2.2 The Architecture
Mobiware is a software-intensive adaptive
mobile networking environment based on distributed object technology. As
illustrated in Figure 2 mobiware promotes the separation between mobile
signaling and adaptation management, and the transport of media. At the
transport layer, an active wireless transport supports the end-to-end transmission
of audio, video and real-time data services based on an adaptive-QOS paradigm.
The active wireless transport is an object-based transport that blurs the
region over which traditional transports (e.g., TCP and RTP) typically
operate to include access points and mobile devices. Built on a set of
Java classes, the transport system binds active and static transport objects
at mobile devices and access points to provide end-to-end transport adaptation
services. Static transport objects include segmentation and reassemble,
rate control, flow control, playout control, resource control and buffer
management objects. These objects are loaded into the mobile device as
part of the transport service creation process. Active transport objects
can be dynamically dispatched to mobile devices and access points to support
valued-added QOS. Currently, two styles of active transport objects have
been implemented: active media filters [Balachandran,97],
which perform temporal and spatial scaling for multi- resolution video
and audio flows and adaptive FEC filters [Mobiware,98]
which protect content against physical radio link impairments by matching
the level of Reed Solomon channel coding to match time-varying error characteristics.
At the network layer, a programmable mobile
network supports the introduction of new mobile adaptive-QOS services based
on the xbind broadband kernel [xbind,96].
The network layer supports switching IP flows over ATM native transport
services. Architecturally, the network comprises a set of CORBA network
objects and adaptation proxies that operate at the mobile device, access
points and at mobile capable switch/routers. Currently, an adaptive-QOS
network service supports the delivery of multi-resolution flows having
a base layer and one or more enhancement layers. The base layer provides
a foundation for better resolutions to be delivered through the reception
of enhancement layers based on the availability of resources in the wireless
environment. Three key mobiware network algorithms include: (i) QOS controlled
handoff, which gracefully scales flows (up and down) based on the semantics
of the adaptive-QOS service during handoff when bandwidth availability
varies; (ii) mobile soft-state, which provides mobile devices with the
capability to respond to changes in wireless-QOS and mobile-QOS; and (iii)
flow bundling, which exploits a common routing representation for all the
flows to and from a mobile device to speed up handoff. The focus of this
overview is the design and evaluation of the programmable mobile network layer
described in Sections 3 and 4, respectively.
Figure 2: The Mobiware Architecture
At the data link layer, a programmable
MAC [Bianchi,98b] combines a set of foundation
services to support more sophisticated adaptive wireless-QOS services.
Foundation services provide sustained rate services used to support minimum
wireless-QOS assurances, and active and passive adaptive services to support
application specific adaptation policy as discussed in Section 2.1. The
"programmable" nature of the data link service provisioning over wireless
networks provides an alternative approach to that found in the literature.
Rather than supporting a specified set of 'hard wired' MAC services (e.g.,
VBR) by means of centralized control schemes, it provides a programmable
air-interface that allows new services to be dynamically created and installed
on the fly. This programmable MAC service support relies on a simple core
architecture that pushes complexity and application specific adaptation
decision making to the mobile device. For full details on the programmable
MAC see [Bianchi,98b].
2.3 The Network Model
Mobiware provides a middleware toolkit
that controls mobinet an experimental programmable broadband cellular access
network. The network model [Footnote #2] comprises a set
of mobile devices, wireless access points and mobile- capable switches/routers
providing broadband cellular and ad hoc communication services to mobile
users. Mobinet is based on ATM switching technology that supports IP switched
flows in the access network. Mobile devices can be connected to mobinet
via broadband cellular or ad hoc wireless access modes. In broadband cellular
mode, mobile devices receive core network services via a set of wireless
access points. Ad hoc devices may operate autonomously without the aid
of any fixed infrastructure and core network services or can connect to
the broadband cellular network via multiple ad hoc hops as illustrated
in Figure 3.
Figure 3: Mobinet
Providing quality of service assurances
in broadband cellular networks is difficult. However, providing quality
of service assurances without the aid of any fixed infrastructure as in
the case of mobile ad hoc networking is more challenging [Corson,98].
We believe there is a need to understand the level of quality of service
that can be supported at different points of attachment in mobinet, e.g.,
at the access point or multiple hops away from the access point. We observe
that quality of service assurances are likely to diminish as a mobile device
moves away from the core network. Providing seamless quality of service
support to mobile devices on the move (e.g., switching between broadband
cellular and ad hoc modes) underpins mobiware's adaptive-QOS design approach
[Footnote #3] .
3. Programmable Mobile
Network
3.1 Programmable Objects
The mobile network comprises a set of programmable
distributed CORBA objects [Footnote #4] that support the
delivery of adaptive-QOS flows to mobile devices over mobinet. The use
of distributed object technology also provides support for interoperability
between mobile devices utilizing different operating systems and protocol
support. Mobiware objects execute on mobile devices, access points and
mobile capable switch/routers supporting a set of mobile signaling and
quality of service adaptation algorithms (viz. QOS controlled handoff,
flow bundling and mobile soft-state) as illustrated in Figure 3. Objects
combine data structure (defining the object's state) with a set of methods
(defining the object's behavior). Methods are executable programs associated
with objects that operate on information in an object's data structure.
Two per-mobile proxy objects support the
adaptation and handoff of flows in mobiware: (i) QOS Adaptation Proxy (QAP)
objects play an integral role in allowing mobile devices to probe and adapt
to changing resource availability over the wireless link; and (ii) Routing
Anchor Proxy (RAP) objects support the 'bundling' (i.e., aggregate) of
flows to and from mobile devices for fast, efficient and scalable handoff.
