52 (IJCNS) International Journal of Computer and Network Security,

Vol. 2, No. 10, 2010

Semi-deflection Routing: A Non-minimal

Fully-adaptive Routing for Virtual Cut-through
Switching Network
Yuri Nishikawa1, Michihiro Koibuchi2, Hiroki Matsutani3 and Hideharu Amano1
1
Keio University,
3-14-1 Hiyoshi Kouhoku-ku Yokohama, Kanagawa 223-8522, Japan
{nisikawa,hunga}@am.ics.keio.ac.jp
2
National Institute of Informatics,
2-1-2, Hitotsubashi Chiyoda-ku Tokyo 101-8430, Japan
3
The University of Tokyo
7-3-1 Hongo Bunkyo-ku Tokyo 113-8656, Japan

it does not guarantee in-order packet delivery, which is

Abstract: In this paper, we propose a routing methodology
called “Semi-deflection” routing, a non-minimal fully adaptive required for some message-passing libraries, and the logic to
deadlock-free routing mechanism for System area networks dynamically select a single channel from among a set of
(SANs), which usually employ virtual cut-through switching. We alternatives might substantially increase the switch’s
present that by adding a simple turn-model-based restriction to complexity. However, simple sorting mechanisms for out-of-
classical deflection (or hot-potato) routing, which is to allow order packets in network interfaces have been researched
non-blocking transfer between specific pairs of routers, Semi- [10][11], and real parallel machines, such as the Cray T3E
deflection routing mechanism guarantees deadlock and livelock- [12], the Reliable Router [13], or BlueGene/L[14] have
free packet transfer without use of virtual channels in a two- shown the feasibility of adaptive routing. A simple method
dimensional Mesh topology. As the result of throughput to support adaptivity in InfiniBand switches has also been
evaluation using batch and permutation traffic patterns, Semi- proposed [15]. We thus consider that these switches and
deflection routing improved throughput by maximum of 2.38
routers will employ adaptive routing not only in
times compared with that of North-last turn model, which is a
typical adaptive routing. The routing also reduces average hop
interconnection networks of massively parallel computers
count compared to deflection routing. but also in SANs.
There are two basic approaches to avoid deadlocks in
Keywords: SAN, virtual cut-through, non-minimal fully adaptive routing. The simpler strategy removes cyclic
adaptive routing, deflection routing, turn-model channel dependencies in the channel dependency graph
(CDG) [6][7][8]. The more complex ones deal with cyclic
1. Introduction channel dependencies by introducing escape paths [9]. It is
Network-based parallel processing using system area difficult to apply the latter strategy to interconnection
networks (SANs) has been researched as potential cost- networks with no virtual channels.
effective parallel-computing environments [1][2][3], as well In general, the most flexible routing algorithm is non-
as traditional massively parallel computers. SANs, which minimal fully-adaptive routing, because it allows each
consist of switches connected with point-to-point links, router to maximize the number of alternative paths. Non-
usually provide low-latency high-bandwidth minimal fully-adaptive routing supports dynamic selection
communications. Modern SANs and interconnection of various paths among multiple routers. The challenge of
networks of massively parallel computers usually use virtual such algorithm is to guarantee deadlock-free. Fully-adaptive
cut-through [4] as their switching technique, and they routing algorithms usually introduce virtual channels to
achieve reliable communications at the hardware level with guarantee deadlock freedom of packet transfers. For
deadlock-free routing. example, a minimal fully-adaptive routing called Duato’s
The following two strategies can be taken when a protocol uses two virtual channels for Mesh topologies, and
deadlock free routing algorithm is designed. Deterministic three for Torus topologies. However, virtual channels
routing takes a single path between hosts, and it guarantees require addition of channel buffers for each router ports, and
in-order packet delivery between the same pair of hosts [5]. they are sometimes used for providing different quality of
On the other hand, adaptive routing [6][7][8][9] services (QoS) classes. In addition, they are not always
dynamically selects the route of a packet in order to make supported in SANs.
the best use of bandwidth in interconnection networks. In In this paper, we propose a non-minimal fully-adaptive
adaptive routing, when a packet encounters a faulty or routing called Semi-deflection routing which does not
congested path, another bypassing path can be selected. require virtual channels in virtual cut-through networks.
Since this allows for a better balance of network traffic, Semi-deflection routing allows non-blocking transfer among
adaptive routing improves throughput and latency. certain routers in order to guarantee deadlock and livelock
In spite of the adaptive routing’s advantages, most freedom. Since non-blocking transfer means that a packet
current SANs do not always employ it[1][2]. This is because that arrived in a router must secure an output port in
 (IJCNS) International Journal of Computer and Network Security, 53
Vol. 2, No. 10, 2010

