MINIMALLY BUFFERED ROUTER USING WEIGHTED DEFLECTION ROUTING FOR MESH NETWORK ON CH

International Journal of VLSI design & Communication Systems (VLSICS) Vol.6, No.3, June 2015
DOI : 10.5121/vlsic.2015.6306 59
MINIMALLY BUFFERED ROUTER USING
WEIGHTED DEFLECTION ROUTING FOR
MESH NETWORK ON CHIP
Simi Zerine Sleeba and Mini M.G.
Department of Electronics Engineering, Government Model Engineering College,
Cochin University of Science and Technology
Kochi, India
ABSTRACT
The scalability, modularity and massive parallelism exhibited by Network on chip(NoC) interconnects
make them highly suitable for the inter core communication framework of multiprocessor system-on-chip
(MPSoC) designs. Routers play the most vital role in transferring flits through the network, hence efficient
microarchitecture and cost effective routing algorithms are highly essential for modern NoC routers.
Elimination of buffers and deflection routing help to achieve energy and area efficiency of these routers.
The advantages of bufferless and buffered designs can be combined by using a minimum number of side
buffers to store a fraction of deflection flits in the router. In this paper, we propose a routing algorithm
based on weighted deflection of flits for minimally buffered deflection routers. Evaluations on 4x4 and 8x8
mesh NoC using synthetic workloads as well as benchmark applications demonstrate that deflection rate
and average network latency are significantly reduced in comparison with the state of the art NoC routers.
Performance analysis of the newly proposed algorithm shows that the network saturation point improves
by 26% compared to earlier designs in this domain.
KEYWORDS
Network on Chip, Deflection routing, Minimal buffering, Average Latency
1. INTRODUCTION
As technology scaling reduces the feature size to nanolevels, a large number of intellectual
property cores are being integrated on to a single chip[1]. Such chips can execute applications
that demand extensive amounts of parallel processing. Hence there is an ever increasing demand
for on chip interconnects capable of handling huge amount of data. Traditional bus based
interconnects are no longer suitable for MPSoCs with hundreds and thousands of cores. NoCs
emerge as a promising design choice for realising efficient on chip interconnections as they
largely alleviate the limitations of bus based interconnects.
Generally, MPSoCs employ a regular mesh topology as in Figure 1. Each Processing Element
(PE) is connected to a local switching element called router(R) by means of a network interface.
Each PE is either a processor core (with built-in L1 cache) or a slice of shared L2 cache. Requests

60
for data transfer between processor and shared cache inject packets into the network. These
packets start from source node and traverse multiple routers and links before reaching the
destination node. Each router has five input ports and five output ports, four of them are
connected to neighbouring routers in the east, west, north and south directions. The fifth input
port accepts packets injected from the PE to the network and output port ejects packets destined
to the PE from the network. The packets arriving at the input ports of a router move through
various functional units and get access to output ports depending on their routing choice and
prioritisation.
Figure 1. Mesh Interconnection Topology Figure 2. MinBD router microarchitecture[8]
In the traditional virtual channel router (VCR), large number of buffers are employed at the input
ports so that flits belonging to different packet transmissions can proceed simultaneously through
the same physical channel[2][3]. Packets that cannot find a productive output port from the
router are temporarily buffered. Thus buffers prevent unnecessary wastage of link bandwidth and
increase saturation throughput of the network. But these buffers consume significant amount of
dynamic power when active and static power when idle[4][5]. Recent research findings indicate
that static power could constitute 80% to 90% of interconnect power in future systems[6][7].
Hence buffer size is an important parameter that affects area, power and performance of NoCs.
The buffered routers provide an overprovisioned buffer size to accomodate the worst case
network traffic. Real applications are found to have a low injection rate[8][9] compared to
synthetic traffic and in such situations, buffer utilisation is very less. Hence the concept of
bufferless and minimally buffered router in NoC are gaining acceptability as they can deliver
similar performance with a lesser power consumption. In these routers, packets are broken down
into smaller flow control units or flits which are independently routed. In bufferless deflection
routers, when more than one incoming flit competes for the same output port, one wins and
traverses through the desired output port, the others are assigned non productive ports ie. such
flits are deflected. So eventhough power and area due to buffers is minimum in bufferless
designs, the flit deflection rate is high and network saturates earlier compared to buffered routers.
Minimally buffered routers combine the advantages of buffered and bufferless designs by
buffering a small fraction of the deflection flits, achieving performance close to that of NoCs
with buffered routers. The area and power consumed by the side buffers are not significantly
higher than that of bufferless designs.

