Upload
alexander-schill
View
213
Download
0
Embed Size (px)
Citation preview
ELSEWIER Computer Networks and ISDN Systems 28 (1996) 1915-1927
Internetworking over ATM: Experiences with IP/IPng and RSVP
Alexander Schill * , Sabine Kiihn I, Frank Breiter ’ Dresden Uniuersity of Technology. Department of Computer Science. 01062 Dresden, Germany
Accepted 2 July 1996
Abstract
This paper describes recent experiences with evaluating and implementing advanced intemetwork communication protocols on top of ATM. First, performance results with conventional TCP/IP over ATM based on Digital Equipment’s Gigaswitch/ATM are reported. It becomes obvious that current protocols must be tuned specifically in order to exploit ATM performance. In order to address advanced quality of service issues based on resource reservation, the paper describes an implementation of IPng (IP next generation) and RSVP (Resource Reservation Protocol) over ATM. Solutions for mapping quality of service and traffic parameters in an adequate way are presented. Moreover, the issue of address mapping from Prig onto ATM is discussed. Implementation results and experiences in these areas are illustrated. Finally, ongoing current work on resource reservation in advance is presented. It is outlined that longer-term resource planning and scheduling provides additional benefits for selected ATM applications.
Keywords: ATM; IPng; RSVP; Quality of service; Resource reservation in advance
1. Introduction
ATM components for local area networks [21] have become widely available as products. With the current UN1 3.1 (User Network Interface) specifica- tion and the emerging UN1 4.0 specification of the ATM forum [3,4], interoperability can also be achieved on a broad basis. It now becomes more and more important to actually support applications on top of ATM with sufficient performance, and also in heterogeneous network environments. Although it is already possible to run applications directly over AAL. (ATM Adaptation Layer 51, additional trans-
* Corresponding author. Email: [email protected]. ’ Email: (kuehn,breiter)@ibdr.inf.tu-dresden.de.
port- and network-level protocols are required in heterogeneous settings, for example with Ethernet, FDDI, and ATM subnetworks being interconnected [I, 191. The typical choice of many vendors is to offer the TCP/IP protocol suite [24] over ATM, based on the IP over ATM recommendation of the ATM Forum [ 181. LAN Emulation [ 15,171 presents another alternative, however with significant limitations. Most important, compatibility of existing applica- tions with ATM is achieved by using IP over ATM.
In Section 2, we first present experiences and performance results for IP over ATM based on a local ATM network using DEC Gigaswitch/ATM. Comparisons with other standard networking tech- nologies are also performed experimentally. It be- comes obvious that current transport protocols are not ideally suited for ATM, and that tuning mecha- nisms and functional improvements are required.
0169-7552/96/$15.00 Copyright 0 1996 Elsevier Science B.V. All rights reserved PII SO1 69.7552(96)00086-4
1916 A. Shill et uI./Computer Nenwrks und ISDN Systems 28 (1996) 1915-1927
Meanwhile, the IETF (Internet Engineering Task similar studies, it is only based on analytical and Force) has also specified a follow-on version of IP, numerical models and does not include practical the IP version 6 (or IPng - IP next generation) measurements. While many research efforts also protocol [ 13,14,20]. The major goal is to enhance the concentrate on QoS specification and supervision IP address space due to the rapid growth of the [6,25], the actual implementation of QoS-related Internet and to offer additional functionality. More- reservation protocols especially over ATM has hardly over, the resource reservation protocol RSVP been addressed yet, with a few exceptions such as [5,11,27] has been developed. Such a protocol is the ST-II work described in [9] and a rough compari- crucial for guaranteeing quality of service (QoS) son of the IETF and ATM services models [8]. characteristics based on explicitly reserved network, Although resource reservation in advance has been memory and processing resources [6]. It is of particu- considered by several authors already [22,23,26], lar importance in heterogeneous networks with par- implementation concepts are still at a very early tial ATM infrastructures. stage.
Section 3 therefore addresses IPng and RSVP and reports concepts and first experiences with IPng and RSVP over ATM. In particular, the problem of mapping QoS and traffic parameters in an adequate way is discussed. This is of specific importance as the kind of QoS specification differs significantly between RSVP and ATM so that the mapping is non-trivial. Moreover, concepts for mapping IPng addresses onto ATM addresses are also presented, and relevant implementation-level aspects are dis- cussed. Based on this work, early experiments with these new protocols have become possible; this way, the new functionality can readily be exploited by emerging ATM applications, especially in heteroge- neous network environments.
2. IP over ATM: Performance evaluation
In this section, we present selected performance results of conventional TCP/IP over ATM and dis- cuss our experiences. First, our experimental envi- ronment is briefly introduced.
2.1. Network/structure
Section 4 discusses ongoing research work on resource reservation in advance. We present new concepts for longer-term management of distributed resource requirements in ATM settings. The require- ments and basic concepts of an adequate resource scheduling in conjunction with reservation protocols are discussed. This way, specific quality of service characteristics for important application scenarios can be guaranteed at a rather early stage. Section 5 presents concluding remarks and summarizes the major findings.