minimum transfer time, it cannot necessarily choose the northward movement. Turn model allows non-minimal path
shortest path to the destination. This is a similar approach to destination nodes, but certain path cannot be taken due to
with deflection (hot-potato) routing[16][17][18] and chaos prohibited turns.
3.10 Deadlock removal
This section describes the mechanism of Semi-deflection
routing. This allows prohibited turns of turn model under
certain conditions, while it guarantees deadlock-free. We
focus on the fact that deadlocks do not occur as long as
packets are moved by non-blocking on every prohibited turn
in virtual cut-through switching, because each packet can
Figure 1. Prohibited turn of North-last turn model independently detour blocked packets. First, we define “non-
blocking packet transfer” as follows.

Definition 1: (Non-blocking packet transfer) A packet

transfer is called “non-blocking” when a packet at a certain
input port is allocated to an output port via router’s crossbar
switch, given priority by the arbiter, and transferred to the
selected output port without being blocked by other packets.

Even when a prohibited turn is taken, deadlock does not

occur as long as it is guaranteed to transfer packets to
adjacent routers in a non-blocking manner. Thus, fully
deadlock-free routing is satisfied by always transferring
Figure 2. Example of non-minimal routing packets that intend to take prohibited turns in a non-
blocking manner.
Figure 2 shows pairs of input and output ports where non-
router[19] where all incoming packets are moved to output blocking transfer needs to be guaranteed in case of North-
ports at every cycle as non-blocking which may introduce last turn model. The non-blocking between these ports can
non-minimal paths in order to avoid packet blocking. These be attained because each packets can independently detour
studies already found that their proposed routing is themselves when virtual cut-through routing is applied. In
deadlock-free and livelock-free. Unlike these existing works, precise, we give a limitation when packets cannot take
some packets wait for the appropriate ports to their prohibited turns as follows. Assume that an incoming packet
destinations in Semi-deflection routing when other packets is expected to take a path along a prohibited turn, but its
occupy them. output channel is locked by another packet. In this case, the
The rest of this paper is organized as follows. In Section packet cannot wait only for that output port to become
2 we propose Semi-deflection routing. Section 3 gives vacant. In other words, the packet must be transferred to
evaluation results, and Section 4 gives related works. Finally another output channel if it is vacant even when that
Section 5 concludes the paper. direction is not towards to the destination. This is a similar
idea to deflection routing [20], which requires all packets to
2. Semi-deflection Routing be moving around the network constantly at a router at
This section proposes the mechanism of Semi-deflection every cycle.
routing. This routing avoids deadlocks using the approach For example, let’s say that the gray-colored node in
of Turn-model, and it requires to update output selection Figure 2 receives a packet from the south input port, and
functions and arbitration mechanism of routers as described applied turn model is North-last model. Since selection of
below. west, east or south output port makes prohibited turn, the
packet must choose north output port if the rest of the ports
3.9 Turn model are locked (occupied) by others. Only in case when North
Turn model is a family of adaptive routing which supports output port is locked, a packet that came from a South input
deadlock-free by removing all cyclic channel dependencies port packet waits for certain output port to become vacant.
in a network topology [6]. North-last model and West-first This is because not prohibited output ports behave as escape
model are representative examples. paths that prevents cyclic channel dependency. Idea of
Figure 1 shows North-last turn model in a two- escape path was originally suggested in adaptive routing
dimensional Mesh topology. A router has four input/output algorithms such as Duato’s protocol [9] where they are
ports in each direction (north, south, east and west), and implemented by using virtual channels.
there are eight possible turns. A turn model prohibits a set Black-colored input ports in Figure 2 indicate that any
of certain turns (which are called “prohibited turns” in this available output port selection becomes a packet transfer
paper) in order to remove cyclic channel dependency, and along the prohibited turns. Thus, packets from these input
they are southwest turn and southeast turn in case of North- ports must be transferred somewhere in a non-blocking
last model. In other words, packets must arrive at a manner.
destination node by finally taking 0-hops or more of Next, we define a term “Non-waiting port” as follows.
54 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 10, 2010