61
In this work, we propose a novel weighted deflection algorithm for output port selection in
minimally buffered routers for mesh NoCs. From evaluations, we observe that the proposed
algorithm outperforms the native routing method of MinBD in terms of reduced deflection rate
and average flit latency. Reduction in deflection rate leads to reduced dynamic power
consumption due to unproductive flit movement through the network.
The rest of the paper is organised as follows. Sections 2 and 3 give an overview of the existing
deflection routers employing minimal buffering and the motivation behind this work. Section 4
and 5 explain the Weighted Deflection routing algorithm and router pipeline in detail. Section 6
describes the evaluation method. We consolidate our experimental results in Section 7 and then
we conclude the paper.
2. RELATED WORK
First among the network on chip routers is the buffered virtual channel router that attains the
required network performance by providing large number of virtual channels and buffers [2] [3].
Bufferless routers are first proposed in [10][11]. Bufferless routing mechanism is either based on
dropping and retransmission of packets or on deflection of packets. SCARAB [12] uses the
policy of packet dropping which incurs a high retransmission overhead. BLESS [13] proposes a
bufferless router microarchitecture using deflection routing. In a deflection router, all flits
arriving at the input ports pass through one of the output ports from the router pipeline [14].
BLESS routers use sequential allocation of output ports to flits which are sorted on the basis of
age; this increases the critical path length inside the router and makes the router slow. A much
superior router microarchitecture is that of CHIPPER [15] which employs parallel port allocation
of flits. CHIPPER is faster compared to BLESS but exhibits higher flit deflection rate. In our
earlier work, WeDBless [16],we use a routing policy which reduces deflection rate significantly
by prioritising output ports for a flit based on Deflection Weights(DW) and ranking flits inside
the router based on Weighted Deflection Count (WDC).
MinBD [8] which is the first minimally buffered router design employs a small side buffer to
store a few number of deflection flits in the router. DEBAR [9] and SLIDER [17] also use
minimal buffering at the routers along with innovative methods for flit ejection, injection and
buffering. Another work using minimal buffering inside router is mentioned in [18]. The study of
various NoCs using deflection routing comes out with the opinion that a priority based deflection
policy which uses global or history related criteria is most suited for boosting the performance of
networks [19].
3. MOTIVATION
The MinBD router is efficient in terms of lower deflection rate compared to CHIPPER and
consumes lesser power and area compared to virtual channel router. In deflection routers, each
incoming flit is transferred to an output port or a side buffer in two cycles of operation ( we use
two cycle routers). Deadlock problem does not arise in deflection routers since cyclic dependency
of resources will not occur [13]. In bufferless routers, it is equally important to ensure livelock
freedom. Both CHIPPER and MinBD use the golden flit priority scheme to resolve livelock
problem. In this scheme, flits belonging to one packet are marked as the highest priority flits
(golden flits) in the entire network and it is guaranteed to win a productive output port at each
router until the packet finally reaches the destination and golden status is passed on to another

