateway at destination
name server registration deletion of query & reply query & reply – —
to name server the entrv.
broadcast — — query & reply query & reply — —
propagating 2 control 1 control — AMT entry AMT entry AMT entry
cache packets packet search search search

of the migrating host and its naming domain, the host is the time taken for the generation of the next key
cannot register its new address with the name server. if collision occurs. The mean number of comparisons,
Thus, the name server method is not tolerant against Ncompare ~ is M follows:
network partition. The propagating cache method is
tolerant against network partition as described in Sec- N ~ompa,e = @#
tion 4.4.
where, a is the load factor, the fraction of the hash table
6.2 Overhead Estimation of VIP which is filled. Using linear probing for collision resolu-
In VIP using the propagating cache method, the over- tion, then a values of 0.25, 0.5, and 0.9 yield IVcornpar.
head of packet transmission from source host to desti- values of 1.17, 1.50, and 5.50, respectively. We Assume
nation host in the stable state, To.$ata, can be shown as that the host address addr, the hash function ll(addT),
follows: and the function generating the next key hk+l from hk
are as follows:
To-data = TSTC+ T~~ x Niwp + l’de,t

addT = A1,A2.A3.A4 (1P’s dotted notation)

where, T8~c, Tgw, and Tde~t are the processing overhead
H(addT) = Al + A2 + A3 + A4 (mod S)
at the source host, the gateway, and the destination
hk+l = hk + 1 (mod S)
host, respectively, and NhOP is the number of hops be-
tween the source and the destination. T8.C, TgW, and where S is the number of the entries in the AMT.
Tde,t can be defined as follows: If the MC68030 CPU with 25MHz clock is used, a is
0.5, and the hop count between the source and the des-
T..c = T.mt-...rch + Th-const
tination is 20, then the data packet transmission over-
Tgw = Tamt-...r.h x 2 + Tarot-update
head To-data is 0.3-0.4 msec.5 The order of the trans-
T~e.t = Tamt_.ea..h i- Tamt-uP~ate i- Th-check
mission time in the Internet, e.g. between Tokyo and
where, Tamt_.....h,
Th-..t,t,
ff’amt.update, and Th.check
the East Coast of the U. S.A., is 100 msec. Therefore,
are the time taken for the AMT entry search, the con- the transmission overhead in the stable state, ToAata, is
struction of the VIP header, the update of the AMT negligible.
entry, and the check of the VIP header, respectively. At The memory usage overhead for VIP compared to
the source host, AMT entries are searched for the ad- conventional 1P is due to the area for the AMT and the
dress conversion of the destination address. At the gate- extra area for the VIP header in the packet buffer. The
way, they are searched for the address conversion of the size of an AMT entry is 16 bytes. If a thousand AMT
destination address and for the possible update of the entries are prepared, they occupy only 16 Kbytes. The
AMT entry for the source host. At the destination host, size of the VIP header is 20 bytes. If a hundred packet
they are searched for the possible update of the AMT buffers are prepared, the memory usage overhead is only
entry for the source host. Th-COn~t and Tamt.UPd~t~ con- 2 Kbytes. Therefore, the extra memory usage of VIP is
sist of several valuable assignments. Th-check consists of negligible.
several valuable comparisons. The traffic overhead of VIP is due to control packets.
A hash function can optimize the AMT entry search There are three kinds of control packets in the VIP sub-
time Tamt-.....h, Wdhg Iayer besides the error packet: VipConn, VipConnAck,
and VipDisc. The VipConn and VipConnAck packets
Tamt.sea.ch = T’hash + Tcompa.e x Ncompa.e +
are transmitted once when a host is plugged in to a
Tkey x (~compa.e – 1) network. These packets travel between the current lo-
where, Tha~h is the time taken to calculate the lookup cation of the migrating host and its native network, The

key, Tc.mpa.e is the time taken for the key comparison, bli’or ~Xample, the hop count between SonYCSL in Tokyo and
Ncompare is the m-n number of comparisons, and Tkev MIT is about 20.

217
VipConn packet is broadcast within the native network. 1. When a host enters the common area of adjacent
Only two control packets are transmitted when a host cells, the data-link layer notifies the network layer.
is plugged in to a network. Therefore, the traffic over- The network layer then notifies the transport layer.
head at plug-in time is negligible. The VipDisc packet The transport layer changes the state of each con-
is broadcast within the network from which a host is nection from Active to Suspend.
about to be unplugged. The gateways with AMT en-
tries for the host relay that packet. In a broadcast type 2. The transport layer sends the TconnSuspend
network such as Ethernet, the cost of broadcast is equal packet to the peer entity of each transport con-
to that of unicast. The VipDisc packet is propagated nection and waits for TconnSuspendAck packets.
only to those networks which were on the communica- After all acks are received, the transport layer is-
tion route to/from the migrating host. The VipDisc sues the migration request to the network layer.
packet is not propagated to unrelated networks. Thus,
the traffic overhead at disconnect time is negligible. 3. The network layer processes the migration proce-
dures described in Section 4.2.