Definition 2: (Non-waiting port) An input port and its 3.3.2 Output selection function
pair connected with a link in the neighboring node is Adaptive routing is characterized by determining adaptive
defined as “non-waiting port” when either of following is routing algorithm and output selection function (OSF).
satisfied: (a) when any available output port selection Adaptive algorithm decides a set of available output ports,
becomes a packet transfer along the prohibited turns, or (b) and OSF prioritizes output ports for assigning to packets.
a port is paired with port defined in (a).
Theorem 1: There is one non-waiting port per router at Definition 4: (Output Selection Function (OSF) of Semi-
most when network topology is two-dimensional Mesh (N × deflection routing) OSF of Semi-deflection routing
N) and applied turn model is either North-last or West-first prioritizes each port based on the following order:
model. 1) Output towards the destination
Proof: Packets that enter a router from a South input port 2) Output that goes away from the destination, but does not
can only take a prohibited turn for North-last model. loop back
Consequently, South input ports have only possibilities to be 3) Output that loops back
non-waiting ports according to Definition 2 (a). On the
other hand, for Definition 2 (b), since there is only one input It is possible to prevent livelock between non-waiting ports
port that can be paired with (a), Theorem 1 is obviously by giving loop-back selection the lowest priority.
satisfied. For West-first routing, packets that go through
routers at coordinate x = N− 1 can only take prohibited turn.
Here, there is an output port which does not make prohibited 2.3.2 Arbitration
turn either in x or y direction in case of y ≠ 0, y ≠ N − 1. An arbiter is in charge of allotting input packets to output
Condition of Definition 2 (a) is satisfied either when x = N ports. Priority of arbitration is described as follows. If none
− 1, y = 0 or x = N − 1, y = N − 1 are satisfied. In the of the rules match, the order is based on the timestamp
former case, condition is satisfied for North input port, and when packets entered the router.
South input port for the latter case. On the other hand, for
Definition 2 (b), since there is only one input port that can Definition 5: (Arbitration order of Semi-deflection
be paired with (a), satisfaction of Theorem 1 is obvious. routing)
■ 1) Packets that were injected from non-waiting input port
2) Packets whose highest prioritized output port by the OSF
Definition 3: (Loopback transfer) When a packet transfers makes prohibited turn
directly from a router’s port to adjacent router’s port which 3) Other packets
the packet was transferred from, it is called “loopback
transfer”. It is possible to prevent livelock by arbitrating (1) with the
Theorem 2: When there is only one non-waiting port, a highest priority. Also, non-blocking loop-back transfer is
packet does not stay at non-waiting port after a certain realized when other output ports are busy. By arbitrating (2)
period of time when following conditions are satisfied: (a) with high priority, non-blocking transfer to the prohibited
loopback transfer is permitted, and (b) highest arbitration direction is satisfied when the output port is available. If it is
priority is given to the non-waiting port. blocked, other vacant ports are allotted in prior to other
Proof: When buffer of forwarding input port is vacant, a input packets.
packet at a non-waiting port is transferred to one of vacant Finally, the arbiter selects other packets that are not
ports. According to Definition 2, a non-waiting port always injected from non-waiting port, or packets that does not
exists on a router that a packet is transferred back to in a wish to take prohibited turn. If only available output port
loopback manner. If any input port of forwarding router is would keep packets away from destination nodes, packets
unavailable, a packet is transferred in a loopback manner. can wait at the router input buffer until other ports to
Even when a packet exists in forwarding non-waiting input become vacant. This suppresses unprofitable increase of hop
port, packets are simply exchanged between pairing non- counts.
waiting ports when highest arbitration priority is given to
these ports. Since these packet transfer and exchange are
2.1 Router structure
done within a certain period, Theorem 2 is satisfied.
■ Based on above studies, routing which adopts the following
rules is defined as Semi-deflection routing mechanism.
Non-waiting ports in Figure 2 is marked with black color
or diagonal lines. Livelock removal of non-waiting ports is Definition 6: (Semi-deflection routing mechanism)
described in the following section. l Prohibited turns and non-waiting ports are determined
based on a turn model in the given network.
l Each router transfers packets based on priorities given by
3.11 Router structure OSFs and arbiters in the previous subsections.
This section describes the modification of router’s output
selection function and arbiters, which are necessary to Because turn model can be applied to arbitrary topology
support Semi-deflection routing mechanism. [21], application of Semi-deflection routing is not limited to
a two-dimensional Mesh topology.
 (IJCNS) International Journal of Computer and Network Security, 55
Vol. 2, No. 10, 2010