62
packet. This scheme is not very efficient because more that 90% of the flits are delivered without
becoming golden. The conflict for port allocation of non golden flits is resolved randomly.
MinBD uses an additional silver flit priority for prioritising a flit within a router. We analyse
certain drawbacks of MinBD which motivate us to propose a new router using weighted
deflection algorithm.
3.1. Problem of complex flit prioritisation scheme
The golden priority scheme requires that all routers in the network incorporate a functional unit
for golden flit identification while most of them donot recieve golden flits. Silver flit scheme is
also inefficient since it cannot ensure that high priority flits in one router get similar priority in
the succeeding routers. This leads to increased number of deflections per flit resulting in early
saturation of the network. The silver flit prioritisation unit also increases the critical path delay
inside router as can be seen from Figure 2.
3.2. Inefficiency of dimension order routing
MinBD uses the Permutation Deflection Network(PDN) proposed in CHIPPER [15] for
allocating output port to flits. Computation of desired output port is done by XY routing method
which computes only one productive output for a flit. In a mesh network, majority of the flits
have two productive output ports in a router. If one of the output ports chosen by the flit is in
demand by a higher priority flit, the lower priority flit can occupy the next productive output port.
But this capability of PDN is not utilised by the existing routing algorithm in MinBD. Hence we
need a routing algorithm which adaptively chooses output ports for flits based on local
congestion in the neighbouring routers and global congestion in the network.
3.3. Additional power and area consumption due to duplicate ejection units
As shown in Figure 2, MinBD router uses two ejection units placed one after another for ejecting
two locally destined flits in a cycle. This costs additional power and area and also increases the
critical path length inside the router.
In this work, we propose a Weighted Deflection routing algorithm for a minimally buffered
router which eliminates flit ranking based on complex global timing and provides local priority
for a flit inside the router based on the value of Weighted Deflection Level. Flits with higher
WDL win the arbitration for output port in a router. Choice of output port for a flit is based on
Weighted Distance to Destination which provides more than one productive routes for a flit in
majority of the cases. Dual ejection is facilitated through a single ejection port by providing an
Eject Buffer which buffers one ejection ready flit. The path delay inside the router caused by
duplicate ejection units and the additional wiring overhead is reduced in this design. We also
propose a simple logic that precomputes the productive routes of outgoing flits in the succeeding
router by adjusting the WDDs of the flit.
4. WEIGHTED DEFLECTION ROUTING ALGORITHM
In this routing method, flits in a router are ranked based on the value of Weighted Deflection
Level(WDL)and output ports for a flit are prioritised based on Weighted Distance to
Destination(WDD). WDL and WDD values are incorporated in the enhanced header of a flit.

63
4.1. Weighted Deflection Level
The WDL is 6 bits long and has an initial value of 0 when injected into the network. When a flit
makes a hop from one router to another, its WDL value changes. When the flit moves through a
productive output port, its WDL is decremented and when it is deflected through non productive
port, its WDL is incremented. So frequently deflected flits have higher WDL value compared to
less deflected ones. During arbitration for port allocation, if two flits compete for the same output
port, the one with higher WDL wins and moves through the desired port whereas the other flit is
deflected through the remaining output port. As the flit moves out of a router, its WDL is updated
with the new value . Since the WDL of deflected flits are incremented, these flits get higher
priority to win productive output ports in subsequent routers. The WDL field maintains an
account of the flit’s deflection and this priority scheme effectively replaces other methods based
on global timing information.
4.2. Weighted Distance to Destination
For a flit, the productive output ports of a router are those which move the flit towards its
destination router. Each flit has four WDDs, each of which are 2 bit values. They represent the
Figure 3. Example 4x4 Mesh Network Figure 4. Enhanced Flit Header
Figure 5: Maximum value of WDL for various flit injection rates for 8x8 mesh