7 Mobile Host Support

4. The transport layer sends the TconnResume packet
In radio networks, off-line migration transparency can to the peer entity of each transport connection and
be achieved in the network layer by the method de- waits for TconnResumeAck packets. After all acks
scribed in Section 4. To achieve on-line migration trans- are received, the transport layer changes the state
parency, the transport layer must preserve the estab- of each connection from Suspend to Active.
lished connections while moving. This section describes
the transport layer procedures which preserve the trans- The state transition diagram of the transport connec-
port connection in on-line migration. tion is shown in Appendix B.
It can be assumed that radio networks consist of many
radio cells, each of which can be thought of as equivalent
to a conventional (wired) subnetwork as shown in Figure
8 Conclusion
7. Each cell is assigned a network number. In the figure, Telephone systems for moving entities such as auto-
a dotted circle is a radio network and the small rectangle mobiles are already available. Also, the construction
at the center of each radio network is a base station or of local area radio network is underway. In these sys-
gateway. These gateways are connected by wire. A tems, however, the usage of radio is limited within the
small circle is a host migrating in the radio networks. subnetwork. That is, the corresponding protocol tech-
We can also imagine the dotted circles in the figure as niques are part of the data-link layer. The near fu-
a cellular phone system consisting of many cells. Such ture will bring the construction of large interconnected
a cellular phone system can be thought of aa equivalent radio networks. Because host migration transparency
to a conventional (wired) subnetwork. is paramount in such an environment, a new network
,--- architecture for radio networks which can handle mo-
bile/portable hosts will be essential.

1--*’ This paper proposes that host migration

should be provided
the 0S1 seven-layer
of a virtual network.
layer into two sublayers:
and the physical
identifiers:
network
a migration
in the network layer
model and introduces
This concept divides the network
the virtual network
sublayer.
independent
sublayer
transparency
according

Each host has two

identifier (the vir-
to
the concept

Figure 7: Host migration in radio networks tual network address) and a migration dependent iden-
tifier (the physical network address). To minimize the
In addition to the normal control packet types in overhead of address conversion, we int reduce the prop-
the transport layer, there must be four extra types agating cache method. This method falls under the
of control packets to preserve transport connections: category of lazy evaluation. This paper also proposes
TconnSuspend, TconnSuspendAck, TconnResume, and VIP derived from DARPA-IP by applying the concept
TconnResumeAck. In on-line migration, the data-link of the virtual network. The overhead of VIP versus
layer is responsible for notifying the network layer when 1P is negligible. If radio networks become widely avail-
the host enters the common area between the current able as data links, a host migrating to another network
radio cell (or cellular phone system) and the adjacent can continue connection-oriented mode communication.
cell. The procedures to preserve the transport connec- VIP is being implemented on the Muse operating sys-
tions are as follows: tem [Yokote 91].

218
Acknowledgments [Xerox 81] Xerox. Internet Transport Protocols. XE-
ROX CORPORATION, December 1981. XSIS
We give our thanks to Dr. Hideyuki Tokuda of
028112.
Carnegie Mellon University, Dr. Jun Murai of Keio
University, the members of the WIDE project, and the [Yokote 91] Yasuhiko Yokote, Fumio Teraoka, Atsushi
members of Sony Computer Science Laboratory Inc. Mitsuzawa, Nobuhisa Fujinami, and Mario Tokoro,
Discussions with them helped us refine the concept of The Muse Object Architecture: A New Operating
virtual network and its adaptation to the 1P. System Structuring Concept. Operating Systems Re-
view, VO1.25, No.2, April 1991. also available as SCSL-
References TR-91-002 of Sony Computer Science Laboratory Inc.

[Accetta 86] Mike Accetta, Robert Baron, David

Golub, Richard Rashid, Avadis Tevanian, and Appendix A
Michael Young. Mach: A New Kernel Foundation The state transition diagram of the virtual network
For UNIX Development. Technical Report, Depart- sublayer is depicted in Figure 8. In the figure, S-n is
ment of Computer Science, Carnegie-Mellon Univer-
a state. ~ means the state is changed by the event
sity, August 1986.
E-n and t%-emoutput O-m is issued. The states, events,
[Cheriton 88] David R. Cheriton. THE V DIS- and outputs are as follows:
TRIBUTED SYSTEM. Communications of the
● States:
ACM, VO1.31, No.3, pp.314-333, March 1988.
– S-1: Disconnected disconnected from the net-
[1S0 84] 1S0. Information processing systems - Open
work.
systems Interconnection - Basic Reference Model.
– S-2: Wait VnConnReq waiting for the VN-
1984. 1S07498.
sublayer connect request.
[1S0 90] 1S0. Intermediate System to Intermediate Sys- – S-3: Wait VnConnDoneInk waiting for the
tem Intra-Domain Routeing Exchange Protocol for VN-sublayer connect done indication.
use in Conduction with the Protocol for Providing the
– S-4: Wait VnPeerConnAck: waiting for the
Connectionless-mode Network Service (ISO 8,/73’).
VN-sublayer connect notification ack from
February 1990. 1S0 DP 10589,
peer.
[Rozier 88] M. Rozier, V. Abrossimov, F. Armand, I. – S-5: Connected: connected to the network,
Boule, M. Gien, M. Guillemot, F. Herrmann, C.
– S-6: WaitVnDiscReq waiting for the VN-
Kaiser, S. Langlois, P. L60nard, and W. Neuhauser.
sublayer disconnect request.
Chorus Distributed Operating Systems. Computing
Systems, Vol.1, No.4, Fall 1988. – S-7: WaitVnDiscDoneInk waiting for the
VN-sublayer disconnect done indication,
[SRI 81a] J. B. Postel. Internet Protocol. September
1981. RFC 791. ● Events:

– E-1: VnConnReq the VN-sublayer connect

[SRI 81b] J. B. Postel. Internet Control Messag Proto-
col, September 1981. RFC 792 request is called by the transport layer.
– E-2: VnConnInd the VN-sublayer connect
[SRI 87] P. V. Mockapetris. Domain names - concepts indication is called by the PN-sublayer.
and facilities. November 1987. RFC 1034.
– E-3: VnConnDoneInd the VN-sublayer con-
[Tanenbaum 91] Andrew S, Tanenbaum. In his lecture nect done indication is called by the PN-
in Sony Computer Science Laboratory Inc. January sublayer.
1991. – E-4: VnDiscReq the VN-sublayer disconnect
request is called by the transport layer.
[Teraoka 91a] Fumio Teraoka. A GeneraZ Protocol Spec-
ification of the Virtual Network Sublayer. Techni- – E-5: VnDiscInd the VN-sublayer disconnect
cal Memo SCSL-TM-91-001, Sony Computer Science indication is called by the PN-sublayer.
Laboratory Inc., January 1991. – E-6: VnDiscDoneInd the VN-sublayer dis-
connect done indication is called by the PN-
[Teraoka 91b] Fumio Teraoka. The Virtual IP Proto-
sublayer.
col Specification. Technical Memo SCSL-TM-91-O07,
Sony Computer Science Laboratory Inc., February – E-7: VnPeerConnAck the VN-sublayer con-
1991. nect notification ack is received from peer.

219
 o outputs: — S-5: WaitTConnDoneInd waiting for the
TConnDoneInd to be called by the lower layer.
– O-1: TConnInd: calls the transport layer con-
. S-6: WaitTconnResumeAck: waiting for the
nect indication.
TconnResumeAck packet from the peer trans-
– O-2: TConnDoneInd calls the transport layer port entity.
connect done indication.
● Events:
– O-3: TDiscInd calls the transport layer dis-
connect indication. – E-1: TDiscInk the transport disconnect indi-
– O-4: TDiscDoneInd calls the transport layer cation is called by the VN-sublayer.
disconnect done indication. – E-2: TDiscDoneInd the transport disconnect
– o-5: VnPeerConn transmits the VN- done indication is called by the VN-sublayer.
sublayer connect notification to peer. – E-3: TConnInd: the transport connect indi-
– O-6: VnPeerdisc: transmits the VN-sublayer cation is called by the VN-sublayer.
disconnect notification to peer. – E-4: TConnDoneInd the transport connect
– O-7: PnConnReg calls the PN-sublayer con- done indication is called by the VN-sublayer.
nect request. – E-5: TconnSuspend the TconnSuspend
– O-8: PnDiscReg calls the PN-sublayer dis- packet is received.
connect request.
– E-6: TconnResume: the TconnResume packet
is received.

– E-7: TconnSuspendAck The TconnSus-

pendAck packet is received,

– E-8: TconnResumeAck the TconnRe-

sumeAck packet is received.

● outputs:

— O-1: TconnSuspend transmits the TconnSus-

pend packet.
Figure 8: State transition diagram of the virtual net-
— O-2: TconnResume: transmits the TconnRe-
work sublayer
sume packet.
— o-3: TconnSuspendAck transmits the
Appendix B TconnSuspendAck packet.
— O-4: TconnResumeAck transmits the Tcon-
The state transition diagram of the transport connec-
tion related to host migration is depicted in Figure 9. nResumeAck packet.

In the figure, S-n is a state. ~ means the state is — O-5: VnDiscReg calls the VN-sublayer dis-
changed by the event E-n and output O-m is issued. connect request.
The states, events, and outputs are as follows: — O-6: VnConnRey calls the VN-sublayer con-
● States: nect request.

– S-1: Connected the transport connection is

connected and available to transmit/receive
packets.

– S-2: WaitTconnSuspendAck waiting for the

TconnSuspendAck packet from the peer trans-
port entity.

– s-3: Wait TDiscDoneInk waiting for the

TDisconnDoneInd to be called by the lower
layer.
Figure 9: State transition diagram of the transport con-
– S-4: Suspended the transport connec- nection
tion is suspended and unavailable to trans-
mit/receive packets.

220