Download A reservation-based mechanism prevents losses in slotted Optical

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Deep packet inspection wikipedia , lookup

Airborne Networking wikipedia , lookup

Distributed operating system wikipedia , lookup

Recursive InterNetwork Architecture (RINA) wikipedia , lookup

Drift plus penalty wikipedia , lookup

Backpressure routing wikipedia , lookup

IEEE 802.1aq wikipedia , lookup

CAN bus wikipedia , lookup

Passive optical network wikipedia , lookup

Kademlia wikipedia , lookup

UniPro protocol stack wikipedia , lookup

IEEE 1355 wikipedia , lookup

Routing in delay-tolerant networking wikipedia , lookup

Transcript
A reservation-based mechanism
prevents losses in slotted Optical Burst
Switching
J. D. Angelopoulos, H.C. Leligou, K. Kanonakis, Ch. Linardakis, I. Pountourakis
Institute of Communication and Computer systems (ICCS), National Technical
University of Athens, Athens, Greece , Tel:.+30-2107721511, Fax:+30-2107722534email: {jangel, nelly, kokan, chlin, ipount}@telecom.ntua.gr,
Abstract: The biggest disadvantage of Optical Burst Switching lies in the high
packet losses which are inevitable due to the lack of adequate optical buffering.
Thus, the advantage of avoiding payload translation into the electrical domain comes
at the cost of dropping packets contending for the same output port whenever this
contention persists for longer than the limited fiber delay lines can accommodate. A
method to avoid these losses is proposed in this paper by sending first a scout
packet over the control channel that simulates the events that the actual packet will
experience. Once the scout message experiences a drop at any of the intermediate
nodes, the actual packet is not sent but the process is repeated. To enable this, a
slotted approach is adopted with all the necessary control tools. The penalty is the
delay of the indirect reservations and the occasional process repetition limiting the
applicability of this approach to a couple of thousand kilometres.
1. Introduction
The proliferation of bursty IP services, has fuelled intensive research on the design
of flexible, fast reconfigurable optical networks, since transferring such traffic over
SDH necessitates wasteful over-provisioning. In an effort to adjust the IP paradigm
to optical core networks, the Optical Burst Switching architecture has emerged [1],
[2]. The basis of the concept is the separation of control information from the data.
Traffic is aggregated in bursts of variable size and a Burst Header is emitted per
burst, to pave the burst’s way through the nodes. One wavelength is devoted to the
control information which is the only one processed in every node; data travel from
source to destination in the optical domain, without undergoing any optical to
electrical conversion, decreasing the node processing requirements and, thus, cost.
The main achievement is dynamic traffic multiplexing directly in the optical domain.
However, its performance suffers high losses due to the lack of buffers of
comparable size to these of the IP routers.
A lot of alternatives regarding the node control algorithms, such as Tell-And-Go
presented in [3], Just-In Time in [4] and Just-Enough-Time [1] have been
investigated and compared [5]. Although the control algorithms are tightly bounded
to the node architectures, the existence of even limited buffering capabilities
strongly impacts performance as well as the node cost. Similarly does the number
of supported wavelengths and the existence of wavelength converters. In an effort
to enhance performance, fast circuit switching techniques have been adjusted to
OBS systems [6] employing either distributed or centralized control [7], one or two
way reservation techniques. The impact of different traffic aggregation techniques
has also been examined [8]. Recently, the adoption of slotted operation in OBS
systems has been pursued as a promising direction and shown to outperform the
pure OBS systems [9], [10].
When packets are launched from peripheral links towards the input ports of a bufferless node, a situation similar to Aloha is created at each output port of the switch
and the very low efficiency of Aloha is what should be expected. However, in OBS
the situation is much worse since each packet has to cross a number of nodes
(typically 3 to 6) to reach its destination and the joint probability of avoiding all
collisions becomes quite low.
In this paper, the focus is on the design of the control architecture for a slotted OBS
system. Our control scheme leads to loss-less operation, and statistical bounds for
the delay are targeted, while achieving high resource utilisation. The novelty in this
work is that two-way reservations are used per data slot. The principle is to simulate