64
relative distance of the flit’s destination from the current router in the North, South, East, West
directions. Each WDD can have a value of -1(least distance and highly preferred output port), +1
(second preference output port), +2 (non productive or least preferred output port). WDD of -1,+1
and +2 are coded as 00, 01 and 10 respectively. We use 8 bits in total to accomodate the four
WDDs in the enhanced flit header. When a flit is injected into the network the WDDs are
calculated based on the position of its destination and stored in its header. During the arbitration
for output port selection, each flit gets its routing preference from the WDD values in its header.
A flit tries to occupy an output port with the least WDD value. Consider the 4 x 4 mesh NoC
shown in Figure 3. For a flit at south input port of router(1,1) whose destination is at router(2,2),
north and east directions are equally productive, hence WDD values in these directions will be -1.
south and west output ports carry the flit away from the destination and hence WDD values in
those directions will be +2 . The 8 bit WDD values of this flit at router(1,1) will be North(00),
South(10), East(00), West(10). As a second example, if the flit destination is at router(1,2), only
the East output port has a WDD of -1, North and South WDDs are +1 and west WDD is +2.
Figure 6. Pipeline diagram of proposed router
When the flit leaves the current router, the flit header is updated with a new set of WDDs
corresponding to the next router. This is done in the route precomputation stage at the end of the
router pipeline. The WDL of a flit is updated by adding the WDD corresponding to the allocated
output port to it. In the first example mentioned above, if north or east output port is alloted to the
flit (WDD value is -1), its new WDL value is obtained by adding -1 to the current WDL value. If
south or west output port (WDD value is +2 ) is alloted, new WDL is obtained by adding +2 to
the current WDL value. As a flit gets deflected away from the destination, the WDL value
increases due to addition of larger WDDs. We maintain the value of WDL between 0 and 26 by
decrementing the value as the flit approaches destination and incrementing it as it goes away.
This priority scheme based on Weighted Deflection largely increases the probability of the older
flits reaching the destination earlier. Thus, the algorithm guarantees livelock freedom of flits
since increase in WDL raises the priority of deflected flits and prevents them from being
deflected continuously.
4.3. Enhanced Flit Header
In order to incorporate the WDL and WDDs, we use an enhanced header of 14 bits for each flit .
The header format is shown in Figure 4. Among the 14 bits; 8 bits represent the four WDDs and
6 bits represent WDL. In Figure 5, we plot the maximum values of WDL obtained for various flit

65
injection rates in an 8x8 mesh network. The WDL value increases from 0 to 38 as the injection
rate is increased. Hence a 6 bit field is sufficient to represent WDL in the flit header.
5. ROUTER PIPELINE
In the router pipeline shown in Figure 6, the incoming flits experience two cycle latency. At the
beginning and end of a cycle, flits are stored in pipeline registers which are shown by solid
vertical lines in Figure 6. One cycle latency is constituted by ejection unit, flit injection from side
buffers and injection from the local PE. The next cycle includes output port allocation using
Permutation Deflection Network(PDN), side buffering of deflection flits and route
precomputation. Function of each of these units is explained in detail.
5.1. Ejection Unit
The function of the Ejection Unit is to remove flits destined for the local core from the network.
The ejection unit consists of a flit identification circuit, an Eject Buffer (EB) and Eject
Multiplexer. We refer to the flits destined for the local core as Ejection Ready Flits (ERF). The
flit identification circuit identifies the ERF from among the flits arriving at the input ports of the
router. We simulate the network using real application traffic and observe that for 10% of the
operation cycles, more than one ERF occur. For very few cases, more than one ERF is present in
two consecutive cycles. In the newly proposed router, we use the dual ejection mechanism
mentioned in [9] and [16] where dual ejection is made possible using a single ejection port. If
more than one ERF is present, the one with the highest WDL qualifies for ejection through the
local port and one is buffered in the EB. The ERF in the Eject Buffer is ejected from the router in
the next cycle. In the next cycle, newly arrived ERF are considered for ejection only if EB is
empty. If more than two ERF are present, one is ejected, one is buffered and the others are
deflected. But such cases occur rarely as we see in evaluations. Buffering of one ERF in a cycle
substantially reduces deflections due to them.
5.2. Injection
There are two injection blocks placed one after another. One is for buffer injection and the other
is for injection from local core. In this router, some of the deflection flits are forced into a side
buffer after port allocation. The Buffer Injection unit re-injects these flits back into the router
pipeline in a subsequent clock cycle, when any of the flits in the four input channels are not
present. In a cycle, atmost one flit can be re-injected into the router pieline. A FIFO side buffer is
used in the routers ie. the flit that entered the side buffer first will be re-injected into the pipeline
first. The local PE can inject flits into the network only when at least one input channel is free
even after re injection from side buffer. In a cycle, atmost one flit can be injected from the local
core.
5.3. Permutation Deflection Network
We use the PDN structure mentioned in CHIPPER and MinBD for parallel output port allocation.
This PDN efficiently maps every input port to every output port of the router. The PDN consists
of four 2x2 Permutation Deflection Units (PDU1, PDU2, PDU3, PDU4) arranged in two stages,
two units per stage. In the first stage, the North and East input channels are connected to PDU1
and South and West input channels are connected to PDU2. In each permuter block, the flit with
higher WDL value has the priority to choose the desired output port and the flit with lower WDL
gets the other port. The winning flit chooses the output port with lesser WDD; the WDD values
are obtained from the flit header.

