[ieee 2013 ieee symposium on computational intelligence for communication systems and networks...
TRANSCRIPT
Efficient Scheme for Load Balancing on
Heterogeneous Biswapped Networks
Liting Sun
School of Automation Engineering
University of Science and Technology Beijing
Beijing, China 100083
Email: [email protected]
Chaonan Tong
School of Automation Engineering
University of Science and Technology Beijing
Beijing, China 100083
Email: [email protected]
Abstract—Existing local iterative algorithms for load-balancing are poor-suited to many large-scale interconnectionnetworks. The main reason is complicated Laplace spectrum com-putations. In view of the large scale Biswapped network(BSN),our paper proposes a more simple and effective solution, namedHCDE-X, which is suitable to BSN of heterogeneous for loadbalancing. The scheme avoids the whole network’s Laplace matrixcalculation, only needs spectral information of the much smallerbasis or factor graph. By example illustrations and algorithmanalysis, the new suggested scheme lowers algorithm and thecomputational complexity, and reduces unnecessary communica-tion cost, and is more simple and effective than the traditionalones.
I. INTRODUCTION
Load balancing problem has been an important matter in
parallel and distributed processing, and which greatly affects
the performance of systems in processing time. The goal of
load balancing process is to migrate loads across edges, so
that all nodes achieve the balanced state after load migrating.
Now there have been many diffusion algorithms proposed
for load balancing on general networks. For homogeneous
networks, Cybenko [1] presented first order scheme(FOS).
Muthukrishnan et al. [2] improved FOS, and presented the
second order scheme (SOS), which speeded up the iteration
process. Diekmann et al. in [3] developed a new iterative
algorithm, called the optimal polynomial scheme(OPS). OPS
improved iteration convergence speed, and balanced the loads
among nodes within a finite number of iterations. For a more
simple and efficient algorithm than OPS, Elsasser et al. [4]
presented optimal diffusion scheme(OPT). In homogeneous
networks, the balanced state is achieved when all nodes of
networks have equal loads. But for heterogeneous networks,
each node has different capability, then the balanced state
means each node takes loads proportional to its capability.
So load-balancing algorithm in heterogeneous networks is a
bit more complicated than that in homogeneous networks.
For the latter one, Elsasse et al. [5] presented diffusive load
balancing scheme. Rotaru et al. in [6] designed an transition
algorithm from homogeneous system to heterogeneous system.
They all proved the effectiveness of the algorithms in theory.
All of above were the classical algorithms of load balancing
in the past. For the recent years, load balancing problem is
still a research hot topic, and it has been researched on more
ample scope. In [7], Alsarhan et al. suggested load balanced
clustering for cognitive radio technology. Warabino et al. in [8]
presented advanced load balancing which was used in cellular
network. Take the future trend of load balancing problem into
consideration, Keyvanpou in [9] proposed two main categories
that were topology dependent and topology independent of
load balancing algorithms in distributed systems.
However, traditional diffusion algorithms are not well
applied to large-scale interconnection networks (such as
biswapped networks), because of the complicated spectrum
computation. Biswapped network (BSN), which is a new type
of interconnection network, is closely related to the OTIS [10,
11]. BSN is of more regularity than the OTIS, and is built
of 2n copies of an n-node basic network (called the cluster)
which are isomorphism by using a simple rule for connectivity.
Recent studies show that BSN has good properties such as
low expansion cost and inherits good characteristics of factor
network which is similar to OTIS [12,13].
In this paper, we introduce a new algorithm HCDE-
X for load balancing to adapt to heterogeneous BSN. Our
paper is organized as follows. In section 2, we describe basic
definitions pertaining to load-balancing, diffusion algorithms
18978-1-4673-5903-0/13/$31.00 c©2013 IEEE
of heterogeneous networks, and the structure of BSN, and
in section 3, we propose algorithm for load balancing on
heterogeneous BSN. In section 4, we analyze the performance
of traditional algorithms and the new one, the results show the
simple and efficient of our approach.
II. DEFINITIONS
Definition1(BSNGraph). Let Ω=(VΩ, EΩ) be an
undirected graph. The BSN graph associated with Ω.
BSN (Ω)= (V,E) is an undirected graph with the vertex set
V (BSN(Ω)) = {< i, g, p > |g, p ∈ VΩ , i = 0, 1} and the
edges set E(BSN(Ω)) = {(< i, g, p1 >,< i, g, p2 >) |(p1,p2) ∈ EΩ, i = 0, 1} ∪ {(< i, g, p >,< 1− i, p, g >) |g, p ∈VΩ, i = 0, 1}.