future events based on the “scouts”, i.e. probing messages, in order to avoid
transmitting data which will be dropped later on during their journey to their
destination. Thus, payload can be launched under the certainty of a collision-free
trip to the destination via all intermediate nodes. The concept extends the
methodology used in reservation-based MAC protocols as employed in PONs [11],
or rings [12]. It can be described as “probe-and-go”, because before emitting a data
slot, the outcome of future contentions in all nodes along the path is first found out
by a scout (instead of the Burst Header). Full wavelength conversion is assumed,
while limited optical buffering capabilities are optional. The rest of the paper is
organized as follows: in section 2 the system architecture is presented, while in
section 3 the node scheduling is described. An assessment of the whole scheme
follows in section 4. The simulation results are presented and commented in section
5 and, finally, conclusions are drawn in section 6.
2. System Architecture
In the OBS situation where, despite the help from WDM channels and FDLs, the
collision situation is quite reminiscent of Aloha protocols, it is natural to seek the
same improvement of performance that slotting brings on the basis of the same
reasons that justified slotted Aloha, i.e. avoidance of the quite large waste and
likelihood of partial collisions. Naturally, the creation of slots out of variable packets
comes at some efficiency penalty, since some slots will inevitably remain partly
filled. In this system, the IP (Ethernet encapsulated) packets are aggregated and
possibly segmented at the system periphery creating fixed size data packets (slots),
which are then transmitted across the OBS domain. Longer bursts are
accommodated in trains of slots.
The architecture on which the two-way slot reservations will apply is similar to any
typical OBS environment and is shown in Fig. 1. Nodes are interconnected via
WDM links consisting of W data wavelengths and the control wavelength. Local
traffic is used to form slots. Electronic buffers are used for this kind of traffic, while
only optical buffering is possible for transit traffic. Queuing is effected per
destination and source routing is employed. Knowledge about each route also
includes the round trip distance measured in slots obtained by ranging after every
route update.
Input λ
Src Addr.
Dest. Addr.
High Class
Queues
F/B
HEC
3
1
Low Class
Queues
Ack
Scout
format
2
Queuing structure
for node 1
Td
7
4
To node 2
6
To node 3
To node 7
To node 2
5
To node 3
To node 7
Figure 1: System architecture
The slotting of the system is similar to the one described in [10] except that no
frames are considered and each slot is accounted for on its own and controlled
without reference to any possible relation to subsequent slots. However, the main
motivation for the choice of a slotted system is not just the avoidance of partial
collisions, but mainly the need to keep an easy accounting of traffic and exploit this
to reach a quite complex closed loop reservation control. To create the fixed-size
slots, it is necessary to include at the entry port of each node a synchronizer like the
one described in [10], which by means of variable delay lines it is capable of
aligning the boundaries of the incoming slots in all input ports and wavelengths, so
as to be able to switch them to the output ports without overlap with the previous
and following slots possibly coming from different input ports. A small guard-band of
a few tens of ns is of course necessary. Each core switch may employ wavelength
converters as well as FDLs in multiples of the slot size. However, while the
aforementioned components remain expensive, it is possible to keep the
wavelengths isolated as extra capacity, at the penalty of the relevant loss of
multiplexing gain.
As in other OBS systems, data slots are preceded by control packets called Burst
Headers; however in this system they will be called scouts, in order to stress the
different functionality and their two-way reservation character. The scouts are used
to test the availability of output links along the path that the data slot will take at a
later slot. The scouts are situated inside control slots of the same size, also
synchronized to data slots. Each scout carries the necessary information for one
data slot including (as shown in Fig. 1) the source and destination address (local
OBS address), the wavelength on which the data slot is to arrive and the offset Td
indicating how many slots later it is to arrive. In addition, in the F/B field, it is marked
whether the scout is traveling in the forward direction (making reservations) or the
reverse (i.e. returning with a positive or negative outcome) while in the Ack field it is