66
At the output of Permutation units PDU3 and PDU4, an output port is alloted to each flit. Most of
the flits have two output ports with the same WDD values. So even if a flit does not win an
arbitration in the first permutation stage, it can still acquire an equally desirable output port in the
second stage. For example, if a flit from the north input port is destined for a router in the south
east direction, both south and east output ports are equally productive for it. If the flit fails to get
to south output port at PDU1 due to lesser WDL value, it will be deflected to PU4 which is
connected to east and west output ports. In this block the flit has a chance of getting the east
output port which is also equally desirable for it. If the flit again loses in this arbitration and is
deflected through the west output port, its WDL value is incremented by + 2 (WDD of west
output port is +2 for this flit). Then the flit has an increased probability of winning the desired
output port in the next router.
5.4. Side Buffering of Deflection Flits
In minimally buffered router, after port allocation by the PDN, some of the flits donot win
productive output ports to their destinations. Deflection of these flits through the links is reduced
by moving some of them into a side buffer. We choose the side buffer size as 4. Atmost one flit
can be side buffered in a cycle. If the side buffer is full, flits in the subsequent cycles are
deflected.
Table 1. Precomputed WDD values for a flit at router (1,1)
Assigned Output Destination at (2,2) Destination at (3,3)
N S E W N S E W
North 1 1 -1 2 -1 2 -1 2
South -1 2 -1 2 -1 2 -1 2
East -1 2 1 1 -1 2 -1 2
West -1 2 -1 2 -1 2 -1 2
5.5. Route Precomputation Unit
The WDL and WDD values in the header of the outgoing flits are recomputed and updated in this
unit which is placed after the port allocation stage. The new WDD values correspond to preferred
output routes in the next router. Pre-computation of routes using this method is less complex
compared to that of MinBD where XY route computation is done for incoming flits. This can be
explained with the help of the example in Figure 3. If south or west output port ( WDD is +2) is
alloted to the sample flit at router (1,1), the flit is being moved further away from destination. So
in the south or west neighbouring router of (1,1), the WDD values have no change ie. they need
not be re-computed. Table 1 shows the precomputed WDD values for a flit at router(1,1) when
destination is at router(2,2) (in columns 2 to 5) and destination is at router(3,3) (in columns 6 to
9). Since the current WDD values of the flit in north, south, east and west directions are 0, +2, 0
and +2 respectively, WDDs need to be re-computed only when a flit having two productive
output ports is alloted one of them and destination is at two hops distance from the current router.
In all other cases, the new values of WDDs are equal to the previous values.
6. EXPERIMENTAL METHODOLOGY
We use the traditional cycle accurate NoC simulator, Booksim [20] which models the VC router
mentioned in [2]. We make suitable modifications on Booksim to model the newly proposed
minimally buffered router using weighted deflection (abbreviated as MinBWD in the evaluation
section) as mentioned in Section 4. We enhance the flit header with 14 bits for representing WDD