To manage the network states introduced
by flow- oriented communications and, more importantly, to gain efficiency
across a wireless link, mobiware deploys a number of network objects that
can execute on network nodes or on servers at the edge of the network.
In the following we outline the function of some of the mobiware objects
and their interface definitions. For full details on the objects and interfaces
see [Mobiware,98].
A mobile device object abstracts the operation
of mobile devices and provides APIs for querying beaconing information,
registering with new access points, establishing flows, renegotiating quality
of service and handing off flows. It also includes the functionality to
dynamically control the transport system at the access points, e.g., to
set the media scaling or error control level for video flows. The mobile
device object state mainly consists of quality of service specification
(viz. utility function and adaptation policy) for all the flows transported
to/from the mobile device and routing information including the source/destination
address and current RAP and access points addresses.
An access point object supports APIs for
binding to wireline network objects (e.g., mobility agent) on behalf of
mobile stations, propagating CORBA calls and for the establishment and
periodic refreshing of local wireless flow state as illustrated in Figure
4. This object plays an important role in QOS controlled handoff and interacts
with the transport system for the injection of active transport objects.
// initialize mobile related states during registration
void mobileRegister(in long fbi, in string<40> coreName,
in string<40> cmName,
in string<40> msName)
raises(Reject);
// flow setup from the current access point to the network, called by
// the mobile device object.
void setupFlow(in long fbi, inout QOSSpecification qosSpec,
inout FlowInfo flowinfo, in string<40> srcname,
inout EndPoint peerEp, inout EndPoint RAP_fix,
inout EndPoint RAP_mobile, inout EndPoint AP_fix,
inout EndPoint AP_mobile, inout EndPointId airIP,
inout double msr_time)
raises(Reject);
// handoff flow bundle for a specific mobile device
void handoffFlowBundle(in long fbi, inout QOSSpecList qosSpec[2],
inout FlowInfoList flowinfo[2],
inout SourceList srcname[2],
inout EndPointList RAP_fix[2],
inout EndPointList RAP_mobile[2],
inout EndPointList AP_fix[2],
inout EndPointList AP_mobile[2],
inout EndPointIDList airIP[2],
inout double msr_time)
raises(Reject);
// refresh mobile soft-state for a flow bundle from the current
// access point to the network.
void refreshFlowBundle(in long fbi, inout EndPointList RAP_mobile[2],
inout EndPointList AP_fix[2],
inout EndPointList AP_mobile[2],
inout double ntw_msr_time)
raises(Reject);
Figure 4: CORBA IDL for Access
Point Object
A mobility agent object provides flow,
adaptation and mobility management services. It interacts with per- mobile
RAP and QAP state in the switch servers and supports APIs for retrieving
network topology information from a router object (e.g., location of the
cross over switch) and for interacting with the switch servers (see below)
to establish, maintain and handoff flows in the cellular network.
A set of switch server objects [xbind,96]
abstract and represent physical ATM switch/routers and are quality of service
programmable. These objects support APIs for the reservation and release
of namespace (viz. VCI/VPI pairs) and the allocation of network resources
(viz. bandwidth). State mainly consists of per-flow connection information,
stored in local hash-tables called switch caches. Switch server objects
have been extended to be mobile capable, i.e., support RAP and QAP functionality.
The General Switch Management Protocol (GSMP [Newman,96])
is used at the access points and switch/routers for accessing the switch
tables.
3.2 QOS Controlled Handoff
QOS controlled handoff gracefully scales
flows during handoff based on the semantics of the adaptive-QOS API described
in Section 2.1. By scaling flows during periods of resource contention
(e.g., during handoff), mobiware improves the wireless resource utilization
and helps reduce the handoff dropping probability. While the style of handoff
is entirely programmable [Angin,98b], the current
implementation style is mobile-initiated, forward handoff with soft-handoff
on the down-link and hard- handoff on the uplink. By mobile-initiated,
we mean that after a suitable dwell time a mobile device initiates a handoff
by first registering with the forward/new access point. By soft-handoff,
we mean that during handoff the mobile device simultaneously receives flows
from the old and new access point on the down-link. In contrast, uplink
flows use a `break and make' approach between the old and new access points.
During handoff, registration to the new access point, rerouting of flows
and quality of service adaptation is accomplished by signaling objects
and associated APIs outlined in Section 3.1. Signaling APIs are programmable
[Footnote #5] allowing various styles of handoff to be
tailored toward particular radio environments.
The QOS controlled handoff object-interactions
are illustrated in Figure 5. Mobile device objects periodically `hunt'for
beacon signals from neighboring access points. Beacons are made programmable
by the mobiware toolkit and carry low-level signal information in addition
to the current bandwidth availability at the sending access point. The
mobile devices' hunting algorithm periodically compares all beacons received
over the current hunt period and cumulatively over multiple hunt periods.
If the wireless-QOS [Footnote #6] indicated in the beacon
from the current access point falls below a pre-determined threshold the
hunt algorithm selects a new access point for handoff. Handoff is initiated
after a suitable dwell period after which the mobile device registers with
the new access point starting the handoff process as indicated by (2) (3)
in Figure 5.
The device registration procedure triggers
the new access point object to bind to a mobility agent object (4). The
mobility agent caches bindings to the per-mobile adaptation proxies that
are implemented as part of switch servers. Mobility agents and proxies
can run anywhere in the mobinet, i.e., mobility agents can operate at fixed
edge devices, mobile devices or on the switches. Mobility agents are the
main controllers for managing handoff in mobiware. Mobility management
is a fully distributed algorithm that includes one or more mobility agents
for scalability. Currently, we allocate a single mobility agent to manage
handoff in the experimental mobinet. When the mobile device initiates a
handoff (5) it passes a unique mobile device identifier called the flow
bundle identifier (FBI) to the access point that allows mobiware to identify
the mobile device's flow bundle in the wireless access network.