Fig. 1 shows our network structure. Several multi- media workstations of type DECstation 3000 AXP 700 and 300 and a server are connected with our DEC Gigaswitch/ATM via fibre optic links. ATM access is implemented by adapter cards in the work- stations and by line cards in the switch. The adapter cards perform cell generation from input packets of variable size and transmit the cells using SONET/SDH frames with standard 155 Mbit/s per channel via multimode fibre. Cell assembly and dis- assembly is done in hardware. Only AALS is cur- rently implemented.
According to the knowledge of the authors, only few results in these areas are found in the current literature so far. In [ 161, the transfer of data over ATM using available bit rate is examined. Memory management is discussed and advanced switching concepts are proposed. However, higher-level proto- cols are not investigated yet. Ref. [7] presents a higher-level investigation of application-level perfor- mance in ATM networks; however, like many other
The switch itself offers a total performance of 10.4 Gigabit/s and is input-buffered. Possible head- of-line-blocking, a potential problem of input buffer- ing, is reduced by a specific output port allocation algorithm (parallel iterative matching). Both PVCs (permanent virtual channels) and SVCs (switched virtual channels) are supported; signalling is based on ITU Q.93B with 4.2931 as the follow-on version. Moreover, CBR (constant bit rate), VBR (variable bit rate), and ABR (available bit rate), both with point-to-point and point-to-multipoint VCs, are basi- cally possible. However, the driver and subsystem
A. Schill et d/Computer Networks und ISDN Systems 28 (1996) 1915-1927 1917
GIGAswitch / ATM DECstation 300 AXP 300
DECstation 300 AXP 700
Fig. 1. Experimental environment: Structural overview.
software currently does not support CBR yet, so that the experiments were mainly based on ABR.
The multimedia workstations are also connected via Ethernet and have access to another Ethernet switch, and also to an FDDI ring and in this way to the Internet via a concentrator. Each station is equipped with typical devices such as cameras, mi- crophones, speakers, etc., using MME (Multimedia Environment) as an internal software platform. Over all, the installation and maintenance of our environ- ment did not create major problems, and existing applications could easily be ported to run within this infrastructure.
2.2. Results
Within our experiments, we evaluated the perfor- mance of TCP/IP over ATM. Fig. 2 shows the associated protocol structure. On top of the ATM hardware. a driver module provides the interface to the connection management module (CMM). The CMM handles connection establishment and coordi- nates the other modules. It also has an interface to the signalling module (implementation of signalling protocol), to the ATM address resolution protocol
(ATM ARP with dedicated ARP server) and to the IP convergence module. This module maps classical IP onto ATM; this includes access to the ATM ARP server, initiation of connection establishment to IP destinations, and transfer of data.
The application consists of a bidirectional client/server interaction. A component named “netserver” receives data while another component at each peer site named “netperf” sends data ac- cording to a load specification. Then on each site performance characteristics are evaluated. The com- ponents interact via a conventional socket layer of- fered by TCP. In alternative tests, Ethernet and FDDI were also used instead of ATM.
ATM. Ethernet. FDDI ,
Fig. 2. Protocol structure: Overview
1918 A. &hill et al./Computrr Networks und ISDN Systems 28 (19961 1915-1927
Fig. 3. TCP/IP over ATM: Throughput with varying buffer and message sues.
Fig. 3 shows a major summary diagram of the results. Both the message sizes transferred via TCP/IP and the buffer sizes at the receiver’s site were varied; initial experiments have already shown a strong influence of both parameters. The major target parameter was the actual throughput that could be achieved. First, transmission was unidirectional only.
The maximum throughput achieved was 135 Mbit/s with optimal parameter values. Although this equals 87% of the physical bandwidth, it also is notable that the CPU load of more than 60% was
significant then, caused both by the sender and re- ceiver applications and by protocol processing. Nev- ertheless, the experiment has shown that the band- width of ATM can only be exploited based on adequate protocol parameter setting and sufficient CPU capacity. This means, that without tuning the protocol parameters, namely message sizes and buffer sizes, higher-level transport and application protocols present major performance bottlenecks in ATM ap- plications. In any case, it becomes obvious that local protocol processing rather than physical communica- tion causes the major performance limits with today’s hardware.
With a buffer size of less than 64 kbytes, perfor- mance dropped significantly, for example down to values between 60 and 90 Mbit/s for buffers of 32 kbytes. The higher values (90 Mbit/s) could only be achieved with significant message sizes. However, many applications often do not deliver messages that are large enough to achieve satisfying performance. For example, we also observed these problems with bulk data transfer via existing remote procedure call protocols over ATM. In these cases, the constant part of the overhead of protocol processing has increasing influence. Of course, especially very small messages resulted in very poor performance. Similarly, very small buffers lead to very poor performance, too, as
Fig. 4. Throughput comparison among heterogeneous networks.