From above, if we regard the basis network as group,
the definition postulates 2n groups, each group is an Ω
digraph. The name ”Biswapped network” (BSN) arises from
two defining properties of the network: one is when groups
are viewed as super-nodes, the resulting graph of super-nodes
is a complete 2n-node bipartite graph; the other one is that
the inter-group links connect processor g in cluster p of part
0 with processor p in cluster g of part 1.
If Ω has n nodes, then BSN (Ω) is composed of 2n
node-disjoint subnetworks Ωij(i = 0/1, 0 ≤ j ≤ n− 1),
which constitute the groups or clusters. Each group is
isomorphic to Ω. Denote the vertex set of Ωij as
Vij = {vijk |0 ≤ k ≤ n− 1, i = 0, 1}
and its edge set can be defined:
Eij = {(vjm, vjn) |(vm, vn) ∈ EΩ, i = 0, 1}.
The vertex set V of BSN (Ω) is
V = ∪viji∈{0,1};0≤j≤n−1
.
The edge set E of BSN (Ω) can be partitioned into two
subsets: the intragroup or basis edge set E1, and the intergroup
or swap edge set E2. Clearly,
E1 = ∪i∈{0,1};0≤j≤n−1
Eij ;
Fig.1 BSN-R4
E2 = {< vijk , v(1−i)kj> |k < j, i = 0, 1}.
Fig.1 contains an example of BSN , which is formed with
4−node ring as its basis or factor network.
Let Wij = (wij1 , wij2 , . . . , wijn)T (i = 0, 1) and
Cij = (cij1 , cij2 , . . . , cijn)T (i = 0, 1) represent the loads
and weight on the jth factor of part i respectively.
Similarly, let Wi = (wi1, wi2, . . . , win)T (i = 0, 1) and
Ci = (ci1, ci2, . . . , cin)T (i = 0, 1) denote the loads and
weight on BSN (Ω). Additionally, C ′ and C ′ij are taken to
be diagonal matrices, with elements of the vectors Ci and Cij
as their diagonal entries, respectively. That is:
C ′ = diag(c011 , c012 , . . . , c0nn , c111 , . . . , c1nn)
C ′ij = diag(c01, c02, . . . , c0n, . . . , c1n)(i = 0, 1)
Denote Af and A be the node-edge incidence matrices of
the basis graph Ω and BSN (Ω) respectively. Take A2 to be the
matrix specifying the incidence of the intergroup edges in E2
to nodes of BSN (Ω). Matrix A, Af and A2 all have in each
column exactly two nonzero entries 1 and -1, which represent
the nodes incident to the corresponding edges. The sign of
these nonzero entries imply directions of the flows produced
in the process of load-balancing on the corresponding edges.
The Laplacian L of a graph is L = AAT . Let L and Lf be
the Laplace matrices of BSN (Ω) and its basis network Ω,
respectively.
For heterogeneous networks, we have to use the general-
ized Laplacian representation: LC−1. In this case, the Lapla-
cian of BSN (Ω) and Ω are defined respectively as L = AAT
and Lf = Af AfT
, where A = AF−1 and Af = Af(F f
)−1.
From the above expression, the matrix F denotes diagonal
matrix with Fii =√fii , where fii indicates the edge-weight
of edges i in BSN (Ω). Likewise, the matrix F f denotes
diagonal matrix with F fii =
√ffii, where ff
ii denotes the edge-
weight of edge i in Ω. Then Laplacian matrix of BSN (Ω)
2013 IEEE Symposium on Computational Intelligence for Communication Systems and Networks (CIComms) 19
is LC ′−1, and the Laplacian matrix of Ω is LfC′−1ij . The
matrix Mf defined by Mf= I−αf LfC′−1ij is called diffusion
matrix of polynomial-based scheme, where α ∈ (0, 1) is a
constant. The similar to Mf , the matrix M is also got as
M = I−αL(C ′)−1. Let λf
1 < λf2 < . . . < λf
m′ be m′
distinct eigenvalues of the Laplacian LfC′−1ij in increasing
order, while λ1 < λ2 < . . . < λm are m distinct eigen-
values of L(C ′)−1. Then M and Mf have the eigenvalues
μi = 1− αλi and μfi = 1− αfλf
i respectively. We define the
diffusion norm γ and γf as γ ={max(
∣∣∣μf2
∣∣∣ , ∣∣μfm
∣∣)} < 1 and
γf ={max(
∣∣∣μf2
∣∣∣ , ∣∣∣μfm′
∣∣∣)} < 1 . For any polynomial-based
diffusion scheme, a small diffusion norm will lead to a fast rate
of convergence. The work load wk in step k can be expressed
in the form of wk = pk(M)w0 for any polynomial-based
load balancing schemes. The convergence of a polynomial
based load balancing scheme depends on whether the error
ek = wk −w between the iterate wk and the balanced load w
tends to zero [3].