67
(8 bits) and WDL values ( 6 bits). We also model the CHIPPER and MinBD architectures as
mentioned in the literature and use MinBD as a baseline in our evaluations. The router delay is
two cycles in all the three architectures. We conduct all evaluations using single flit packets.
6.1. Synthetic Workloads
We use synthetic traffic patterns to analyse the robustness of our algorithm. Real applications
exhibit self throttling, hence synthetic traffic is best suited to test the network saturation point of
an algorithm. Uniform random traffic is used to assess the adaptability and load balancing
capability of the routing algorithm where as patterns like transpose, bit-complement and tornado
are network intensive. Using these synthetic workloads, we conduct simulations on 4x4 and 8x8
mesh network. After sufficient warm up time, we collect the average flit latency, deflection rate
and throughput values by varying the flit injection rates from 0 to network saturation point.
Table 2. Percentage of different network injection intensity applications in various benchmark mixes.
Benchmark Mix M1 M2 M3 M4 M5 M6
% of Low 100 0 0 50 0 31
% of Medium 0 100 0 0 50 31
% of High 0 0 100 50 50 38
Figure 7. Average flit latency comparison under various synthetic traffic patterns in 8 × 8 mesh network.
Figure 8: Average flit deflection comparison under various synthetic traffic patterns
in 8 × 8 mesh network
6.2. Real Applications
We simulate a 64 (8x8) core multiprocessor system using Multi2sim [21]. Each core has an out-
of-order x86 processor with 64KB 4-way set associative L1 cache and 512KB 16-way set
associative shared L2 cache. Each core runs one of the applications from SPEC CPU 2006
benchmark application suite [22]. Based on the misses per kilo instructions(MPKI) from L1
cache, these applications are classified as low, medium or high MPKI . We run low MPKI
applications like calculix, gombk and h264ref; medium MPKI applications like bwaves,

68
bzip2,gcc and high MPKI applications such as hmmer, leslie3d and matlab in our simulations.
We generate 6 workload mixes by combining applications from the suite as given in Table 2. The
network packets generated by these workloads are used in our NoC simulator and average
deflection rate and average latency of MinBWD and MinBD are compared.
7. EXPERIMENTAL ANALYSIS
We compare the average deflection rate of MinBWD with that of MinBD for synthetic and real
workloads to analyse the role of weighted deflection algorithm in reducing deflections. We also
compare the average flit latencies of MinBWD with MinBD as well as a bufferless router,
CHIPPER. We analyse average latencies for 4x4 and 8x8 mesh inorder to study the variation in
performance with network scaling. Various performance parameters like average deflection rate,
average latency and average throughput are analysed in detail.
Figure 9. Average flit latency comparison under various synthetic traffic patterns
in 4 × 4 mesh network.
Figure 10: Average deflection rate and latency for real applications.
7.1. Effect on Average Deflection Rate
Deflection of flits through the network causes unnecessary dissipation of dynamic power. The
aim of our algorithm is to minimise these deflections and achieve energy efficiency for the NoC.
Average deflection rate is computed as the average number of deflections encountered per flit.
The average deflection plot for three typical synthetic traffic patterns ie., uniform-random,
transpose and bit complement are shown in Figure 8. MinBWD reduces the average deflection
rate by 56% compared to MinBD for uniform-random traffic. For non-uniform traffic distribution

69
like transpose and bit complement, MinBWD reduces deflection rate by 33% and 65%
respectively. Unique features of the proposed algorithm which lead to reduction in deflection rate
are: (1) providing more than one productive paths for a flit (2) incrementing WDL value for
deflected flits giving them high priority to win arbitration in successive routers (3)providing Eject
Buffer to store ERF for dual flit ejection. From evaluations using real applications as shown in
10, MinBWD shows a maximum of 36% reduction in deflection rate which is obtained for work-
loads with a larger share of high MPKI applications. This explains that the effectiveness of the
MinBWD algorithm is more evident under higher network load.
7.2. Average Flit Latency
It is obvious that the reduction in deflection rate obtained for MinBWD reflects in the average flit
latency also. Flit latency is measured as the average number of cycles taken by the flit from the
time it is generated at the PE upto its ejection at the destination core. A lower value of latency
means that the data flits traverse the network in lesser number of cycles and this results in
application speed up. We plot the average latencies for synthetic traffic patterns for 8x8 and 4x4
mesh in Figure 7 and Figure 9 respectively. For uniform traffic, the adaptivity and reduced
deflections of MinBWD improves the network saturation point by 26% compared to MinBD.
Figure 11: Average throughput comparison under various synthetic traffic patterns
in 8 × 8 mesh network.
7.3. Average Throughput
The throughput is computed as the number of flits ejected from the network per router per cycle.
Average throughput variations for various synthetic traffic patterns is shown in Figure 11. At
lower injection rates, the throughput delivered by MinBWD and MinBD are similar. MinBD
reaches saturation earlier and a dip in throughput is observed for injection rates higher than
saturation point.The average throughput of MinBWD is higher for all synthetic workloads and a
steady value is maintained further beyond saturation.
8. ROUTER TIMING, POWER AND AREA
We implement each of the functional units of the minimally buffered router using Verilog HDL
and synthesise using an EDA tool. We also model additional control logic required for weighted
deflection algorithm and compare the overall timing latency of MinBWD and MinBD router
architectures. Since MinBWD uses dual ejection with Eject Buffer, the first pipeline stage has
37% gain in timing latency compared to MinBD which uses two ejection Units. Silver flit
prioritisation stage of MinBD is not used in our design and the timing delay of PDN in both
MinBD and MinBWD is the same. The delay of the Route precomputation unit at the end of the
second pipeline stage in MinBWD is lesser than that of silver flit priority block of MinBD, hence
we conclude that the timing latency of the second pipeline stage is same for MinBWD and