A. &hill et d/Computer Networks urul ISDN Systems 28 (1996) 1915-1927 1919
buffer overflow resulted in loss of data and subse- 3. IPng and RSVP: Concepts and implementation quent retransmissions. over ATM
Moreover, buffer sizes may be limited by the hardware, especially if multiple ATM applications coexist on a system. Therefore, it can be recom- mended to consider the possible use of ATM already during application development, for example for de- signing the data transfer phases and the mechanisms to be used. Automatic adaptation of protocol parame- ters according to the underlying network would also be a reasonable goal.
Fig. 4 shows the results of further experiments with bidirectional traffic and with Ethernet, com- pared with ATM, using buffer sizes of 128 kbyte. First, ATM performance drops to 110 Mbit/s and less per connection although a 155 Mbit/s VC is available for each direction. This is caused by signif- icantly growing CPU load, because of the fact that both the sender and receiver applications and proto- cols have to be handled on each machine. The comparison with Ethernet shows that much lower throughput (a maximum of 8.5 Mbit/s) is achieved as expected, and that bidirectional traffic leads to even more significant reductions (down to 3 Mbit/s per connection) due to the shared medium with collisions. For FDDI, similar effects were observed, i.e. a more than 50% performance reduction per connection for bidirectional traffic. The major reason is that the FDDI bandwidth of 100 Mbit/s is to be divided among all communication partners while ATM VCs can be provided exclusively for each pair of stations, and also for each direction.
The deployment of the new Internet Protocol on separate data link technologies is very important in view of a broad propagation concerning both already existent technologies such as Ethernet and new tech- nologies such as ATM. This part of the paper de- scribes an implemented adaptation of Internet Proto- col next generation on ATM. Further, it introduces an approach to solve the problem of mapping QoS parameters required by applications onto ATM pa- rameters to utilize the advantages of resource reser- vation in ATM. These considerations are especially based on the Resource Reservation Protocol (RSVP) as a future constituent of an Integrated Services Internet. The major characteristic of RSVP is an extended support of QoS for the Internet Protocol, whereby IPng is especially suitable supporting such reservation protocol. Therefore we compare the emerging integrated traffic services that are being developed by IETF and ATM Forum.
3.1. Major concepts of IPng
Under normal load conditions (with regular back- ground load), we also observed that the quality of video transmission varied significantly due to the current network and local load conditions. This prob- lem can only be addressed by offering CBR and VBR mechanisms with guaranteed bandwidth. How- ever, this requires resource reservations in the end systems and in the active network components (i.e. switches, routers, bridges etc.). Important resources are bandwidth, memory or buffers, and CPU cycles.
IPng is a new version of the Internet Protocol which is assigned IP version number 6 and is for- mally called IPv6. Because IPng is an evolutionary step from the Internet Protocol (IPv4), it comprises on one hand numerous mechanisms of IP which have been expanded or kept in IPng. On the other hand, it contains new mechanisms and characteristics. The improvements of IPng in contrast to IPv4 fall primar- ily into the following categories:
Vers. Prio. Flow Label
Payload Length Next Header Hop Limit
Source Address .___..___~___.___..__________________
I For these reasons, we are currently working on
the implementation of RSVP (resource reservation protocol) over ATM. Major problems and concepts are discussed below. This work is coupled with the implementation of IPng over ATM which is also described.
----_-----__---__--_________________ I------------------------------------ Destination Address
4 w 32 Bit
Fig. 5. IPng basic header.
1920 A. Schill et ul./Computer Networks und ISDN Systems 28 (1996) 19151927
Packet structure. The IPng header (see Fig. 5) distinguishes from the IP header in such a way, that some elements have been reduced and other ele- ments will be optional. The distinction of options into difference extension headers will reduce the bandwidth needed for the IPng header. Therefore, the IPng basic header size is only the twofold of the IP header even though IPng increases the IP address size from 32 to 128 bit. In addition, the time of packet processing in routers will be reduced because the extension header generally contains information concerning the endsystems. Changes in the way IPng header options are encoded allow for more efficient forwarding, less stringent limits on the length of options, and greater flexibility for considering future options.
Addressing. The extension of the IPng addresses provides more addressing hierarchy levels and a much greater number of addressable nodes. In con- junction with these aspects IPng offers new address formats. One of these is the Anycast Address identi- fying sets of nodes, whereby a packet sent to an Anycast Address is delivered to one of the nodes only.
Security and authentication. IPng includes addi- tional options for the definition of extensions which provide support for authentication, data integrity, and confidentiality. These basic elements of IPng will be included in all its implementations.