III. HCDE-X FOR HETEROGENEOUS BSN
For large-scale BSN, it is difficult for us to compute all
eigenvalues of an 2n2 × 2n2 matrix before load balancing
process starts, and the calculation is large time costed. There-
fore we introduce a hybrid scheme of diffusion and dimension
exchange called HCDE-X scheme, where X represents any
general load balancing scheme. For the suggested scheme, the
balancing processes can be divided into four stages.
For doing inter-group links having much larger bandwidth
than doing intra-group links, we denote the edge weights of
the inter-group links, and let the weights on intra-group links
be infinite close to zero. Now, take lgp to be the corresponding
edge weight on the edge which linking the group g in part 0
and the group p in part 1. Then, lgp is expressed as follows.
lgp =(∑n
i=1 cgi)0 (∑n
i=1 cpi)1(∑ni=1
∑nj=1 cij
)0+
(∑ni=1
∑nj=1 cij
)1
(1)
Among the formula above, cgp is denoted as the weight
of node 〈g, p〉 of one part.
For the flowing discussion, we take zi and zfi to be
the corresponding eigenvectors of L(C ′)−1and LfC
′−1ij , and
define xi as the flows on the edges generated from load
balancing process in the i stage.
Now, the algorithm of HCDE-X is described as follows.
In the first stage, suggested scheme diffuses node loads
iteratively until all factor networks achieve balanced status. In
other words, the initial loads W 0ij of the jth factor network
in part i achieves balanced status W 0ij locally. Following this
stage, the workload W kij in step k can be expressed as
W kij =
(In ⊗ pk(M
f ))Wij . (2)
After this stage, the flows x1 produced from the factor
network can be got referencing in [14] as this:
x1 =(Af
(F f
)−1)T m′∑
i=2
1− (μi
f)k
λif
zfi . (3)
In formula 3, k is the iterations in the diffusion stage.
In the second stage, we perform a dimension and simple
diffusion strategy over all intergroup links. After the process,
the immigrated loads wΔ on every lgp can be expressed
wΔ =
∣∣∣∣ lgpcgp× wgp − lgp
cpg× wpg
∣∣∣∣ . (4)
For the formula, we denote the cgp and wgp as the weight
and loads on the node 〈g, p〉 of part 0, whereas cpg and wpg
are on behalf of the same meanings on the node 〈p, g〉 of part
1. The similar to formula (3), the flows produced after the
stage 2 is
x2 =
(I ⊗
(A2(F
′)−1)T
)∑m2
i=2
1−(μ
′i
)k∗
λ′i
z′i. (5)
We have specified the matrix A2 , now make supplement
explanation to the formula (5): we denote the λ′i (0 ≤ i ≤ m2)
to be the Laplacian eigenvalues of the graph, with which the
vertex set is V and the edge set is E2. Let z′i be the orthogonal
eigenvectors corresponding to λ′i, and μ
′i be the eigenvalues of
the corresponding to the diffusion matrix.
In the third and fourth stage, we proceed with diffusion
and dimension exchange using the same iterative polynomial-
based load balancing as in the first and second stage, and the
flows x3 generated from each factor network is got as
x3 =(Af
(F f
)−1)T m′∑
i=2
1− (μi
f)k∗∗
λif
zfi , (6)
the flows x4 is obtained
x4 =
(I ⊗
(A2(F
′)−1)T
)∑m2
i=2
1−(μ
′i
)k∗∗∗
λ′i
z′i. (7)
After the load balancing process finished, the whole
network achieves a load balancing status. The flows x
20 2013 IEEE Symposium on Computational Intelligence for Communication Systems and Networks (CIComms)
produced using the HCDE-X scheme can be calculated as
x =2n∑j=1
x1 + x2+2n∑j=1
x3 + x4
= x2 + x4 +2n∑j=1
(x1 + x3)
=(Af
(F f
)−1)T 2n∑
j=1
m′∑
i=2
2−(μfi )
k−(μfi )
k∗∗
(λi)f zfi
+I ⊗(A2(F
′)−1)T
m2∑i=2
2−(μ
′i
)k∗
−(μ
′i
)k∗∗∗
λ′i
z′i. (8)
If we use the general diffusion scheme X, the generated flows
x′
on BSN will be
x′=
(I ⊗ (
AF−1)T)∑m
i=2
1− (μi)k′
λizi. (9)
From our suggesting algorithm, we can see that most
of the migrations occurre in factor networks. Compared with
formula (8) and (9), the calculations of flows x resulting from
HCDE-X avoid the whole networks’s large scale complicated
Laplace spectrum computation, only just knowing the factor
networks’s Laplace eigenvalues. Therefore, our new scheme
lowers algorithm and the computational complexity.