indicated whether the required resources have been successfully reserved. At the
departure of the scout, the parameter Td is set one slot above the known slot
number of the round trip time to the destination. The source and destination address
are both 10 bits wide, covering a network of 1024 edge nodes. Another 10 bits are
used to code the wavelength on which the data slot is to arrive and an equal
number of bits are devoted to the offset Td. Finally, each scout contains a HEC field
and spare bits, to reach an overall size of 64 bits. For a transmission rate of
2.5Gbps in the control wavelength and 100μs slots, each control slot can host more
than 3900 scouts, making congestion in the control channel unlikely.
As the scout enters each core node, the node switch learns that Td slots later a
data slot will enter from the same port and in the indicated wavelength, claiming the
also indicated output, thus implying a request to reserve a targeted future slot in this
output port in the same wavelength (or another available wavelength, depending on
the number of converters in the switch). In contrast with ordinary switches, which
are at any time scheduling the next slot, the nodes in this system are given an early
warning by the scouts and they program future switching actions many slots later in
time, while simultaneously executing current switching actions that were planned
some slots earlier in time.
When a new scout claims an output, any contentions from other scouts entering
from other input ports are resolved within the present slot time considering also any
previous claims for the same slots that are already marked in the scheduling log. (It
is completely irrelevant whether they are actually currently occupied since this is on
the basis of previous planning while the scheduler is concerned with the status of
the FDLs at the future slot time in question).
Switch matrix
Output 2
Slot i+H
Slot i
Slot i+1
Output 1
Slot x
Output K
switch matrix)
Switch matrix
Output 1
Output 2
Slot i+H+1
by the optical
Slot i+2
(to be executed
Slot i+1
Input-output
Permutation
Slot x+1
Output K
Empty slot
“To be assigned”
Figure 2: The switch matrix handled by the node
In the event that no WDM conversion is used or all WDM channels will be busy at
that time in future, and that no delay lines are available either, the unsuccessful
scouts will be notified that their data slots would be dropped. In the latter case, the
scout is switched around and returns home without completing its trip to the other
end carrying a negative reservation, prompting the dispatch of a new scout by the
source edge node. No data slot will be sent, avoiding losses of data and this is the
most important difference to other OBS systems. The winning scouts on the other
hand, continue their journey towards their destination trying to plan a secure
complete trip for the data slot they represent. The end result is that the scouts
execute a preview of what the future would hold if their data slots were to be sent,
probing the traffic conditions along the path and accepting only successful transit
patterns. Thus, positive reservations are used to cause a departure of data and
negative ones to cause the issue of a new scout.
The round trip time of control slots is larger than that of data slots by the extra time
needed in each node for processing. Due to the simplicity of the switching
assignment algorithm, making possible hardware implementations, it is expected
that one slot processing delay in every node is enough. So, for each reservation, at
least the round trip time of control slots is required and each departing scout is
marked with a Td of the round trip time to the destination of the control slot. Each
node reduces this number by one to subtract the one-slot processing needed by the
control and not by the data (irrespective of other changes it may do). Control is
distributed with each node executing the scheduling and accepting or not
reservations on the basis of local information.
3. The scheduling algorithm
Every node acts as source node for its local traffic and in parallel as switch for the
rest of the traffic. Hence, it is responsible for the generation of scouts to serve its
local traffic and for the handling of transit scouts, scheduling its future switching
actions.
The scouts are managed per queue, with each queue associated with a destination.
The scout departs as soon as a payload fitting the slot has been prepared. Once a
scout with a certain Td value has been launched, the relevant data will depart Td
slots later under the condition that a scout with positive reservation will have
returned before. The scout will either return at planned time (i.e. Td-1) or earlier,