Quality of service aspects. The Flow Label and the Priority fields in the IPng header may be used by a host to identify those packets for which a special handling by IPng routers is requested, such as non- default quality of service or “real-time” service. The characteristics of this special handling for the corresponding labelled packets belonging to the same flow may be conveyed by a resource reservation protocol or IPng options. This capability is important in order to support multimedia and realtime applica- tions which require some degree of consistent throughput, delay, and/or jitter.
Transition mechanisms. To facilitate the migra- tion from IP to IPng, IPng includes transition mecha- nisms, allowing an adoption and deployment of IPng in a highly diffuse fashion, and a direct interoperabil- ity between IP and IPng. Examples of such transition mechanisms are the dual IP layer, automatic and configured tunnelling, and the IP header translation
as a special case of the communication of pure IPng with IP hosts.
In conjunction with the adaptation of IPng onto different link layers, IPng hosts need to resolve or determine the neighbour link layer address which is known to reside on attached links. The neighbour discovery protocol will be applied for that, using ICMP messages for information interchange. This is generally done via multicasting the addresses of neighbouring routers, the reachability of neighbour hosts, and some additional information. Moving the address resolution up to the ICMP/IP layer makes IPng more independent from the underlying link layer. However, although the neighbour discovery protocol was designed for the deployment of differ- ent link layers, it is mainly suited for broadcast media like Ethernet.
3.2. Integrating IPng and ATM
Recently, formats and methods were specified for the transmission of IPng packets over different net- works like Ethernet, FDDI and Token Ring. In the following, we outline basic concepts for an initial implementation of IPng over ATM. As a general basis, the following parameters must be derivable from IPng parameters to allow support of applica- tions via ATM networks: - ATM address of destination. - Quality of service and traffic parameters. - Connection states and identifier of the ATM vir-
tual channel. The first three parameters represent information
which are needed for the signalling protocol to estab- lish a virtual connection. For making reservations for dedicated flows, it is required to determine the asso- ciated virtual channel identifier. Facilitating the as- signment of IPng packets to ATM VCs, the IPng packet header contains the flow label, where packets with flow label zero are sent as best effort data.
To determine the ATM destination address ac- cording to the neighbour discovery protocol, the multicast capabilities of ATM must be exploited. That is, before sending multicast messages for neigh- bour solicitation, an address translation of a multicast IPng address into a corresponding ATM address has to be performed. To reduce the overhead which will occur in conjunction with using a central Multicast
A. &hill er al,/Computer Networks and ISDN Systems 28 (1996) 19/5-1927 1921
Address Resolution Server (MARS) [2], the IETF currently discusses several solutions. Currently, ad- dress resolution for our implementation of IPng over ATM is realized based on the principle of classical IP over ATM.
The specification of QoS and traffic parameters for dedicated flows has to be performed by the application. However, with the assistance of RSVP, application-level QoS and traffic parameters can be mapped onto ATM parameters explicitly as dis- cussed below.
3.3. RSVP - Basic concepts
Proposed as an internet draft, RSVP is a known constituent of the Integrated Services Internet. It provides especially real-time applications with guar- anteed, predictable and controlled end-to-end perfor- mance across networks. It is a receiver-oriented pro- tocol, which may be classified as a control protocol of the Internet Protocol. Therefore, it offers a flexi- ble handling of heterogeneous receivers as well as an adaptation to dynamically changing multicast groups. The main task performed by RSVP is signaling of resource requirements at connection setup time. To be precise, RSVP does not actually reserve or allo- cate resources but rather indicates reservation re- quests to the underlying systems.
The transmission of RSVP control information is implemented by encapsulating RSVP packets into IP or IPng packets. Reservations may be performed for both unicast and multicast connections. The basis of a reservation is a detailed description of the flow traffic and the QoS characteristics. In accordance with this fact, RSVP defines the so-called flow and
Sender Router Receiver
I I I 1
ReSV RCW PathE% PathElT ResvTear ResvTear
Fig. 6. Exchange of RSVP messages.
filter specification. The flow specification specifies the traffic (TSpec) using token bucket parameters for describing bursty traffic, and also determines the required QoS parameters (RSpec). The filter specifi- cation contains an identification of the flow for which reservations have been performed. Even though the reservation is receiver-oriented, the initia- tor of a reservation is the sender, which informs appropriate receivers about the characteristics of the flow to be sent (see Fig. 6 for an example, also giving the main protocol elements).
3.4. Integration of RSVP and ATM
As discussed above, reservation protocols are characterized by reservation direction, with RSVP as a receiver-oriented protocol. As opposed to that, the ATM signalling protocol is sender oriented. More- over, using soft states to manage the state informa- tion in nodes (endsystems, routers> facilitates the renegotiation of resource reservations within RSVP. Accordingly to hard states used in ATM, there is no possibility to change the specified QoS parameters after call setup (Q.93b, 4.2931 signaling). In con- trast to the RSVP model, the QoS setup time is combined with connection establishment. These dif- ferent concepts make an integration of RSVP and ATM more difficult. In addition to the general func- tionality differences of RSVP and ATM, RSVP sup- ports sessions of several senders within a multicast group. ATM does not support such kind of be- haviour.
