IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO.

8, AUGUST 2011

5227

Optimal Exact-Regenerating Codes for Distributed

Storage at the MSR and MBR Points via a
Product-Matrix Construction
K. V. Rashmi, Nihar B. Shah, and P. Vijay Kumar, Fellow, IEEE

AbstractRegenerating codes are a class of distributed storage

codes that allow for efficient repair of failed nodes, as compared
to traditional erasure codes. An [n; k; d] regenerating code permits
the data to be recovered by connecting to any k of the n nodes in
the network, while requiring that a failed node be repaired by connecting to any d nodes. The amount of data downloaded for repair
is typically much smaller than the size of the source data. Previous
constructions of exact-regenerating codes have been confined to the
case n = d + 1. In this paper, we present optimal, explicit constructions of (a) Minimum Bandwidth Regenerating (MBR) codes for all
values of [n; k; d] and (b) Minimum Storage Regenerating (MSR)
codes for all [n; k; d  2k 0 2], using a new product-matrix framework. The product-matrix framework is also shown to significantly
simplify system operation. To the best of our knowledge, these are
the first constructions of exact-regenerating codes that allow the
number n of nodes in the network, to be chosen independent of the
other parameters. The paper also contains a simpler description, in
the product-matrix framework, of a previously constructed MSR
code with [n = d + 1; k; d  2k 0 1].
Index TermsDistributed storage, interference alignment, network coding, node repair, partial data recovery, product-matrix
framework, regenerating codes.

I. INTRODUCTION

N a distributed storage system, information pertaining to

a data file (the message) is dispersed across nodes in a
network in such a manner that an end-user can retrieve the data
stored by tapping into a subset of the nodes. A popular option
that reduces network congestion and that leads to increased
resiliency in the face of node failures is to employ erasure
coding, for example, by calling upon maximum-distance-separable (MDS) codes such as Reed-Solomon (RS) codes. Let
be the total file size measured in terms of symbols over a
of size . With RS codes, data is stored across
finite field
nodes in the network in such a way that the entire message can
be recovered by a data-collector by connecting to any nodes,
Manuscript received May 23, 2010; revised March 29, 2011; accepted March
29, 2011. Date of current version July 29, 2011. The results in this paper were
presented in part at the Information Theory and Applications Workshop, San
Diego, CA, Feb. 2011.
K. V. Rashmi and N. B. Shah are with the Department of Electrical Communication Engineering, Indian Institute of Science, Bangalore-560012, India
(e-mail: [email protected]; [email protected]).
P. V. Kumar is with the Department of Electrical Communication Engineering, Indian Institute of Science, Bangalore-560012, India. He is also with
the Electrical Engineering Systems Department, University of Southern California, Los Angeles, CA 90089-2565 USA (e-mail: [email protected]).
Communicated by N. Kashyap, Associate Editor for Coding Theory.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIT.2011.2159049

a process that we will refer to as data-reconstruction. Several

distributed storage systems such as RAID-6 [1], OceanStore
[2] and Total Recall [3] employ such an erasure-coding option.
A. Regenerating Codes
Upon failure of an individual node, a self-sustaining datastorage network must necessarily possess the ability to regenerate (i.e., repair) a failed node. An obvious means to accomplish this is to permit the replacement node to connect to any
nodes, download the entire message, and extract the data that
was stored in the failed node. But downloading the entire
units of data in order to recover the data stored in a single node
that stores only a fraction of the entire message is wasteful, and
raises the question as to whether there is a better option. Such an
option is indeed available and provided by the concept of a regenerating code introduced in the pioneering paper by Dimakis
et al. [4].
Conventional RS codes treat each fragment stored in a node as
a single symbol belonging to the finite field . It can be shown
that when individual nodes are restricted to perform only linear
, the total amount of data download needed
operations over
to repair a failed node can be no smaller than , the size of
the entire file. In contrast, regenerating codes are codes over a
vector alphabet and hence treat each fragment as being com. Linear operations
prised of symbols over the finite field
in this case, permit the transfer of a fraction of the data
over
stored at a particular node. Apart from this new parameter ,
two other parameters and are associated with regenerating
codes. Under the definition of regenerating codes introduced in
[4], a failed node is permitted to connect to an arbitrary set of
of the remaining nodes while downloading
symbols
from each node. This process is termed as regeneration and the
of data downloaded for repair purposes as the
total amount
repair bandwidth. Further, the set of nodes aiding in the repair
are termed as helper nodes. Typically, with a regenerating code,
the average repair bandwidth is small compared to the size of
the file .
It will be assumed throughout the paper, that whenever menregenerating code, the code is such
tion is made of an
that and are the minimum values under which data-reconstruction and regeneration can always be guaranteed. This restricts the range of to
(1)
The first inequality arises because if the regeneration parameter
were less than the data-reconstruction parameter then one