Figure 5: QOS Controlled Handoff
Object Interaction
Mobility agents are responsible for re-routing
a mobile device's flow bundle from an old access point to a new one as
illustrated in Figure 5. This entails the mobility agent invoking the route
object (7) to determine the location of the cross over switch as illustrated
in Figure 5. Switch server objects are used to re-establish new flow state
at all switches between the cross over switch and the new access point.
The re-routing phase includes name space reservation (viz. outgoing VCI/VPI)
and bandwidth value at each network switch and the new access point. The
final process of re-routing a flow bundle through a switch includes the
use of GSMP (9) (9') (12) to set up the switch table and reserve resources.
However, GSMP does not support the concept of flow bundling. While the
mobile agent informs the switch server objects to establish state for a
complete flow bundle, the switch server interacts with GSMP on a flow by
flow basis. In the evaluation section we describe enhancements to GSMP
to support flow bundling. The mobility agents interact with mobile capable
switch servers and the new access point in parallel (8) (8') (11) resulting
in a speed up of the re-routing phase of the handoff algorithm over conventional
hop-by-hop signaling. After the re-routing of the flow bundle is complete
the mobile agent informs the new access point of the negotiated quality
of service and flow bundle VCI/VPI mappings (11). The new access point
interacts with the active wireless transport to provide active media filters
based on the available bandwidth at the air interface [Balachandran,97].
To keep the name space binding between
the mobile device and access points constant with mobility we have implemented
the notion of virtual wireless ports. As mobile devices connect to different
access points their VPI/VCIs mapping remain constant. The flow bundle to
VPI/VCI bundle is resolved by a virtual wireless port, which is dynamically
allocated by the new access point during handoff. This approach minimizes
the impact of re-negotiation in comparison to a full name space re- negotiation
which would be disruptive during handoff.
3.3 Flow Bundling
QOS controlled handoff and mobile soft-state
exploit flow bundling to speed up handoff and minimize the signaling overhead
associated with maintaining the network state. Flow bundling [Porter,95]
provides a common routing representation for all the flows to and from
a mobile device as illustrated in Figure 3. This is similar to the virtual
path concept in ATM networks or tunneling in IP networks. Flow bundling
is a general method for encapsulating and routing. Flow bundling is motivated
by the need to reduce the complexity of re- routing multiple independent
flows to and from mobile devices during handoff. By aggregating flows in
this manner we can speed up handoff, simplify mobile soft- state probing
and minimize signaling overhead. Using flow bundling, QOS controlled handoff
simply discovers a single crossover switch then re-routes all flows to
the new access point in a single object-level operation.
During handoff the flow bundle object interaction
is as follows. The mobile device object invokes the access point's HandoffFlowBundle()
method once for the flow bundle minimizing the signaling overhead at the
air- interface as illustrated in Figure 5. Mobiware supports the option
of enabling or disabling bundling when flows are established. Note that
when flow bundling is disabled the mobile device, access point, mobility
agent objects treat each flow independently during handoff. As discussed
in the evaluation this increases the signaling overhead and the handoff
latency. During each invocation a separate cross over switch needs to be
located, using a shortest path algorithm and individual flows need to be
re-routed and signaled independently.
The mobile agent interacts with the switch
servers to re-establish flows and update switch tables for all switches
between the cross over switch and the mobile device. GSMP is used to update
the switch table at each traversed switch. In order to support the atomic
re-routing of flow bundles we have enhanced the GSMP protocol. Flow aggregation
at the switch control level has been implemented as a modification to the
GSMP invocation mechanism. Two enhancements to GSMP to support flow bundling
are considered in Section 4. The first enhancement has no impact on the
current GSMP client server interaction. The mobile agent simply invokes
the GSMP client for each flow in a flow bundle without waiting for acknowledgement
from each GSMP command. This results in the switch server sending a burst
of GSMP setup messages to the switch then waiting for acknowledgements.
The second enhancement augments the GSMP setup message to allow up to 256
VCI pairs to be passed across the interface in one client- server interaction.
This allows the switch server to set up the switch table for a flow bundle
in one operation. We discuss the performance benefits of flow bundling
in Section 4.
3.4 Mobile Soft-state
During handoff a flow bundle must be re-routed
to a new access point, resources need to be reserved, and the old flow
bundle state between the old access point and the cross over switch removed.
Mobile devices resident in cells also need to scale flows in accordance
with channel conditions, whether new flows are established or released,
and when new mobile devices enter and leave cells. Mobile soft-state provides
quality of service adaptation support to cater to a number of these conditions.
Mobile soft-state results in the periodic establishment of bandwidth and
name space resources for flow bundles between a mobile device and a per-mobile
QAP. Mobiware supports the idea of soft-state [Clark,90]
in the mobinet to refresh the network state. The periodic refresh messages
sent by a mobile device as part of the soft-state probing mechanism are
sent on a per-flow bundle basis not a per-flow basis. This means a mobile
device sends a single probe message for all flows supported at the mobile
device rather than one probe per-flow. During the refresh phase mobile
devices respond to any changes in allocated bandwidth (based on utility
functions) to a flow bundle.
In [Campbell,97]
we argue that a soft-state approach is well suited to supporting QOS adaptation
in mobile networks. Mobile soft-state supports a distributed probing mechanism
based on flow bundles allowing mobile devices to compete fairly for bandwidth
in a completely decentralized and scalable manner. During handoff mobile
devices do not explicitly remove the old flow bundle-state between the
old access point and the cross over switch. In this case, mobile soft-state
timers located at the switches and old access point timeout and release
resources automatically. Mobile devices resident at the old access point
compete for these available resources thereby potentially improving their
utility.