In order to map RSVP onto ATM, there are two possibilities. As RSVP only reserves unidirection- ally, it seems to be possible to establish an ATM connection for sending the RSVP reserve message from the receiver to the sender 1281. This would be possible as ATM is able to reserve in both directions. However, this emulation of a receiver oriented estab- lishment of the virtual channel is only applicable to unicast VCs and is in conflict with the ATM multi- cast environment allowing only unidirectional alloca- tion.
We propose another, more efficient way; receiv- ing a reservation message from the downstream host, the appropriate router or host establishes an ATM connection to the next downstream hop according to the reservation information. On this basis. it will also
1922 A. Schill et al./ Computer Networks und ISDN Systems 28 (1996) 191551927
be possible to establish ATM point-to-multipoint connections, at least to homogeneous receivers (as opposed to RSVP where heterogeneous receivers are supported, too).
The differences between the service classes of the Integrated Services IP and ATM also require detailed analysis of mapping service classes as well as traffic and quality of service parameters. Considering both ATM and IETF service classes, the latter are charac- terized by delay, whereas, in ATM classes of bit rates (ABR, CBR) are specified.
Translating traffic and QoS parameters is an addi- tional service for the layer-to-layer communication during the call establishment phase. After the ATM link layer gets its own QoS and traffic parameters through negotiation and translation, the admission control is performed to control resource availability. The description of data streams by an application should be enabled in such a way that the reservation at the ATM layer complies with the actual require- ments. Therefore, overestimating the resource needs of applications due to an improper mapping has to be avoided to achieve an efficient bandwidth utilization.
Considering the traffic description parameters of RSVP and ATM, ATM allows both variable bit rate (burst, peak) and constant bit rate for transmission of non bursty traffic as opposed to RSVP. The traffic parameters of ATM accordingly comprise Maximum Burst Size (MBS), Sustainable Cell Rate @CR) and Peak Cell Rate (PCR). The mapping of the particular RSVP traffic parameters is discussed in conjunction with an assignment of the service classes (see Table 1). Obviously, it would be possible to transmit all IP flows with requested guaranteed service over con- stant bit rate channels, however this implies an inef- ficient utilization of resources e.g. for transmission of bursty traffic. In such cases, transmission via a variable bit rate channel is more convenient.
Moreover, it is apparent that the IP/RSVP ser- vice classes are not in direct relation to the service classes of ATM. The basis of this consideration are the service classes of ATM (specified in the UN1 4.0 draft of the ATM Forum) which mapping problems that have to be recognized in the future.
For mapping the guaranteed service class, only two ATM classes are possible: the CBR and VBR
Table I RSVP and ATM service classes with resoect to 00s and traffic Dammeters
ATM RSVP
QoS class
CBR 00s
VBR-NRT
Traffic VBR-RT QoS
Traffic
Qos
Traffic
ABR Qos Traffic
UBR Traffic
a Cell Delay Variation - CDV. b Cell Transfer Delay - CTD. ’ Mean Cell Rate - MCR.
CLR CDV a MaxCTDb PCR CLR CDV Max CTD PCR SCR MBS CLR Mean CTD PCR SCR MBS CLR PCR MCR ’ PCR
Service class
Guaranteed service
Predictive service
Controlled delay
Best effort
n/a n/a n/a Average Token Rate (RSpez)
n/a n/a n/a n/a Average Token Rate (R of RSpec) Token Bucket Depth (TSpec)
n/a Delay as part of RSpec
n/a Average Token Rate (TSpec) Token Bucket Depth (TSpec) n/a n/a Average Token Rate (TSpec) No declarations are necessary
A. Schill PI cd./ Cornpurer Nrtnvrks and ISDN Sysrrms 28 11996) 1915-I 927 1923
real time due to guaranteed bandwidth and end-to-end delays. In conjunction with the token bucket depth, the assignment of the appropriate ATM service class may be performed. That means, a token bucket depth with a value equal to one deals with a constant data stream. Considering the CBR service class, peak cell rate is the only traffic parameter which actually has to be specified. However, it can directly be derived from the RSVP average token rate, which will be calculated based on end-to-end delay.
Variable bit streams combined with guaranteed service can be mapped onto the VBR real time class, expecting a token bucket depth value greater than one. Parameters of RSpec and TSpec could be trans- lated onto sustainable cell rate and maximum burst size. PCR as a necessary traffic parameter of the VBR-RT service is currently not derivable from RSVP parameters. Basically, RSVP does not define any QoS parameters for the guaranteed service class yet.
Considering the predictive service of the Inte- grated Services Internet, it provides a very high probability that the end-to-end delay experienced by packets in a flow will not exceed a known limit. The VBR non real time class is especially suitable for this service class because it expects a bounded end- to-end delay, too. The mapping of traffic parameters corresponds to that of VBR real time with exception of end-to-end delay as a component of RSpec.