Fig.2 illustrates the preceding four-stage load balancing
process. In the subfigure (a), the bigger integers represent the
loads of the nodes, while the decimal above the node denotes
the weight of each node. The decimals on the intergroup links
are the edge weights. Analyzing the figures, we can see that
in the subfigure (b), each factor network achieves the load
balancing status after the first diffusion stage. Following the
second stage, loads on nodes are diffused and exchanged on
the intergroup links in figure (c). The immigrated loads via
each inter-group link can be computed by using formula (4).
The similar to stage 1 and stage 2, stage 3 and stage 4 are
illustrated in subfigure (d) and (e). From subfigure (e), we can
see that the nodes with the same weights have the nearly work
loads, and the whole network is in a load balancing status.
The HCDE-X algorithm is outlined given below.
IV. ALGORITHM ANALYSIS
We know the convergence speed of HCDE-X is decided
by the diffusion norm of corresponding factor network, not
depends on the distribution of node weights, no matter X is
FOS, SOS, OPS or OPT.
Theorem1. Let Ω be the factor network of BSN (Ω),
then μf (Ω) ≺ μ (BSN (Ω)).
Proof : Let μ and μf be the Laplace eigenvalues of
BSN (Ω) and Ω, let λ and λf be the eigenvalues of
BSN (Ω) and basic network Ω . Denote D be degree
diagonal matrix, denote A be adjacency matrix, so we have
μf (Ω) = λf1 (D (Ω) +A (Ω)) (10)
BSN (Ω) can be described as a complete bipartite graph, Ω
is the connected generate support subgraph of BSN (Ω) ,then
μf (Ω) = μf1 (Ω) = λf
1 (D (Ω) +A (Ω))
≺ λ1 (D (BSN (Ω)) +A (BSN (Ω)))
= μ (BSN (Ω)) = μ1 (BSN (Ω))
(11)
μf (Ω) ≺ μ (BSN (Ω)) implies that when applying the
HCDE-FOS scheme and FOS scheme to BSN (Ω) at the
same time, HCDE-FOS will have a smaller upper bound
of error than FOS in the kth iteration according to∥∥ek∥∥
2≤ rk.
Theorem2. Let ψ′′t and ψ′t be potential of BSN (Ω)
and Ω after t steps iterations, then ψ′′t ψ′t.
Proof : We know ψt = μ2tψ0, ψ0 is the initial potential. By
using Theorem 1, we find that ψ′′t ψ′t.
Theorem3. To achieve balanced status on BSN, HCDE-OPS
has a smaller upper bound of iterations required.
Proof : We know that, for the OPS scheme, iterations
are not depending on the diameter of a graph. For the
symmetrical of BSN, the upper bound k of iterations of the
balancing flows satisfies k ≺ D(BSN (Ω)). By using HCDE-
X, d (Ω) ≺ D(BSN (Ω)) , then it implies when applying
the HCDE-OPS scheme and OPS scheme to BSN (Ω),
HCDE-OPS will have a smaller upper bound of iterations of
the balancing flows.
2013 IEEE Symposium on Computational Intelligence for Communication Systems and Networks (CIComms) 21
(a) Unbalanced initial node distribution on a Heterogeneous BSN-
R4
(b) Node loads after the first step on a Heterogeneous BSN-R4
(c) Node loads after the second step on a Heterogeneous BSN-R4
(d) Node loads after the third step on a Heterogeneous BSN-R4
(e) Node loads after the forth step on a Heterogeneous BSN-R4
Fig.2 An example of the HCDE-X scheme applied to heterogeneous
BSN
Algorithm 1 HCDE-X
Require: BSN (Ω) network consists of Ωij , Wij , Cij ;
Ensure: Balanced load vector, wlij ;
1: for all groups Ωij of BSN (Ω);
2: run the diffusion procedure X on Ωij ;
3: end for;
4: for all intergroup edges
5: e:(< i, g, p >,< 1− i, p, g >);
6: lgp =(∑n
i=1 cgi)(∑n
i=1 cpi)(∑n
i=1
∑nj=1 cij)
0+(
∑ni=1
∑nj=1 cij)
1
;
7: wgp = wgp +(
lgpcgp× wgp − lgp
cpg× wpg
);
8: end for
9: for all groups Ωij ;
10: run the diffusion procedure X on Ωij ;
11: end for;
12: for all intergroup edges
13: e:(< i, g, p >,< 1− i, p, g >);
14: lgp =(∑n
i=1 cgi)(∑n
i=1 cpi)(∑n
i=1
∑nj=1 cij)
0+(
∑ni=1
∑nj=1 cij)
1
;
15: wgp = wgp +(
lgpcgp× wgp − lgp
cpg× wpg
);
16: end for
17: for all groups Ωij ;
18: return blancing load vector as wlij ;
19: end for.