Theorem 3: Semi-deflection routing is deadlock-free. ports, no deadlock occurs unless a packet transfer takes
Proof: (i) If a packet does not take a prohibited turn, it prohibited turn. Thus, buffers of forwarding input port will
does not deadlock because no cyclic-dependency occurs. (ii) be vacant according to (i), and packet transfer will be
In case of taking prohibited turn, a packet can be transferred possible. According to the above, only possible condition for
if buffers among forwarding input ports are vacant. (iii) causing a deadlock is when all packet transfer to forwarding
Even when packets exist in all buffer of forwarding input input ports to take prohibited turns, and packets

Figure 3. Uniform (1-flit) Figure 4. Matrix transpose (1-flit) Figure 5. Bit reversal (1-flit)

Figure 6. Uniform (4-flit) Figure 7. Matrix transpose (4-flit) Figure 8. Bit reversal (4-flit)

Figure 9. Uniform (8-flit) Figure 10. Matrix transpose (8-flit) Figure 11. Bit reversal (8-flit)

also exist in their buffers. However, such input ports

fulfilling the above conditions are non-waiting ports 3. Evaluations
according to Definition 2, and a packet does not stay after a
certain period of time according to Theorem 1 and This section shows throughput evaluation results of Semi-
Definition 5. Consequently, Semi-deflection routing is deflection routing for 8x8 Mesh topology with various
deadlock-free. ■ packet lengths.
3.1 Simulation conditions
Theorem 4: Semi-deflection routing is livelock-free.
The throughput was evaluated with 64 nodes in a two-
Proof: According to Definition 4 and Definition 5, an
dimensional Mesh topology using irr sim[22], a C++ flit-
output port toward the destination is given higher priority.
level network simulator. Here, latency is number of cycles
Also, Semi-deflection routing is a fully-adaptive non-
between source core injects a packet in the network and
minimal routing which assumes virtual cut-through routing.
destination core receives it. Accepted traffic is the average
The above two conditions satisfy conditions for a livelock-
number of flits each core receives at a cycle in average.
freedom of a chaos-router [19]. Consequently, Semi-
Maximum throughput is defined as maximum value of
deflection routing is livelock-free. ■
accepted traffic when latency is 1000 cycles or lower. For
evaluation, we used following batch and permutation
routing problems [23].
56 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 10, 2010