The controlled delay service imposes relatively low requirements on network components and does not provide any quantified level of assurance about packet delays. Instead, it merely promises to avoid overloads by turning excess traffic away. A compari- son of both the controlled delay service and Avail- able Bit Rate shows that the two classes support a best effort service in assistance with congestion con- trol. As opposed to RSVP, ATM does not provide any end-to-end delay specification. Currently, only the average token rate may be mapped onto the mean cell rate. Pure best effort service without any require- ments on network resources may be mapped onto unspecified bit rate which offers the same character- istics.
In summary, a full translation or derivation of QoS parameters is impossible. For example, the cell loss rate (CLR) can generally not be mapped from RSVP parameters onto ATM parameters. As the
traffic parameters of RSVP only consist of token bucket depth and average token rate, an accurate mapping or interpretation of traffic parameters is even more difficult. Introducing a peak rate as part of TSpec may facilitate the mapping onto ATM parameters. A problem which has to be taken into consideration is the additional overhead. Applica- tions have to consider overhead calculations of the transport and network layers. Introducing a minimum packet size as a component of TSpec would facilitate calculation of overhead. This overhead occurs within the ATM layers in order to reserve the requested bandwidth of applications while considering the ac- tual overhead.
3.5. IPng over ATM: An implementation
IPng over ATM has been implemented within our group based on the concepts discussed above. Ac- cording to the prototypical implementation of IPng and the development stage of ATM, it has first been realized for transmission of IPng packets over ATM using best effort service. This implies that packets are multiplexed over an ABR virtual channel, as reservations based on CBR channels are not yet supported by the ATM subsystem.
The adaptation of IPng to ATM is realized as a kernel module, which is implemented as a conver- gence module providing an adaptation of higher- layered network protocols onto the ATM subsystem. To realize an automatic resolution of the IPng ad- dresses into the corresponding ATM addresses, we currently still use the address resolution protocol CARP). According to the principle of classical IP over ATM, the differentiation of the ARP and IPng packets which are transmitted over the same channel is performed by encapsulating both packets with LLC/SNAP. The basis for such an address resolu- tion is an adapted ARP which is able to handle the 16 byte IPng addresses.
In the future, the LLC encapsulation will not be necessary when using the neighbour discovery proto- col. This will reduce the appropriate overhead and should resolve the problem of single point of failure which the ATM ARP server represents. The adapta- tion of IPng onto ATM will be an evolutionary process. With the possibility to reserve CBR in one of the next releases, the analyses concerning the
1924 A. SchiN et al./Comptuer Networks and ISDN Systems 28 (19961 19/S-1927
adaptation of RSVP onto ATM will be integrated at the implementation level. Concerning reservation for IPng flows, mechanisms for demultiplexing of IPng packets have to be considered. The following strate- gies will be the basis for mapping the IPng flows onto ATM channels: - One-to-one mapping allows the transmission of an
IPng flow via a dedicated virtual channel. - Allowing also a many-to-one mapping, more than
one IPng flow can be assigned to one virtual channel. This will be meaningful for flows with the same QoS service requirements.
- All-to-one mapping is currently used because no reservations are possible. In this way, all packets are delivered via ABR using best effort with flow control. Testing this implementation of IPng over ATM,
no noticeable performance losses were recognized in spite of the double sized IPng header. Such tests were performed using the known network program ttcp for IP and ttcpv6 for IPng to get an overview about the performance differences of both protocols. In summary, our experiences were rather positive, so that this implementation will be used as a reference work for further integrating the remaining concepts discussed in this paper, i.e. flexible address mapping, and integration of the full RSVP mechanisms.
9.n.
ATM Network
Fig. 7. ATM subsystem in conjunction with RSVP
Moreover, for realizing a mapping of RSVP onto ATM, an appropriate API is required. Using the ATM subsystem as a part of the Digital Unix operat- ing system, such an API will be constituent of the convergence modules. With IPng over ATM, this API will be available in the IPng convergence mod- ule and will support real time handling and transport of IPng flows (see Fig. 7). This API will be able to receive reservation information from a local daemon as a part of the RSVP implementation. Based on information contained in a given flow specification, a new ATM VC reservation is feasible after perform- ing the mapping of service classes and parameters as discussed before. In addition to calculating and adding the appropriate overhead (e.g. of the AALS trailer) to traffic parameters and converting RSVP parameters [bit/s] into ATM parameters [cells/s], it is necessary to consider the granularity of cell reser- vations in ATM. Obviously, a rather fine granularity will be necessary between endsystems and ATM switches. The admission control will enforce the implementation of reservations.
Moreover, a classification of IPng flows belong- ing to a dedicated reserved VC will be done in the convergence module, too. This implies the fact, that packets which cannot be identified based on filter specification information will be transmitted over an available bit rate channel only.