albeit with a negative reservation and a new scout will be launched starting a new
count down to the new Td. One will notice that scouts making reservations for long
paths have increased chances of success at first go, since their Τd values are
bound to be higher, addressing future slots that have not yet been reserved by
scouts crossing fewer nodes. However, this imbalance is rectified by the fact that
scouts corresponding to shorter paths have the chance of repeating their attempts
more promptly in case of a negative outcome.
The scouts are managed per queue and do not belong to specific data slots. So, the
first payload slot in line departs with the first scout returning with a positive
reservation, irrespective of whether the specific scout was launched with the arrival
of the payload at the head of the queue. This can happen whenever the scout sent
with the first data slot returned with a negative reservation and hence the new scout
launched will now arrive out of order. Thus, the FIFO principle is never violated in
the same queue, but also as many scouts as payload slots waiting are always in
transit.
4. Assessment
At first glance, the two-way reservations seem to trade-off larger queuing delays
and heavier control for no loss. However, it must be borne in mind that the
equivalent to unsuccessful attempts by scouts in the one-way “classical” OBS
solutions is a quite high percentage of dropped data packets, particularly with bursty
traffic generators. These dropped packets need to be re-transmitted by higher
layers (essentially TCP) at a delay much larger than in this scheme. Such reappearing packets constitute a significant percentage of system load, although they
appear in performance evaluations misleadingly as original load. This can have
significant side-effects on admission control and policing functions at the entry of
the network, since an initially conforming traffic stream can appear non-conforming
by extra load of retransmissions at no fault of the user. Applications and services
can no more calculate the bandwidth they need for proper operation and negotiation
with service providers. Unfortunately, during congestion the problem is aggravated
with losses increasing and causing further load, creating an avalanche effect. In
contrast, the scouts reduce the offered load during congestion and they do so
selectively only for congested paths. They act as backpressure to edge nodes not to
emit more traffic towards bottlenecked queuing points along the path, in addition to
finding valid crossing patterns. This happens because at congested outputs
successful reservations are less likely. Similarly, packets re-transmitted by TCP do
not enter the performance evaluation to bring a significant increase of average
delay. Furthermore, it is more important for many services to enjoy a bounded delay
and this proposal keeps the quantiles of delay at much lower levels.
The experienced delay is dominated by the round trip propagation delay of the twoway reservations. Hence, the tolerable round trip limitation (e.g. 30 to 50ms
translating to 3000 to 5000km distance limit) and the concomitant buffer size at the
edge node restrict the size of each OBS domain that can be covered without resort
to electronic buffering. In addition, signal deterioration after a number of switching
actions may introduce the need for regeneration.
5. Performance Evaluation
The network model used for the verification of the probe-and-go scheme was a 4x4
Torus topology, where all nodes act both as edge and core nodes. There are 64
wavelengths at 10Gps for data and one control wavelength at 2.5Gbps in every link
of the network. The slot size is 100μs, thus accommodating 1Mbit of data. The
length of links between nodes is 220km, corresponding to 11 slots, or 1.1ms.
Hence, the round-trip times are given by 2·n·1.1ms, where n is the number of hops.
Given that the maximum number of hops in the network is 4, the round-trip time can
reach 8.8ms. Full wavelength conversion was assumed in all nodes, but no FDLs
for delaying data. In order to provide a realistic pattern for the input traffic of the
network, a traffic aggregation unit receiving bursty IP traffic from a large number of
sources was simulated, generating 1Mbit payloads. The routing is determined
beforehand choosing those shortest paths that create a symmetric distribution of the
routes over the links. The loading in the scenarios presented in this paper is
symmetric (i.e. all nodes produce the same amount of traffic destined to all other
nodes with equal probability).
Different simulation scenarios have been carried out and the results have shown
that using the probe-and-go concept, the system can handle up to 65% load,
despite the complete lack of fiber delay lines in the nodes, keeping the queuing