0018-9448/$26.00 2011 IEEE

5228

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 8, AUGUST 2011

could, in fact, reconstruct data by connecting to any

nodes
(treating the data stored in every other node as a function of that
stored in these nodes) thereby contradicting the minimality of
. Finally, while a regenerating code over
is associated with
the collection of parameters

it will be found more convenient to regard parameters

as primary and
as secondary and thus we will make
frequent references in the sequel, to a code with these six paregenerating code having parameter set
rameters as an
.
B. Cut-Set Bound and the Storage Versus Repair-Bandwidth
Tradeoff
A major result in the field of regenerating codes is the proof
in [5] that uses the cut-set bound of network coding to establish
that the parameters of a regenerating code must necessarily satisfy
(2)
It is desirable to minimize both as well as since, minimizing results in a minimum storage solution, while minimizing (for fixed ) results in a storage solution that minimizes repair bandwidth. As can be deduced from (2), it is not
possible to minimize both and simultaneously and thus there
is a tradeoff between choices of the parameters and . The
two extreme points in this tradeoff are termed the minimum
storage regeneration (MSR) and minimum bandwidth regeneration (MBR) points respectively. The parameters and for the
MSR point on the tradeoff can be obtained by first minimizing
and then minimizing to obtain

C. Striping of Data
The nature of the cut-set bound permits a divide-and-conquer
approach to be used in the application of optimal regenerating
codes to large file sizes, thereby simplifying system implementation. This is explained below.
regenerating code with parameter
Given an optimal
, a second optimal regenerating code with paramset
for any positive integer
eter set
can be constructed, by dividing the
message symbols into
groups of symbols each, and applying the
code to
each group independently. Secondly, a common feature of both
MSR and MBR regenerating codes is that in either case, their pais such that both and are multiples of
rameter set
and further that , are functions only of , and . It follows
MSR or MBR
that if one can construct an (optimal)
, then one can construct an (optimal)
code with
MSR or MBR code for any larger value of . In addition, from a
practical standpoint, a code constructed through concatenation
of codes for a smaller will in general, be of lesser complexity
(see Section VI-C). For these reasons, in the present paper we
. Thus, throughout the redesign codes for the case of
. In the termimainder of the paper, we will assume that
nology of distributed storage, such a process is called striping.
We document below the values of and of MSR and MBR
codes respectively, when
:
(5)
(6)
for MSR codes and
(7)
(8)
in the case of MBR codes.
D. Additional Terminology

(3)
Reversing the order leads to the MBR point which thus corresponds to

(4)
regenerating code as a code
We define an optimal
with parameters
satisfying the twin requirements that:
achieves the cut-set bound with
1) the parameter set
equality;
2) decreasing either or will result in a new parameter set
that violates the cut set bound.
regenerating code
An MSR code is then defined as an
satisfy (3) and similarly, an MBR
whose parameters
satisfying (4). Clearly,
code as one with parameters
both MSR and MBR codes are optimal regenerating codes.