4. Resource reservation in advance
Recently, distributed multimedia applications have become more and more popular. In order to grant the required QoS for multimedia applications (e.g. videoconferencing) and to utilize resources of the network effectively, it is necessary to provide mech- anisms for resource reservation in advance (ReRA). As an example, it shall be possible to schedule a videoconference with several partners for a given time in the future. The system shall then calculate and virtually reserve the required resources for that time, however without immediately blocking them. The advantage of ReRA is to prevent rejections of application requests at the time they need to use network resources. That means, using ReRA, a fu- ture lack of resources during actual reservation can be avoided. However, this is in contrast with imme-
A. S&ill et aI./ Computer Networks rend ISDN Systems 28 ((9961 1915-1927 I925
diate reservation where reservations are being per- formed in conjunction with connection establish- ment. This implies, that a connection will only be accepted if there are sufficient resources available to reserve them.
Despite the fact that the necessity of mechanisms and algorithms for ReRA is obvious, this problem has been addressed only by few authors yet. There- fore, currently no standardization is done in this area yet. In summary, it can be stated that most of the known papers briefly describe more general prob- lems to be solved: [ 121 concentrates on resource partitioning and manage a table of advance reserva- tions based on the Tenet protocol suite. Ref. 1221
describes a possible exchange of advance reservation information using the ST-II reservation protocol. Dealing with Integrated Services Internet, [lo] pre- sents an explicit admission control algorithm for predictive service. Moreover, [26] already introduces a possible ReRA architecture and [23] compares the suitability of the Internet-based reservation protocols ST-II and RSVP to support ReRA and considers ReRA as summary of several reservation phases. However, implementations have obviously not reached a mature stage yet.
As a basis for the implementation of ReRA, we are currently working towards the completion of the full RSVP/IPng/TCP protocol stack in cooperation
situation at time t
other immediate and advance
situation at time t+x Fig. 8. Reservation in advance on ATM.
1926 A. &hill et d/Computer Networks atuI ISDN Systems 28 (1996) 1915-1927
with Digital Equipment. One of our major goals is to augment the “immediate” reservation mechanisms of RSVP with advance scheduling and planning techniques. If reservation is performed a rather long time in advance, intermediate topology changes, as well as, short and longer term router failures also have to be considered. Many other specific problems have to be solved, for example, the cooperation of advance and immediate reservation protocols, the management of application priorities, effects of “overbooking” resources, and advance reservation for weakly-specified scenarios, for example with dy- namic integration of new partners. Moreover, we are working towards a cost-based resource reservation in advance: This means that an integration of account- ing management from the user’s and the network provider’s view is to be achieved.
However, as opposed to other related approaches, another major goal is to achieve a mapping of ReRA mechanisms onto actual ATM networks. Mecha- nisms to realize resource reservation for dedicated virtual channels without actually establishing a con- nection yet are required; this is in contrast to the current ATM signalling protocol. Fig. 8 shows an example of ATM connections via two switches where a physical link in between is being shared; in this scenario, only an advance reservation in ATM end systems is not sufficient. It is important to especially reserve resources in the switches and bandwidth on the shared physical channel, however, without estab- lishing the connections in advance. The considerable differences between the RSVP and ATM model, e.g. the architecture of intermediate nodes (RSVP as- sumes routers; ATM expects switches/crossbars) makes ReRA on ATM even more complicated. Con- cepts in these areas will be validated by implementa- tion and concrete applications.
5. Conclusions
This paper has presented performance measure- ments of TCP/IP over ATM, and has described the concepts and implementation of RSVP and IPng over ATM. It became obvious that current protocols have to be tuned explicitly in order to provide satis- fying performance over an ATM network. Applica- tions will also have to explicitly consider the use of
ATM, for example for engineering their data transfer phases to achieve maximum performance. New reservation protocols, namely RSVP, outline interest- ing directions towards guaranteed quality of service. We have shown that a mapping onto ATM is possi- ble but that also new problems arise concerning adequate mapping of QoS parameters. Additional work, both at the conceptual level and at the stan- dardization level, will be required in this area.
In the near future, we plan to do further experi- ments with different applications. We also completed a video transfer tool with explicit scaling mecha- nisms that can be used in this area. Concerning implementation, we are currently working towards the completion of the full RSVP/IPng/TCP proto- col stack in cooperation with Digital Equipment. The presented concepts for resource reservation in ad- vance are also under refinement and we expect to have further implementation experiences soon.
Acknowledgements
We would like to thank Digital Equipment Corpo- ration (European Applied Research Center Karls- ruhe) for extensively supporting the work described in this paper within the European External Research Program. Moreover, we would like to thank all students and colleagues who contributed to the pre- sented results, namely Samer Habib, Sascha Kiimmel, and Andreas Wentland.