70
MinBD. Among the two pipeline stages, second stage dominates the timing latency and hence it
decides the router’s operating frequency.
We use Orion 2.0 tool[23] for modelling the power and area consumed by MinBWD as well as
MinBD. We choose the operating frequency as 1GHz, technology parameter as 65nm and link
delay as one cycle. The enhanced flit header in MinBWD uses additional 14 bits for transmitting
WDL and WDD values. Hence static power and wiring area at each of the links increases by 10%
compared to MinBD. The scheme for dual ejection using single ejection port reduces channel
wiring overhead in MinBWD by 26%.
9. CONCLUSION
In this paper we propose a novel routing algorithm for minimally buffered deflection routers
which reduces the flit deflection rate and average latency of flits using a unique mechanism of
Weighted Deflection. Flits are prioritized using Weighted Deflection Level which is directly
proportional to the amount of deflections encountered by the flit. Route computation for a flit
based on Weighted Distance to Destination and predetermination of these values for a router
contribute significantly to adaptive routing and reducing critical path lengths. From this work, we
conclude that the new algorithm combined with minimal buffering for deflection routers promise
better performance for NoCs in terms of higher network saturation points and lower latencies.
ACKNOWLEDGEMENTS
This work is supported in part by grant from UGC under MOMA-MANF scheme.
REFERENCES
[1] S. S. I. Association, International Technology Roadmap for Semiconductors Interconnect, Technical
Report, 2011 (http://public.itrs.net/).
[2] W. Dally, B. Towles, Principles and Practices of Interconnection Networks, Morgan Kaufmann
Publishers Inc., USA, 2003.
[3] W. Dally, Virtual-channel flow control, IEEE Transactions on Parallel and Distributed Systems 3 (2)
(1992) 194–205.
[4] G. Kim, J. Kim, S. Yoo, Flexibuffer : Reducing Leakage Power in On-Chip Network Routers, in:
Proceedings of the 48th Design Automation Conference, San Diego, USA, 2011, pp. 936–941.
[5] S. Vangal et al., An 80-tile sub-100-W TeraFLOPS processor in 65-nm CMOS, IEEE Journal of
Solid State Circuits 43 (1) (2008) 29–41.
[6] M. B. Taylor et al., Evaluation of the Raw Microprocessor: An Exposed-Wire-Delay Architecture for
ILP and Streams, in: Proceedings of the 31st Annual International Symposium on Computer
Architecture,Munchen, Germany, 2004, pp. 2–13.
[7] Y. Hoskote et al., A 5-GHz Mesh Interconnect for a Teraflops Processor, IEEE Micro 27 (5)(2007)
51–61.
[8] C. Fallin et al., MinBD: Minimally-Buffered Deflection Routing for Energy-Efficient Interconnect,
in: Proceedings of the 6th ACM/IEEE International Symposium on Networks on Chip, Lyngby,
Denmark, 2012, pp. 1–10.
[9] J. Jose, B. Nayak, M. Mutyam, DeBAR: Deflection Based Adaptive Router with Minimal Buffering,
in: Proceedings of the Design Automation and Test in Europe Conference, Grenoble, France, 2013,
pp. 1583–1588.
[10] E. Nilsson et al., Load distribution with the proximity congestion awareness in a network-on-chip, in:
Proceedings of the Design Automation and Test in Europe Conference, Munich, Germany, 2003, pp.
1126–1127.