Mobile devices periodically probe the path
between the mobile device and the per-mobile QAP and contend for resources.
Note that per-mobile QAPs can be located at an access point or any mobile
switch/router between the mobile and its corresponding RAP. If the QAP
is located at the access point then mobile soft-state is only active over
the air-interface; that is, between the mobile device and access point.
The position and configuration of where these proxies reside is programmable.
In many cases, the access point is the most suitable location because radio
resources are generally the bottleneck in broadband wireless or wireless
LAN systems.
Mobile devices independently probe the
wireless access network. The probe includes a list of flow requirements
for the complete flow bundle, which includes the utility function for each
flow. In the current system, we have implemented discretely-adaptive flow
support (see Figure 1) where a base layer (BL) and one or more enhancement
layers (viz. E1, E2) would be signaled in the probe message. The probe
interacts with the access point and all mobile capable switches/routers
between the mobile device and its QAP. The response is returned to the
mobile device indicating the available bandwidth reserved for each flow
in the bundle over the next time interval. This time interval is called
the refresh interval. If a probe message is received along the path before
the refresh interval expires then the reservation is `refreshed' based
on available resources. In contrast, if no refresh message is received
and the timer expires, then the resources are deallocated and states are
removed. This reservation-adaptation style of probing and response is implemented
at the object level as a set of refreshFlowBundle() method invocations
(2) (3) (3') as illustrated in Figure 6. The mobile device issues a refresh
to the mobility agent object which then refreshes all switches and the
current access point on the path between the mobile and QAP.
Figure 6: Mobile Soft-state Object
Interaction
4. Evaluation
To evaluate the programmable mobile network
performance, we have built the mobinet testbed, developed a wireless network
management tool and designed a set of experiments to analyze QOS controlled
handoff, flow bundling, mobile soft-state algorithms. Through the course
of measurement, evaluation and re- design we managed to substantially improve
upon our initial baseline design.
4.1 The Experimental Environment
At the network level, mobinet provides
wireless access to the Internet and comprises four ATM switches [Footnote
#7] (viz. ATML Virata, Fore ASX/100, NEC model 5, Scorpio Stinger switches)
and four wireless access points as illustrated in Figure 3. The access
points run the mobiware code and are based on a set of multi-homed 200
MHz pentium PCs that provide radio access to a wireline IP/ATM switched
access network. High performance notebooks (viz. IBM, Gateway and NEC)
provide support to mobile applications and mobile access to the access
network. Wireline ATM links operate at OC-3 rates between the switches
and fixed end-systems, and at 25 Mbps between the switches and access points.
Currently, the air-interface is based on WaveLAN operating at 2 Mbps in
the 2.5 GHz band. A future version of the mobinet will operate at 25 Mbps
spectrum in the 5GHz NII/SUPERNet unlicensed band. Mobile capable switches,
access points and mobile devices are abstracted as programmable CORBA objects.
Mobiware requires IIOP CORBA for mobile signaling and adaptation management.
For the results provided in this overview the mobile devices and access points
use IONA's Orbix CORBA running under Windows NT, and the mobile capable
switches/routers use IONA's Orbix running under UNIX or proprietary operating
systems.
We have designed and implemented mobiman
a Java- based management tool for wireless networks. Mobiman drives experiments,
displays recorded statistics, and provides network management capability
to view the network and inspect the state of any mobile device. To measure
the performance of the programmable mobile network we inserted measurement
checkpoints throughout the code and recorded performance statistics during
the experiments. Mobiman, which runs on any fixed or mobile device, can
remotely target any mobile device operating in mobinet displaying mobile
device object statistics. It can set up flows, turn on/off flow bundling
and mobile soft-state, interact with media scaling and adaptive FEC control
at the transport level, and force handoff operations to occur. Measured
information displayed by the mobiman wireless management tool comprises
flow information, quality of service measurements (e.g., signal level and
access point bandwidth availability) and experimental checkpoint measurements.
Mobiman displays measured information, wireless network topology and mobile
device location in a control window as illustrated in Figure 7. When a
mobile device is selected by mobiman a control window indicates the state
of the mobile, e.g., three flows are delivered to `mobile-air#1' as illustrated
in Figure 7. In this example, the mobile-air#1 is running mobiman and the
three flows correspond to the playout of the 'True Lies' video and two
low-resolution text windows. A flow setup panel appears in the top-left
corner of Figure 7.
Figure 7: Mobiman: A Java-based
Management Tool
Microsoft's Active Movie is used for the
reception, decoding and rendering of digital video. It provides a software
tool capable of controlling and processing streams of multimedia data.
Active movie uses modular components called filters and filter graphs.
Typically, a filter graph consists of a source filter that provides the
system with multimedia data, a transform filter that performs data decompression
and a rendering filter. Active Movie's filter graph has been enhanced with
an appropriate mobiware static transport object to perform synchronization
of flows during handoff, controls delay- jitter control and rate control.
4.2 The Experiments
Our evaluation methodology is based on
a set of experiments designed to investigate the performance of the mobiware
programmable mobile network in supporting mobile multimedia communications.
The use of CORBA for mobile signaling, wireless adaptation and mobile network
programmability is a novel aspect of our implementation. CORBA objects
run at the edges and in the mobile network to support wireless-QOS and
mobile- QOS. An important aspect of our evaluation was to determine if
such distributed object technology is viable in supporting mobile signaling
and adaptation management.