References
111 l21
[31
141
[51
l61
A. Alles, ATM Internetworking, Cisco Systems, Inc., 1995. G. Annitage, Using the MARS to support IP Unicast over ATM, draft-armitage-ipatm-mars-unicast-01; Work in progress, Bellcore, 1995. ATM-Forum, UN1 Signaling 4.0, ATM Forum 94-1019R6, Work in progress, 1995. ATM-Forum, Traffic Managment Specification Version 4.0, ATM Forum 950013R6, Work in progress, 1995. R. Braden and L. Zhang, Resource Reservation Protocol (RSVP) - Version I Functional Specification, Work in progress, 1995. A. Campbell, G. Cot&on and D. Hutchison, A quality of service architecture, Cornput. Commun. Rcu. 1 (2) (1994) 6-27.
[71 L. Cidon. R. Guerin and A. Khamisy, An investigation of
A. Schill et ul./Cmnpurer Networks ud ISDN Systems 28 11996) 1915(927 I927
[Sl
[91
[lOI
II II
[I21
[I31
(141
[l51
[I61
application level performance in ATM networks, lEEE hzJ;,- corn, Boston (1995) 845-852. I. Crowcroft, 2. Wang and A. Smith, A rough comparison of the IETF and ATM services models, 1995. L. Delgrossi, R.G. Herrtwich and F. Hoffmann, An imple- mentation of ST-II for the Heidelberg transport system, Internet-working: Reseurch und Experience 5 (1994) 43-69. M. Degermark, T. KBhler, S. Pink and 0. SchelCn, Advance reservation for predictive service, in: Proc. 5th Inrertwat. Workshop on Network and Operating System Support for Digrfal Audio und Video, Durham, 1995. D. Estrin and S. Shenker, Routing support for RSVP, Work in progress, 1995. D. Ferrari. A. Gupta and C. Ventre, Distributed advance reservation of real-time connections, in: Proc. 5th Internat. Workshop on NOSSDAV, Durham, 1995. R. Gilligan and E. Nordmark, Transition mechanism for IPv6 hosts and routers, Work in progress, 1995. R. Hinden and S. Deering, IP Version 6 addressing architec- ture, Work in progress, 1995. R. Jeffries, ATM LAN emulation, Data Commun. 9 (1994) 95-100.
[I71
[I81
Ll91
DO1
[2ll D21
G. Konstantoulakis and G. Stassinopoulos. Transfer of data over ATM networks using Available Bit Rate (ABR), in: Proc. IEEE Symp. on Compurers and Communicutions, Alexandria, Egypt (1995) 2-8. ATM-Forum, LAN Emulation over ATM SpeciIication-Ver- sion 1, 1995. M. Laubach, Classical IP and ARP over ATM, RFC 1577, 1994.
WI
v.41
1251
C. Malamud, inreroperabiliry in To&y’s Computer Net- bvork.s, Prentice-Hall, Englewood Cliffs, NJ, 1992. T. Narten and E. Nordmark, Neighbor discovery for IP Version 6 (IPv6). Work in progress, 1995. C. Partridge, Gigubit Networking, Addison-Wesley, 1994. W. Reinhardt, Advance reservation of network resources for multimedia applications, in: Proc. 2nd M/ACA94, Heidel- berg, 1994. W. Reinhardt, Advance resource reservation and its impact on reservation protocols, Tech. Rept., 1995. R. Stevenson and G. Wright, TCP/IP Illus~rared: The Im- plemenrurion, Vol. 2, Addison-Wesley, 1995. A. Vogel, B. Kerherve, G. Bochmann and J. Gecsei, Dis- tributed multimedia and QoS: A survey, IEEE Mubimedia 2 (2) (1995) 10-19.
051 L.C. Wolf, L. Delgrossi, R. Steinmetz, S. Schaller and H.
1271
[281
Wittig, Issues of reserving resources in advance, IBM Euro- pean Network Center Heidelberg, Tech. Rept. 43.9503, 1995. L. Zhang, S. Deering, D. Estrin, S. Shenker and D. Zappala, RSVP: A new resource reservation protocol, IEEE Nenwrk (September 1993) S- 18. M. Bordon, E. Crawley, B. Davie and S. Batsell, Integration of real-time services in an IP-ATM network architecture, RFC 1821, 1995.
Alexander &hill is a professor of com- puter networks and distributed systems at the Department of Computer Science of Dresden University of Technology. He received an MS in 1986 and a PhD in 1989 from the University of Karls- ruhe, Germany, and spent a year as a post-dot at the IBM Thomas J. Watson Research Center in 1990. His research interests include mobile computing, high performance networks and distributed systems. Professor Schill is a member of
the IEEE Computer Society and the ACM.
Sabine Kbhn and Frank Breiter re- ceived their MS’s (Diploma) in Com- puter Science from the Dresden Univer- sity of Technology in 1995. They are currently Scientific Staff Members in Computer Science at the University of Dresden and their research focuses on construction and deployment of resource reservation in advance in high perfor- mance networks especially in broadband integrated networks.