1) Exact Versus Functional Regeneration: In the context of a

regenerating code, by functional regeneration of a failed node ,
we will mean, replacement of the failed node by a new node in
such a way that following replacement, the resulting network of
nodes continues to possess the data-reconstruction and regeneration properties. In contrast, by exact-regeneration, we mean
that
replacement of a failed node by a replacement node
stores exactly the same data as was stored in node prior to
failure. We will use the term exact-regenerating code to denote a
regenerating code that has the capability of exactly regenerating
each instance of a failed node. An exact-regenerating code is to
be preferred over a functional-regenerating code wherever possible, due to the following reasons. In a system where the code
coefficients are globally known, under functional-regeneration
there is need for the network to inform all nodes of the replacement. Moreover, the repair and decoding algorithms also need
to be re-tuned for the new set of coefficients. These additional
overheads are clearly unnecessary under exact-regeneration. In
addition, exact-regeneration permits the code to be systematic,
as described below.

RASHMI et al.: OPTIMAL EXACT-REGENERATING CODES FOR DISTRIBUTED STORAGE AT THE MSR AND MBR POINTS

2) Systematic Regenerating Codes: A systematic regenerating code can be defined as a regenerating code designed in
such a way that the message symbols are explicitly present
amongst the
code symbols stored in a select set of nodes,
termed as the systematic nodes. Clearly, in the case of systematic regenerating codes, exact-regeneration of (the systematic
portion of the data stored in) the systematic nodes is mandated.
3) Linear Regenerating Codes: A linear regenerating code is
defined as a regenerating code in which:
a) the code symbols stored in each node are linear combinaof the message symbols
;
tions over
b) the symbols passed by a helper node to aid in the rein the
generation of a failed node are linear over
symbols stored in node .
It follows as an easy consequence, that linear operations sufcode
fice for a data-collector to recover the data from the
symbols stored in the nodes that it has connected to. Similarly, the replacement node for a failed node , performs linear
operations on the symbols passed on to it by the helper nodes
aiding in the regeneration.
E. Results of the Present Paper
While prior work is described in greater detail in Section II,
we begin by providing a context for the results presented here.
Background: To-date, explicit and general constructions for
exact-regenerating codes at the MSR point have been found only
. Similarly at the MBR
for the case
point, the only explicit code to previously have been constructed
. Thus, all existing code constructions
is for the case
. This
limit the total number of nodes in the system to
is restrictive since in this case, the system can handle only a
single node failure at a time. Also, such a system does not permit
additional storage nodes to be brought into the system.
A second open problem in this area that has recently drawn
attention is as to whether or not the storage-repair bandwidth
tradeoff is achievable under the additional requirement of exactregeneration. It has previously been shown that no linear code
with
,
can achieve the MSR point for any
when (and hence
but is achievable for all parameters
as well) is allowed to approach infinity.
Results Presented in Present Paper: In this paper, (optimal)
explicit constructions of exact-regenerating MBR codes for all
and exact-regenerating MSR codes for all
values of
are presented. The constructions are of a
product-matrix nature that is shown to significantly simplify
operation of the distributed storage network. The constructions
presented prove that the MBR point for exact-regeneration can
be achieved for all values of the parameters and that the MSR
.
point can be achieved for all parameters satisfying
In both constructions, the message size is as dictated by cut-set
bound. The paper also contains a simpler description, in the
product-matrix framework, of an MSR code for the parameters
that was previously constructed in
[6], [7].
A brief overview of prior work in this field is provided in
Section II. The product-matrix framework underlying the code
construction is described in Section III. An exact-regenerating
MBR code for all feasible values of the parameters