Figure 8a: Handoff with Flow bundling
ON
Figure 8b: Handoff with Flow Bundling
OFF
4.2.1 Handoff Analysis
An important objective of this experiment
is to measure the handoff latency and understand how the signaling system
delays breakdown. In this experiment we investigated the handoff of a single
flow. Handoff with flow bundling is described in the next section. For
this experiment we streamed a single video flow from a fixed network server
(S1) to a mobile device (M1) as illustrated in Figure 3. The mobile device
moved repeatedly between access points AP2 and AP3 with the crossover switch
located at the ATML2 switch. The average handoff latency for the baseline
code was measured to be 171 msec. This measurement broke down into 102
msec for mobile registration and object binding, 30 msec for wireless ATM
connection setup and 39 msec for wireline connection setup. The greatest
portion of the total latency time being absorbed by the binding process
between objects during handoff. As described in Section 3.2, the mobile
device object remotely binds to the access point object at the forward
access point. Following this, the access point locally binds to a QOS mapper
object and remotely binds to a mobility agent object for handoff control.
The following enhancements were made to
the baseline code to reduce binding and Remote Procedure Call (RPC) overhead.
First, by collapsing unnecessarily independent CORBA objects into a single
object the binding overhead was reduced. To reduce binding over the air-interface
the mobile proxy and QOS mapper located at the access point object were
collapsed into a single CORBA object. This reduced the number of CORBA
requests across the air-interface reducing the binding time from 102 msec
to 42 msec. Collapsing objects in this manner reduced the handoff latency
to 111 msec as illustrated in Figure 9. Next, by reducing the number of
CORBA RPCs the overhead between objects during handoff was further reduced.
An RPC across the air-interface between the mobile and access point took
an average of 15 msec to complete. The number of RPCs between the mobile
and access point was reduced from four to two (viz. registration, handoff
request). Reducing the number of CORBA RPCs during handoff reduced the
handoff latency by 28 msec to 83 msec as illustrated in Figure 9. The final
enhancement to the baseline handoff code exploited the concept of caching
object bindings. In order to eliminate the binding latency, we set up and
cached bindings between remote CORBA objects prior to handoff being initiated;
we call this pre-binding. All access points periodically broadcast their
beacons that include address information, signal strength and available
bandwidth resources at the access point. When mobile devices receive these
beacons in promiscuous mode they register the signal quality in lieu of
a possible handoff to a new access point. A pre-bind capability was added
to the programmable mobile network to allow mobile devices to pre-bind
to neighboring access points in advance of handoff. The pre-binding criterion
is based on the signal strength and available resources. The pre-binding
algorithm issues a pre-bind to an access point object on- the-fly establishing
TCP connections for the CORBA IIOP between the mobile device and the access
points. Another enhancement establishes bindings between all access point
objects in a domain and its associated mobile agent object. This final
enhancement reduced the average handoff latency from 83 msec to 41 msec.
Figure 9: Handoff Latency Measurement
Results
4.2.2 Flow Bundling Analysis
This experiment evaluates the performance
gains using flow bundling during handoff. We observe the performance of
handing off multiple flows with flow bundling disabled and then enabled.
In the experiment video is streamed from three independent sources (viz.
S1, S2, S3) across the network to a single mobile device, which is repeatedly
moving between access points AP2 and AP3. When flow bundling is disabled
(as illustrated in Figure 8b) each flow S1, S2 and S3 is independently
re-routed during handoff via the ATML1, ATML2 and ATML3 crossover switches,
respectively. When flow bundling is enabled all flows are bundled at a
per-mobile QAP located at the Scorpio switch and re-routed during handoff
via a single cross over switch located at the ATML2 switch as illustrated
in Figure 8a.
In this experiment, we vary the number
of video streams transported to/from a mobile device from one to ten flows
with bundling enabled and disabled and measure the handoff latency. We
observe that the results from the baseline measurement highlight the performance
improvement (i.e., speed up in handoff) gained using flow bundling techniques
in the access network as illustrated in Figure 10. As indicated in the
figure the benefit of flow bundling becomes more pronounced as the number
of flows increases. For example, the handoff latency for two flows is 200
msec with flow bundling enabled and 250 msec when bundling is disabled.
For ten flows with and without flow bundling enabled the handoff latency
is 320 msec and 780 msec, respectively. The benefit of flow bundling reduces
the handoff latency, and importantly, simplifies the state management of
flows in the cellular access network. The adoption of flow bundling provides
an improvement of 20% for two flows and 59% for ten flows. With flow bundling
enabled (as illustrated in Figure 8a) the handoff latency converges whereas
with flow bundling disabled, the latency increases almost linearly as the
number of flows increases. These results indicate the benefit of using
flow bundling to reduce handoff latency and signaling overhead. This is
mainly due to the fact that all interactions between objects during handoff
deals with aggregated signaling rather than per- flow signaling.
The baseline code only provides flow bundling
support between the mobile device, access point and mobility agent objects.
The interface between the mobility agent and the switch server is per-flow.
Another observation is that the GSMP interface between the switch server
object and switch does not provide any support for aggregation, i.e., GSMP
client cannot update the switch table for more than one VCI pair. To address
this we enhanced the GSMP interfaces used by the switch server object.
This resulted in seamless support for flow bundling aggregation from the
mobile device to the switch tables providing some level of speed up. The
GSMP enhancements include a `parallel' enhancement, which did not require
any changes to the GSMP code. In this case, for two flows the latency for
total GSMP messages is 849 sec without aggregation, and 511 sec with aggregation,
showing a 40% improvement. With increasing number of flows, the total gain
obtained by aggregation increases to 70% for ten flows (3907 sec vs. 1184
sec).
Figure 10: Performance Gain with
Flow Bundling
The baseline code was enhanced to support
the optimization discussed in section 4.2.1. In addition, the GSMP messaging
between the switch server and GSMP provided some incremental improvements
as discussed above. Considering the enhanced code the handoff latency was
reduced to 56 msec with bundling and to 67 msec without bundling for two
flows showing a 16% improvement. In contrast, the handoff latency for ten
flows was 155 msec and 420 msec with and without bundling, respectively,
showing a 63% improvement.