advantage of Semi-deflection routing using uniform traffic

l Batch routing: all processors generate a single packet pattern was moderate. Maximum throughput was lower than
simultaneously. North-last routing when packet length was four, was equal
Ø Uniform traffic when packet length was eight, and slightly better for single-
All processors generate a single packet to identically flit packet.
distributed random destinations. Traffic loads of each Comparing with classical Hot-potato routing, maximum
node are equally balanced. throughput of Semi-deflection routing was equal or
l Permutation routing: each node is the destination of a outperformed according to Figures 3, 4, and 5. This is a
single packet striking result because Hot-potato routing intuitively
Ø Matrix transpose provides higher flexibility compared to Semi-deflection
When array size is k, node (x, y) transmits data to routing. In addition, throughput of Hot-potato routing
node (k - y - 1, k - x -1). Nodes on the diagonal axis decreased when network load surpassed certain amount,
(x + y = k - 1) transmit data to node (k - x -1, k - y - indicating that it requires stricter injection limitation
1). technique which presumably would reduce maximum
Ø Bit reversal traffic throughput of Hot-potato routing.
When index values of source nodes are given as (a 0, a 1, ...,
a n− 1), each sends data to nodes (a n-1, ..., a 1, a 0 ). 3.3 Average hop count
Table 1. Average hop count (8x8 Mesh) Table 1 shows average hop count of each traffic pattern by
applying North-last, Semi-deflection, and Hot-potato routing
Traffic pattern Workload level to 8×8 two-dimensional Mesh topology. Again, here we
Routing assume a single-flit packet. Traffic load was set to the
low moderate very high
Uniform 5.34 5.36 5.29 following.
North Mtx. 6.25 5.60 5.06 l Low traffic: when packets are injected per 50 cycles to
-last transpose all routers.
Bit reversal 6.25 5.67 5.78 l Moderate traffic: when R is round-off value of average
Uniform 5.37 5.51 5.81 hop count in case of low traffic, packets are injected
Semi per R cycles to all routers. In concrete, R = 5 was
Mtx. 6.39 6.61 7.22
-Deflec applied for uniform traffic, and R = 6 for other traffic
transpose
-tion patterns.
Bit reversal 6.34 6.37 7.31
l Very high traffic: when packets were injected per one
Uniform 5.63 6.46 5.72
Deflec cycle to all routers.
Mtx. tranpose 6.25 7.84 5.14
-tion As Table 1 indicates, Semi-deflection routing effectively
Bit reversal 6.84 8.07 6.70
suppresses the increase of hop counts. Semi-deflection and
the other routing algorithms show contrasting
3.2 Throughput characteristics as traffic load increases. The case of Semi-
Throughput of Semi-deflection routing is shown in Figure 3 deflection routing is intuitive; when packets have lower
to Figure 11. The throughput was compared with North-last chances to take minimal path, they take detour route and
adaptive routing, and exclusively for 1-flit packets, we also average hop count increases.
made comparison with the most classical Hot-potato routing Contrary, hop count decreases for North-last routing and
algorithm, where input packets are deflected from the router Hot-potato routing. This is due to difference in packet
at the next cycle, and cannot wait for output ports to take delivery capacity of both routing mechanisms. When traffic
shortest path. Other parameters are specified as follows. becomes closer to the maximum capacity, packet movement
l Topology: 8×8 two-dimensional Mesh slows down and eventually terminates. Thus, hop count
l Buffer size: 16 flits indicates the average value when traffic was below
l Packet size: 1, 4, and 8 flits saturation. Since Semi- deflection routing provides escape
l Throttle threshold value: 3 ports paths, packet movement does not slow down, and this leads
To maintain performance of Semi-deflection routing and to the higher throughput as shown in previous section.
Hot-potato routing, we applied a very simple throttle
mechanism to each router; when more than three input ports
of a router are busy, its local host core does not inject a 4. Related Works
packet. Throughput evaluation with different packet length
Deterministic routing is a routing that always selects a
and traffic patterns revealed following features of Semi-
single path between each pair of source and destination
deflection routing. First of all, comparing the routing with
nodes. A most typical example is dimension-order routing,
North-last methodology, it performed quite well on
which routes packets first in the x-dimension, and then in
permutation routing patterns even with long packet length.
the y-dimension to reach the destination node. Although this
With any packet length, Semi-deflection routing provided
routing is advantageous in terms of average hop count
nearly twice or even higher throughput, as Figures 4, 5, 7, 8,
because it only allows minimal path, it may cause imbalance
10 and 11 indicate. According to Figure 7, approximately
for some traffic patterns. However, because of its simplicity
2.38 times higher throughput was obtained when packet
and feasibility to avoid deadlocks, dimension-order routing
length was one flit and traffic pattern was Matrix transpose.
was applied to many supercomputers such as Cray T3D[24].
On the other hand, according to Figures 3, 6, and 9,
 (IJCNS) International Journal of Computer and Network Security, 57
Vol. 2, No. 10, 2010