delay in controlled levels and of course without any losses. The latter is of extreme
importance, since classical OBS strategies are not only constrained to much lower
loads, but also, as mentioned earlier, suffer significant losses that translate in
additional delays in the form of retransmissions by the upper layers. Under such
circumstances, guaranteeing upper limits of delays as is necessary for supporting
modern delay sensitive traffic becomes questionable. In fig. 3 a and b, the queuing
delay for payloads destined to 1-hop and 4-hop destinations respectively is shown
for two different system load values (60% and 66%).
10
10
9
9
60% Load
60% Load
66% Load
8
66% Load
7
7
6
6
delay (ms)
delay (ms)
8
5
4
5
4
3
3
2
2
1
1
0
0
0
1
2
3
time (s)
4
5
0
1
2
3
4
time (s)
Figure 3: Queuing delay for (a) 1 hop and (b) 4 hop destinations
For loads up to 65%, most scouts return with a positive acknowledgement, hence
the queuing delay is close to the round trip time since each burst departs as soon
as the relevant scout returns back. In the case for the 60% load, both 1 and 4 hop
queues show delay that slightly deviates from the respective round-trip times. As
load increases beyond 65%, the congestion is first mirrored in 1-hop queues, since
the relevant scouts are addressing future slots that are already reserved by higherhop length queues. As a result, although the 1-hop queue experience the
congestion even for 65% load, the delay of the 4-hop queue remains intact
regardless of the increasing load in the system. (In the figure, the delays for 4-hop
destinations in 60% and 66% load coincide.)
6. Conclusion
By first simulating the travel of a packet along an OBS path the scouts can, upon
their return to the sending node, inform the node controller about the fate of the
packet, thus avoiding the emission of those destined to be dropped. This is
translated in further delay but such a delay is also to be found in any other
competitive approach (even if the typical approach in OBS literature is to leave the
issue of retransmission out of the performance since it also involves higher network
layers outside of the optical infrastructure). As a result, the probe-and-go protocol
can avoid the intolerable losses of one-way OBS and the extra delays of higher
layer processing, at the cost of expending on heavier control processing of slot-byslot reservations.
Acknowledgements
This work was funded by the European Commission IP project NOBEL. The views
expressed herein are those of the authors and do not necessarily represent the
position of the whole NOBEL consortium.
5
References
[1] J.S. Turner, “Terabit Burst Switching”, Journal of High Speed Networks, 8, pp. 3-16, 1996.
[2] Qiao, C.; Yoo, M. “Optical burst switching (OBS) – a new paradigm for an optical internet”, Journal of
High Speed Networks, No. 8, 1999, pp. 69-84.
[3] I. Widjaja. Performance analysis of burst admission-control protocols. IEE proceedingscommunications, 142(1):7{14, February 1995.
[4] G.C. Hudek and D.J. Muder. Signaling analysis for a multi-switch all-optical network. In Proceedings
of Int'l Conf. on Communication (ICC), pages 1206{1210, June 1995
[5] Myungsik Yoo, Chunming Qiao and Sudhir Dixit, “A comparative study of contention resolution
policies in optical burst switched WDM networks", Conf. on Terabit Optical Networking: Architecture,
Control and Management Issues, Boston, MA, Nov. 2000, SPIE Vol. 4213, pp.124-135.
[6] John Y. Wei, and Ray I. McFarland, Jr., “Just-In-Time Signaling for WDM Optical Burst Switching
Networks”, IEEE Journal of Lightwave Technology, Vol. 18, No. 12, Dec. 2000.
[7] Michael Düser, and Polina Bayvel, “Analysis of a Dynamically Wavelength-Routed Optical Burst
Switched Network Architecture”, Journal of Lightwave Technology, vol. 20, no. 4, April 2002
[8] Vinod Vokkarane, Qiong Zhang , Jason Jue, and Biao Chen, " Generalized Burst Assembly and
Scheduling Techniques for QoS Support in Optical Burst-Switched Networks," Proceedings, IEEE
Globecom 2002, Taipei, Taiwan, Nov. 2002.
[9] Shun Yao and S. J. Ben Yoo, Biswanath Mukherjee, “A comparison study between slotted and
unslotted all-optical packet-switched network with priority-based routing”, Proc. Of OFC 2001.
[10] J. Ramamirtham, J. Turner, “Time Sliced Optical Burst Switching”, Proc. of IEEE INFOCOM 2003,
San Francisco USA, March 30-April 3, 2003.
[11] J.D.Angelopoulos, G.C. Boukis, I.S.Venieris, "Delay priorities enhance utilization of ATM PON Access
Systems", Computer Communications Journal, Elsevier, Vol. 20, No. , December 1997, pp. 937-949.
[12] Ch. Linardakis, H-C. Leligou, A. Salis, J.D. Angelopoulos, “Control of slotted traffic among
interconnected WDM rings by explicit reservation”, Photonics in Switching 03, September 03,
Versailles, France.