71
[11] P. Baran, On distributed communication networks, IEEE Transactions on Communications
12(1)(1964) 1–9.
[12] M. Hayenga, N. E. Jerger, M. Lipasti, SCARAB: A Single Cycle Adaptive Routing and Bufferless
Network, in: Proceedings of the 42nd Annual IEEE/ACM International Symposium on
Microarchitecture,NewYork, USA, 2009, pp. 244–254.
[13] T. Moscibroda, O. Mutlu, A case for bufferless routing in onchip networks, in: Proceedings of the
36th International Symposium on Computer Architecture, Texas, USA, 2009, pp. 196–207.
[14] G. Michelogiannakis et al., Evaluating bufferless flow control for on-chip networks, in: Proceedings
of the 4th ACM/IEEE International Symposium on Networks on Chip, 2010, pp. 9–16.
[15] C. Fallin, C. Craik, O. Mutlu, CHIPPER: A Low complexity Bufferless Deflection Router, in:
Proceedings of the 17th IEEE International Symposium on High Performance Computer Architecture,
Texas, USA, 2011, pp. 144–155.
[16] S. Z. Sleeba, J. Jose, M. G. Mini, WeDBless: Weighted Deflection Bufferless Router for Mesh NoCs,
in: Proceedings of the 24th Great Lakes Symposium on VLSI, Houston, Texas, USA, ACM, 2014,
pp. 77–78.
[17] B. Nayak, J. Jose, M. Mutyam, SLIDER: Smart Late Injection DEflection Router for Mesh NoCs,
in: Proceedings of the 31st IEEE International Conference on Computer Design,Asheville, USA,
2013, pp. 377–383.
[18] J. Lin, X. Lin, L. Tang, Making-a-stop: A new bufferless routing algorithm for on-chip network,
Journal of Parallel and Distributed Computing 72 (2012) 515–524.
[19] Z. Lu, M. Zong, A. Jantsch, Evaluation of on-chip networks using deflection routing, in: Proceedings
of the 16th Great Lakes Symposium on VLSI, Philadelphia, USA, 2006, pp. 296–301.
[20] N.Jiang et al., Booksim 2.0 user’s guide (http://nocs.stanford.edu).
[21] R. Ubal et al., Multi2sim: A simulation framework to evaluate multicore-multithreaded Processors,
in: Proceedings of the 26th International Symposium on Computer Architecture and High
Performance Computing, Munich, Germany, 2007, pp. 62–68.
[22] SPEC2006 CPU Benchmark suite: (http://www.spec.org).
[23] A. B. Kahng et al., Orion 2.0: A fast and accurate NoC power and area model for early stage design
space exploration., in Proceedings of the Design Automation and Test in Europe Conference, Nice,
France, 2009, pp. 423–429.
AUTHORS
Simi Zerine Sleeba recieved her B.Tech degree in Electronics & Communication
Engineering from Mahatma Gandhi University, India in 1997 and M.Tech degree in
VLSI & Embedded Systems from Cochin University of Science and Technology,
India in 2010. Currently, she is pursuing her PhD research at Cochin University of
Science and Technology. Her research interests include On chip interconnection
network architectures and algorithms and low power MPSoC design.
Mini M.G. recieved her B.Tech degree in Electronics & Communication Engineering
from Kerala University, M.Tech degree in Digital Electronics and PhD in Image
Processing from Cochin University of Science and Technology, India. She is presently
working as the Academic Dean at Govt. Model Engineering College, Kochi, India.
Her research interests include low power MPSoC design, network on chip
architectures and digital image processing.

MINIMALLY BUFFERED ROUTER USING WEIGHTED DEFLECTION ROUTING FOR MESH NETWORK ON CH

More Related Content

What's hot (19)

Viewers also liked (19)

Similar to MINIMALLY BUFFERED ROUTER USING WEIGHTED DEFLECTION ROUTING FOR MESH NETWORK ON CH (20)

Recently uploaded (20)

MINIMALLY BUFFERED ROUTER USING WEIGHTED DEFLECTION ROUTING FOR MESH NETWORK ON CH