On the other hand, an adaptive routing can select blocking transfer among certain routers in interconnection
multiple possible paths, and make much of path diversity of networks of massively parallel computers and SANs.
the topology. The main challenge of adaptive routing is to Evaluation results show that throughput improvement
provide a large number of alternative paths so as to maintain was 2.38 times in maximum in case of 8×8 two-dimensional
average network load constant, balance the load among Mesh topology.
multiple routers, and keep average hop count low as much As the future work, we are planning to test our proposed
as possible. Another challenge is to guarantee deadlock- and routing with other topologies such as torus topology, and
livelock-freedom. A turn model described in Section 2 study other tuning techniques.
defines prohibited turns to eliminate all cyclic channel
dependency of the topology. Duato’s protocol is an adaptive
routing which makes the best use of path diversity, and is a References
fully-adaptive routing. It allows all possible paths by adding
one virtual channel for each path. [1] N.J.Boden and et al., “Myrinet: A Gigabit-per-Second
In contrast, Semi-deflection differs from the above in Local Area Network,” IEEE Micro, vol. 15, no. 1, pp.
terms of the following features. 29–35, 1995.
l Router design becomes lightweight because it does not [2] I.T.Association, “Infiniband architecture. specification
require virtual channels. volume1,release 1.0.a,” available at the InfiniBand
l It is a non-minimal fully-adaptive routing that has Trade Association, http://www.infinibandta.com, Jun.
largest path diversity. 2001.
Another approach to cope with deadlocks is deadlock- [3] F.Petrini, W. Feng, A.Hoisie, S.Coll, and
recovery-based routing, which usually employs minimal E.Frachtenberg, “The Quadrics network: high-
fully adaptive routing. It is useful only when deadlocks are performance clustering technology,” IEEE Micro, vol.
infrequent, and recently, the techniques that are applicable 22, no. 1, pp. 46–57, 2002.
for efficient deadlock recovery-based routing have been [4] P. Kermani and L. Kleinrock, “Virtual cut-through: A
proposed. When deadlock is found, one of the packets is new computer communication switching techniques,”
removed from the deadlocked paths. In overall, there are Computer Networks, vol. 3, no. 4, pp. 267–286, 1979.
two types in deadlock removal strategies: a progressive [5] W. Dally and C. Seitz, “Deadlock-Free Message Routing
method that prepares escape path, and regressive method in Multiprocessor Interconnection Networks,” IEEE
which discards and retransmits packets. The former requires Transaction on Computers, vol. 36, no. 5, pp. 547–
virtual channels, and the latter requires control mechanics 553, May 1987.
for packet disposal and retransmission. Semi-deflection [6] C. J. Glass and L. M. Ni, “The Turn Model for Adaptive
routing is a deadlock-free mechanism, so it does not require Routing,” Proceedings of International Symposium on
such control mechanisms. Computer Architecture, pp. 278–287, 1992.
Also, some methods handle deadlocks more proactively. [7] W. J. Dally and H. Aoki, “Deadlock-Free Adaptive
For example, Deflection routing guarantees deadlock-free by Routing in Multicomputer Networks Using Virtual
making a router sending out larger number of flits than it Channels,” IEEE Transaction on Parallel and
has received per cycle. Of course, each router cannot Distributed Systems, vol. 4, no. 4, pp. 466–475, 1993.
necessarily selects proper output port of each packet to the [8] A.A.Chien and J.H.Kim, “Planar-adaptive routing: low-
destinations, however deadlock does not occur because cost adaptive networks for multiprocessors,” Journal
packets do not make collisions. Also, studies of Chaos router of the ACM, vol. 42, no. 1, pp. 91–123, Jan. 1995.
have proved its livelock-freedom. Drawback of Deflection [9] J. Duato, “A Necessary And Sufficient Condition For
routing is that it cannot be applied to wormhole routing, Deadlock-Free Adaptive Routing In Wormhole
increasing number of average hop count, and increasing Networks,” IEEE Transaction on Parallel and
hardware amount due to addition of dedicated buffer for Distributed Systems, vol. 6, no. 10, pp. 1055–1067,
holding one packet per router[25]. Semi-deflection routing 1995.
differs from Deflection routing in terms of the following. [10] J.C.Martinez, J.Flich, A.Robles, P.Lopez, J.Duato, and
l Semi-deflection routing only requires non-blocking M.Koibuchi, “In-Order Packet Delivery in
transfer for selected ports of partial routers. Other Interconnection Networks using Adaptive Routing,” in
packets can wait for output ports to take appropriate Proceedings of IEEE International Parallel and
paths. Thus, average hop count is smaller than Distributed Processing Symposium, Apr. 2005, p.
Deflection routing, as shown in Section 3. 101a.
l No dedicated buffer is required for implementation. [11] M.Koibuchi, J.C.Martinez, J.Flich, A.Robles, P.Lopez,
and J.Duato, “Enforcing In-Order Packet Delivery in
System Area Networks with Adaptive Routing,”
5. Conclusions Journal of Parallel and Distributed Computing, vol.
65, pp. 1223–1236, Oct. 2005.
In this paper, we proposed Semi-deflection routing, which is [12] S. L. Scott and G. T.Horson, “The Cray T3E network:
a non-minimal fully-adaptive routing and makes the best use adaptive routing in a high performance 3D torus,” in
of virtual cut-through switching. Semi-deflection routing Proceedings of Hot Interconnects IV, Aug. 1996, pp.
does not require the use of virtual channels by allowing non- 147–156.
58 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 10, 2010