4.2.3 Mobile Soft-State Analysis
This experiment demonstrates the ability
of mobile devices to adapt their bandwidth needs to changes in wireless-QOS
and mobile-QOS based on mobile soft- state. Mobile devices periodically
probe and adapt to changes in available resources in wireless access networks.
Users characterize flows using an adaptive- QOS API (described in Section
2.1) that includes a utility function and adaptation policy. In this experiment
we present two scenarios that illustrate the benefit of mobile soft-state
and QOS adaptation management in wireless and mobile environments.
Figure 11a: QOS Adaptation within
a Single Cell
Figure 11b: Handoff Driven QOS
Adaptation
The first scenario illustrated in Figure
11a shows the QOS adaptive behavior of two mobile devices M1 and M2 operating
within a single wireless cell. Mobile devices M1 and M2 receive the `True
Lies' and `Star Wars' video streams, respectively. Both video flows are
based on discretely-adaptive utility functions, i.e., multi- resolution
flows. Initially, M1 receive a base layer (BL) at 80 Kbps and M2 a base
and enhancement layer (E1) at 150 Kbps. Currently, the adaptive-QOS service
gives priority to support the minimum bandwidth requirements of multi-resolution
flows [Campbell,95]. During the scenario, M2
registers an increase in bit error rate as it moves away from its current
access point. Adaptive FEC is applied to the video between the access point
and M2 based on the observed SNR and the measured bit error rate.
An adaptive FEC object selectively codes
the base and enhancement layers of the ôstar warsö video increasing
the bandwidth consumed by M2 from 150 Kbps to 250 Kbps. For the experiment,
the maximum capacity of the air-interface is set to 330 Kbps and approximately
45 seconds into the scenario the M2 video is adapted back to the base layer
with FEC only. Resources released by M2 are consumed by M1 increasing its
utility at 50 seconds [Footnote #8] into the scenario.
This situation remains constant until M2 handoffs to a new access point
after 80 sec allowing the access point to deliver another enhancement layer
to M1. Note that mobile initiated adaptation to released resources (i.e.,
scaling up) is somewhat dependent on the refresh/probe interval.
When a new mobile device M3 enters the
cell around 120 sec mobile device M1 explicitly scales back by dropping
an enhancement layer. Towards the end of the scenario M3 probes and scales
up to a better perceptible quality as M1 hands off to a new access point.
At 140 sec into the scenario, mobile device M3 sets up a new flow to access
web services downloading a GIF file at a rate of 70 Kbps scaling up to
135 Kbps. In related work [Bianchi,98a] we are
investigating a generalized adaptation policy mechanism where applications
can specify application specific adaptation semantics. For example, some
applications would not wish to experience the adaptation observed by M1
while others may be as aggressive as M3 in exploiting any available resources.
The second experiment highlights a number of different QOS adaptation scenarios
that can take place during hand off. In this experiment, mobile devices
hand off to the access point AP2 from AP1 and AP3. In this experiment,
QOS adaptation is not, however, based on the mobile soft-state refresh
mechanism described and evaluated in the previous section. Rather, as part
of the QOS re-negotiation phase during handoff, mobile devices scale their
quality of service needs to match the available resources.
The handoff point at which each of the
four mobile devices (viz., M1, M2, M3, M4) enter the new access point AP2
is illustrated in the trace shown in Figure 11b. The type of adaptation
that takes place after handoff points, which is marked as H1 through H4,
is illustrated. During handoff a number of adaptation scenarios may occur
depending on the available resources and the ability of existing mobile
devices to adapt. For example, the new access point may force existing
mobile devices to drop enhancement layers to allow a new mobile to enter
the cell with minimum QOS assurances. In this experiment, mobile device
M1 enters the new cell at H1 and scales up its utility to take advantage
of available resources. M1 adaptation policy is to only scale after handoff.
At point H2 mobile device M2 hands off to the access point AP2 and is forced
to scale down to its base layer. Mobile device M3 has an adaptation policy
of never adapting. At H3 the mobile hands off to AP2 and maintains its
current utility. In the final part of the experiment, M4 hands off to AP2
at point H4. In this instance, insufficient resources are available to
support the base layers of M1, M2, M3 and M4 forcing the access point to
deny the handoff.
5. Discussion
We have analyzed the performance of mobiware's
QOS controlled handoff, flow bundling and mobile soft- state algorithms.
While the baseline code raised some initial performance concerns about
the viability of using distributed CORBA object technology for controlling
mobile networks the enhanced software is extremely competitive in relation
to existing work. The latency measured for QOS controlled handoff was reduced
from 171 msec to 41 msec for the handoff of a single flow through two ATM
switches making handoff through a single switch in the order of 20 msec.
While it is difficult to compare results
from different testbeds running different signaling systems we highlight
some measurements from comparable systems found in the literature for the
purpose of qualitative comparison only. In [Naylon,97]
and [Mishra,97] handoff latencies were measured
to be 10 and 30 msec for a single flow through a single cross over switch.
The use of flow bundling techniques in mobile networks shows great performance
increases as the number of flows increase during handoff. The handoff latency
for ten flows is 155 msec when flow bundling was enabled and 420 msec when
disabled. This clearly shows the advantage of such aggregation techniques.
Mobiware's flow bundling compares favorably to the literature. In [Mishra,97]
Mishra reported a handoff latency of 125 msec for ten flows using native
ATM signaling code. Mobile soft-state also exploits aggregation techniques
provided by flow bundling. This allows resource probing to be based on
flow bundles rather than per-flow. QOS adaptation techniques
clearly demonstrate the benefit of mobile soft-state in sharing resources
among competing mobiles in a cell.