V. CONCLUSION
The suggested scheme described in this paper is based
on a common idea of hybrid diffusion-exchange-diffusion-
exchange strategy, but they take advantage of the special
structure of BSN to reduce the iterations and reduce the
required communication cost. A main focus of our ongoing
work is whether we can partition the intensive tasks into the
same groups by using the special structure of BSN, for purpose
of futher reducing the commucation cost.
REFERENCES
[1] G.Cybenko, Dynamic load balancing for distributed memory multipro-cessors, Journal of Parallel and Distributed Computing, 1989.7(2):279-
301.
[2] S.Muthukrishnan,B.Ghosh and M.H. Schultz, First- and second-orderdiffusive methods for rapid, coarse,distributed load balancing, Theory
of Computing Systems, 1998.31(4):331-354.
[3] R.Diekmann,A. Frommer and B. Monien, Efficient schemes for nearestneighbor load balancing, Parallel Computing, 1999.25(7):789-812.
[4] R.Elsasser,B.Monien,R.Preis,A.Frommer, Optimal and alternating-direction load balancing schemes, in Proceedings of Euro-Par’99, 31
Aug -3 Sept, 1999, Berlin, Germany: Springer-Verlag.
[5] R.Elsasser,B. Monien and R.Preis, Diffusive load balancing schemeson heterogeneous networks, Theory of Computing Systems,
2001.35(3):305-320.
22 2013 IEEE Symposium on Computational Intelligence for Communication Systems and Networks (CIComms)
[6] T.Rotaru,H.H. Nageli, Dynamic load balancing by diffusion in het-erogeneous systems, Journal of Parallel and Distributed Computing,
2004.64(4):481-97.
[7] A. Alsarhan, and A. Agarwal, Load balancing for spectrum managementin a cluster-based cognitive network, in 2011 Canadian Conference on
Electrical and Computer Engineering, CCECE 2011, 8 May-11 May,
2011, Niagara Falls, Canada: Institute of Electrical and Electronics
Engineers Inc.
[8] T.Warabino, S.Kaneko,S. Nanba, Advanced load balancing in LTE/LTE-A cellular network, in 2012 IEEE 23rd International Symposium on
Personal, Indoor and Mobile Radio Communications, PIMRC 2012, 9
September-12 September, 2012, Sydney, NSW, Australia: Institute of
Electrical and Electronics Engineers Inc.
[9] M. R.Keyvanpour,H.Mansourifar,B.Bagherzade, A novel classificationof load balancing algorithms in distributed systems, 2011 SSITE
International Conference on Computers and Advanced Technology in
Education, ICCATE 2011,3 November- 4 November,2011,2012, Beijing,
China: Springer Verlag.
[10] B.Parhami, Swapped interconnection networks: Topological, perfor-mance, and robustness attributes, Journal of Parallel and Distributed
Computing, 2005.65(11):1443-52.
[11] W.Chen,W.Xiao,B.Parhami, Swapped (OTIS) networks built of connect-ed basis networks are maximally fault tolerant, IEEE Transactions on
Parallel and Distributed Systems, 2009.20(3):361-366.
[12] W.Chen,W. Xiao, Topological Properties of Biswapped Networks (B-SNs): Node Symmetry and Maximal Fault Tolerance, Chinese Journal
of Computers, 2010.33(5):822-32.
[13] W.Xiao, Biswapped networks and their topological properties, in S-
NPD 2007: 8th ACIS International Conference on Software Engineering,
Artificial Intelligence, Networking, and Parallel/Distributed Computing,
30 July - 1 August ,2007.
[14] H.Arndt, On finite dimension exchange algorithms, Linear Algebra
and Its Applications, 2004. 380: p. 73-93.
2013 IEEE Symposium on Computational Intelligence for Communication Systems and Networks (CIComms) 23