[13] W.J.Dally and et al., “Architecture and implementation Author’s Profile

of the reliable router,” in Proceedings of Hot
Interconnects Symposium II, Aug. 1994.
[14] P. Coteus, H. R. Bickford, T. M. Cipolla, P. G. Yuri Nishikawa recieved the B.E. and
M.E. degrees from Keio University, Japan,
Crumley, A. Gara, S. A. Hall, G. V. Kopcsay, A. P. in 2006 and 2008. She is currently a Ph.D.
Lanzetta, L. S. Mok, R. Rand, R. Swetz, T. Takken, P. candidate at Keio University. Her research
L. Rocca, C. Marroquin, P. R. Germann, and M. J. interests include the areas of
Jeanson, “Packaging the Blue Gene/L supercomputer,” interconnection networks and high-
IBM Journal of Research and Development, vol. 49, performance computing.
no. 2/3, pp. 213–248, Mar/May 2005.
[15] J. Martinez, J. Flich, A. Robles, P. Lopez, and J. Duato,
“Supporting Adaptive Routing in IBA Switches,”
Journal of Systems Architecture, vol. 49, pp. 441–449,
Michihiro Koibuchi received the B.E.,
2004.
M.E., and Ph.D degrees from Keio
[16] P. Baran, “On Distributed Communication Network,” University, Japan, in 2000, 2002, and
Communications Systems, IEEE Transactions on, 2003. He was Visiting Researcher in
1962. [Online]. Available: Technical University of Valencia, Spain,
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=108888 and Visiting Scholar in University of
3 Southern California, in 2004, and 2006. He
[17] C. Kaklamanis and S. Rao, “Hot-Potato Routing on is currently Assistant Professor in National
Processor Arrays,” in Fifth annual ACM symposium Institute of Informatics (NII) and the
on Parallel Algorithms and Architectures, Velen, Graduate University for Advanced Studies, Japan. His research
interests include the areas of high-performance computing and
Germany, 1993, pp. 273 – 282.
interconnection networks.
[18] T. Moscibroda and O. Mutlu, “A Case for Bufferless
Routing in On-Chip Networks,” in Proceedings of the
International Symposium on Computer Architecture Hiroki Matsutani received the B.A.,
(ISCA’09), June 2009. M.E., and Ph.D degrees from Keio
[19] S. Konstantinidou and L. Snyder, “The chaos router,” University, Japan, in 2004, 2006, and
IEEE Transactions on Computers, vol. 43, no. 12, pp. 2008. He is currently Cooperative
1386–1397, 1994. Researcher in the Research Center for
[20] E. Nilsson, “Design and Implementation of a Hot- Advanced Science and Technology
(RCAST), The University of Tokyo, Japan,
Potato Switch in Network On Chip,” Master’s thesis,
and Visiting Researcher in Keio
Laboratory of Electronics and Computer Systems, University. His research interests include the areas of Networks-
Royal Institute of Technology (KTH), June 2002. on-Chips and interconnection networks.
[21] A. Jouraku, M. Koibuchi, and H. Amano, “An Effective
Design of Deadlock-Free Routing Algorithms Based
on 2-D Turn Model for Irregular Networks,” IEEE Hideharu Amano received the Ph.D
Transactions on degree from Keio University, Japan, in
Parallel Distributed Systems, vol. 18, no. 3, pp. 320–333, 1986. He is currently Professor in the
2007. Department of Information and Computer
[22] H. Matsutani, M. Koibuchi, D. Wang, and H. Amano, Science, Keio University. His research
“Adding Slow-Silent Virtual Channels for Low-Power interests include the areas of parallel
processing and reconfigurable systems.
On-Chip Networks,” in Proceedings of the
International Symposium on Networks-on-Chip
(NOCS’08), Apr. 2008, pp. 23–32.
[23] W.J.Dally and B.Towles, Principles and Practices of
Interconnection Networks. Morgan Kaufmann, 2003.
[24] S. L. Scott and G. Thorson, “Optimized routing in the
Cray T3D,” in PCRCW ’94: Proceedings of the First
International Workshop on Parallel Computer Routing
and Communication. London, UK: Springer-Verlag,
1994, pp. 281–294.
[25] J. Duato, S. Yalamanchili, and L. Ni, Interconnection
Networks: an engineering approach. Morgan
Kaufmann, 2002.