6. Conclusion
In this overview we have discussed the design,
implementation and evaluation of an open programmable mobile network based
on distributed object technology called mobiware that dynamically exploits
the intrinsic scalable properties of adaptive mobile applications. A number
of researchers have applied distributed object technology to mobile systems.
Our work, however, differs from these efforts. First as part of the open
signaling community [OPENSIG,96] we are deeply
interested in identifying open programming interfaces for mobile and wireless
networking. In this work we have identified a number of objects, APIs and
algorithms that provide QOS support for adaptive mobile networking. The
mobiware technology we have developed over the last two years marks a considerable
software effort. To our knowledge we are one of the first groups to apply
distributed object technology as a mobile middleware solution for adaptive
mobile networking. Mobiware objects execute on the mobile devices, at the
access points and on switch/routers exposing open APIs that can be programmed
to support mobile signaling, QOS adaptation management and wireless transport.
We observe that once the wireless and mobile APIs have been designed the
programming of new network algorithms, e.g., QOS controlled handoff, flow
bundling and mobile soft-state is straightforward engineering. The source
code distribution for mobiware v1.0 can be freely downloaded from [Mobiware,98]
for experimentation.
7. Acknowledgements
First, we would like to thank the
COMET
Group's
industrial participants.
for their kind support. Next, we would like to thank Aurel A. Lazar and
the xbind team for providing the
xbind broadband
kernel which mobiware is built on. Also, we would like to thank the following
colleagues for their major contributions toward the implementation of
mobiware: Oguz Angin
implemented flow bundles, Anand Balachandran implemented the active media
filters, Javier G. Castellanos, retooled the beacon to include QOS hints,
Michael E. Kounavis implemented the transport system and prepared the
release of the code, Raymond Liao implemented
the signaling system, Chien- Ming Yu
implemented the adaptive error control and finally Yasuro Shobatake
(Toshiba Corp. Japan) implemented the wireless transport management.
We would also like to thank Laura Zhou
for implementing mobiman and Mun Choon Chan (Lucent Technologies) who
implemented the xbind connection management system from where the mobiware
mobility agent grew. Finally, we would like to thank Lucent Technologies for
providing a beacon API and our colleagues in the
OPENSIG
community for their valuable input over the last few years.
8. References
[Angin,98a] O. Angin, A.T. Campbell,
M. E. Kounavis and R. R.-F. Liao, "Open Programmable Mobile Networks" ,
Proc. Eight Intl Workshop on Network and Operating System Support for Digital
Audio and Video (NOSSDAV), Cambridge, England July, 1998.
[Angin,98b] O. Angin, A.T. Campbell,
M. E. Kounavis and R. R.-F. Liao, "Enabling the Creation, Control and Management
of Adaptive Wireless Services in Programmable Mobile Networks" , Center
for Telecommunications Research Technical Report submitted for publication,
June 1998.
[Balachandran,97] A. Balachandran,
A.T. Campbell and M.E. Kounavis, "Active Filters: Delivering Scalable Media
to Mobile Devices", Proc. Seventh Intl. Workshop on Network and Operating
System Support for Digital Audio and Video (NOSSDAV), St Louis, May, 1997.
[Bianchi,98a] G. Bianchi, A.T. Campbell
and R. R.-F. Liao, "Supporting Utility-Fair Adaptive Services in Wireless
Networks", Proc. of 6th International Workshop on Quality of Service (IWQOS
'98), Napa Valley, May, 1998.
[Bianchi,98b] G. Bianchi and A.
T. Campbell, ôA Programmable MACö, to appear: ICUPC '98, Florence,
October, 1998.
[Campbell,95] A.T. Campbell, D.
Hutchison, and C. Aurrecoechea, "Dynamic QOS Management for Scalable Video
Flows" , Proc. Fifth Intl Workshop on Network and Operating System Support
for Digital Audio and Video (NOSSDAV), Durham, New Hampshire, 1995.
[Campbell,97a] A.T. Campbell and
G. Coulson, "QOS Adaptive Transports: Delivering Scalable Media to the
Desk Top", IEEE Network., March 1997.
[Campbell,97b] A.T. Campbell, "Mobiware:
QOS-Aware Middleware for Mobile Multimedia Communications", 7th IFIP International
Conference on High Performance Networking, White Plains, NY, April, 1997.
[Clark,90] D. Clark and D. Tennenhouse,
"Architectural Considerations for a New Generation of Protocols", Proc.
Sigcomm'90, Sept. 1990.
[Corson,98] M.S. Corson and A.T.
Campbell, "Toward Supporting Quality of Service in Mobile Ad hoc Networks",
work in progress session, First IEEE OPENARCH'98, San Francisco, CA, April
1998.
[Daedalus,96] Daedalus/BARWAN project
at UC Berkeley http://daedalus.cs.berkeley.edu/index.html
[Katz,94] R. H. Katz, "Adaptation
and Mobility in Wireless Information Systems", IEEE Personal Communications
Magazine, Vol. 1, No.1, First Quarter 1994.
[Lazar,97] A. A. Lazar, "Programming
Telecommunication Networks", IEEE Network, October 1997.
[Lee,95] K. Lee, "Adaptive Network
Support for Mobile Multimedia", Proc. of Mobicom'95, Berkeley, CA, Nov.
1995.
[Lee,98] S-B Lee and A.T. Campbell,
"INSIGNIA: Inband Signaling for Mobile Ad Hoc Networking", Center for Telecommunications
Research Technical Report submitted for publication, June, 1998.
[Lu,97] S. Lu, K.-W. Lee, V. Bharghavan,
"Adaptive Service in Mobile Computing Environment", Proc. of 5th IFIP Intl.
Workshop on Quality of Service (IWQOS'97), New York, May 1997, pp25-36.
[Merwe,97] Van der J.E. Merwe and
I. Leslie,"Switchlets and Dynamic Virtual ATM Networks", Integrated Network
Management V, A.A. Lazar, R. Saracco and R. Stadler, eds., Chapman and
Hall, New York, 1997.
[Mishra,97] P. Mishra, "Implementation
and Experimental Evaluation of Mobility-enhanced ATM Signaling", OPENSIG
FALL '97 Workshop Open Signaling for ATM, Internet and Mobile Networks,
New York, October, 1997.
[OMG,93] Object Management Group
(OMG), The Common Object Request Broker: Architecture and Specification,
Rev. 1.2, Dec. 1993.
[OPENSIG,96] http://comet.columbia.edu/opensig/
[Mobicom,97] Panel on QoS in the
Next Generation Mobile Internet: What is Feasible?, Chaired by A.T. Campbell,
The Third Annual ACM/IEEE International Conference on Mobile Computing
and Networking (MobiCom'97), Budapest, October, 1997.
[Mobiware,98] Mobiware v1.0 Source
Code Distribution: http://comet.columbia.edu/mobiware
[Naghshineh,97] Naghshineh M., and
M. Willebeek-LeMair, "End-to-End QOS Provisioning in Multimedia Wireless/Mobile
Networks" IEEE Network, March 1997.
[Naylon,97] Naylon, J., "Radio Handover
Measurements", OPENSIG SPRING '97 Workshop Open Signaling for ATM, Internet
and Mobile Networks, Cambridge, England, April 1997.
[Newman,96] P. Newman et al., "Ipsilon's
General Switch Management Protocol Specification," IETF RFC 1987, Aug.
1996.
[Porter,95] J. Porter, A. Hopper,
D. Gilmurray, O. Mason, J. Naylon, and A. Jones,"The ORL Radio ATM System,
Architecture and Implementation", ORL Technical Report, 1995
[Satya,96] M. Satyanarayanan, "Mobile
Information Access", IEEE Personal Commun. Mag., Vol. 3, No. 1, Feb. 1996.
[Tennenhouse,97] D. L. Tennenhouse,
J. M. Smith, W. D. Sincoskie, D. J. Wetherall, G. J. Minden, "A Survey
of Active Network Research", IEEE Communications Mag., Vol. 35, No. 1,
pp 80-86. January 1997.
[xbind,96] xbind Broadband Kernel
source code: http://comet.columbia.edu/xbind
[Zenel,97] B. Zenel and D. Duchamp,
"A General Proxy Filtering Mechanism Applied to the Mobile Environment",
Proc. of Mobicom'97, Budapest, Hungary, Sept. 1997.
9. Footnotes:
[1] A generalization of this approach
is detailed in [Bianchi,98a]. Adaptive mobile applications supply adaptation
handlers, which implement application-specific adaptation policies supporting
more sophisticated levels of adaptation than the current menu options (e.g.,
fast, handoff) offered in the existing system. Mobiware exposes a set of
low level programming APIs to allow the application to control its adaptation
strategy.
[2] It is envisioned that the Internet
will provide interconnectivity between a set of these access networks providing
a programmable cellular internetworking environment. In this case the Internet
will support macro-level mobility using Mobile IP and the programmable
access networks will support QOS controlled handoff, flow bundling, mobile
soft-state and active wireless transport services to mobile devices.
[3] In this overview
we focus on adaptive quality of service support for programmable
broadband cellular communications. Currently, mobile ad hoc support is
under investigation by the COMET group [Coroson,98] [Lee,98].
[4] Mobiware programmable objects
also include active transport objects based on Java classes that æpluginÆ
to the wireless transport data path at access points to provide value-added
quality of service adaptation support. Currently, the active wireless transport
supports active media scaling and adaptive FEC support.
[5] The programmable object oriented
nature of the mobiware signaling system makes it easy to æprogramÆ
different styles of handoff algorithm (e.g., network initiated handoff
that is hard in nature) that can operate in parallel to other styles of
handoff over mobinet. For full details of the programming interfaces and
results from the different styles of programmable handoff see [Angin,98b].
[6] In addition to monitoring delay,
loss and bandwidth characteristics of flows a mobile device object receives
beacons that inform it of the state of the wireless link. The beacon informs
the mobile device of the channel conditions upon the reception of each
packet arrival reporting the signal level, silence level signal quality
and antenna selected. The signal and silence level area derived from the
receiver's automatic gain control settings. Beacon messages are augmented
with a 16-bit field that indicates the available resources at the access
point. The mobile device can use this to scale down its request for bandwidth
resources during handoff given that the bottleneck node is typically the
access point. Our radios are based on WaveLAN operating in the 2.4-2.8
GHz ISM band, which we have low level access to for programming the beacon.
[7] We modify the concept of switchlets
[Merwe,97] to provide an extended network for mobiware evaluation. Switchlets
allow multiple virtual network elements to be operational within the same
physical nodes. Three ATM switches (ATML 1, 2, and 3 as shown in Figure
3) in our network are switchlets physically co-located at the same physical
switch. For example, packets traversing three switchets located at the
one physical switch travel across the physical switch three times via cables
that directly connect one port to another in the same switch. Each switchlet
corresponds to a different CORBA switch server object with different name
space, and manages its own resources and controls connections independently
from others.
[8] Mobile device M1 probes and
adapts to the available bandwidth within a single refresh interval that
is currently set to 10 sec. There are a number of tradeoffs in setting
the probe interval. A smaller duration allows mobile devices to aggressively
grab resources on a fast time scale. However, this increases the signaling
load overhead. Currently, we are investigating the coupling of the probing
and adaptation time scales to the application level adaptation policy [Bianchi,98a].