5229

is presented in Section IV, and an exact-regenerating MSR

is presented in Section V.
code for all
Implementation advantages of the particular product-matrix
nature of the code constructions provided here are described in
Section VI. The final section, Section VII, draws conclusions.
Appendix A contains a simpler description, in the product-matrix framework, of an MSR code with parameter satisfying
, that was previously constructed in
[6] and [7].
II. PRIOR WORK
The concept of regenerating codes was introduced in [4],
where it was shown that permitting the storage nodes to store
units of data helps in reducing the repair bandmore than
width. Several distributed systems were analyzed, and estimates
of the mean node availability in such systems obtained. Using
these values, it was shown through simulation, that regenerating
codes can reduce repair bandwidth compared to other designs,
while simplifying system architecture.
The problem of minimizing repair bandwidth for functional
repair of a failed storage node is considered in [4], [5]. Here,
the evolution of the storage network through a sequence of failures and regenerations is represented as a network, with all possible data-collectors represented as sinks. The data-reconstruction requirement is formulated as a multicast network coding
problem, with the network having an infinite number of nodes.
The cut-set analysis of this network leads to the relation between the parameters of a regenerating code given in (2). It
can be seen that there is a tradeoff between the choice of the
and this is termed as the
parameters and for a fixed
storage-repair bandwidth tradeoff. It has been shown ([5], [8])
that this tradeoff is achievable under functional-regeneration.
However, the coding schemes suggested are not explicit and
require large field size. The journal version [9] also contains
a handcrafted functional-regenerating code for the MSR point
.
with
A study of the computational complexity of regenerating
codes is carried out in [10], in the context of random linear
regenerating codes that achieve functional repair.
The problem of exact-regeneration was first considered independently in [11][13]. In [11], it is shown that the MSR point is
.
achievable under exact-regeneration when
The coding scheme proposed is based on the concept of interference alignment developed in the context of wireless communication. However, the construction is not explicit and has a large
field size requirement. In [13], the authors carry out a computer
search to find exact-regenerating codes at the MSR point, resulting in identification of codes with parameters
.
The first, explicit construction of regenerating codes for a
general set of parameters was provided for the MBR point in
and arbitrary . These codes have low re[12] with
generation complexity as no computation is involved during the
exact-regeneration of a failed node. The field size required is of
the order of . In addition, [12] (see also [14]) also contains
the construction of an explicit MSR code for
, that
performs approximately-exact-regeneration of all failed nodes,

5230

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 8, AUGUST 2011

i.e., regeneration where a part of the code is exactly regenerated, and the remaining is functionally regenerated (it is shown
subsequently in [6], [14] that exact-regeneration is not possible,
, for the set of parameters considered therein).
when
MSR codes performing a hybrid of exact and functional-regeneration are provided in [15], for the parameters
and
. The codes given even here are nonexplicit, and
have high complexity and large field-size requirement.
A code structure that guarantees exact-regeneration of just the
systematic nodes is provided in [6], for the MSR point with pa. This code makes use
rameters
of interference alignment, and is termed as the MISER code
in journal-submission version [14] of [6]. Subsequently, it was
shown in [7] that for this set of parameters, the code introduced
in [6] for exact-regeneration of only the systematic nodes can
also be used to repair the nonsystematic (parity) node failures
exactly provided repair construction schemes are appropriately
designed. Such an explicit repair scheme is indeed designed and
presented in [7]. The paper [7] also contains an exact-regener.
ating MSR code for parameter set
A proof of nonachievability of the cut-set bound on exact-regeneration at the MSR point with linear codes, for the paramwhen
, is provided in [6], [14].
eters
On the other hand, the MSR point is shown to be achievable in
the limiting case of approaching infinity (i.e., approaching
infinity) in [16], [17].
A flexible setup for regenerating codes is described in [18],
where a data-collector (or a replacement node) can perform
data-reconstruction (or regeneration) irrespective of the number
of nodes to which it connects, provided the total data downloaded exceeds a certain threshold.
In [19], the authors establish that essentially all points on the
interior of the tradeoff (i.e., points other than MSR and MBR)
are not achievable under exact-regeneration.

This common structure of the code matrices leads to common

architectures for both data-reconstruction and exact-regeneration, as explained in greater detail below. It also endows the
codes with implementation advantages that are discussed in
Section VI.
Data-reconstruction amounts to recovering the message mafrom the
symbols obtained from an arbitrary set of
trix
storage nodes. Let us denote the set of nodes to which the
. The th node in this set
data-collector connects as
to the data-collector. The
passes on the message vector
data-collector thus obtains the product matrix

where

is the submatrix of
consisting of the rows
. It then uses the properties of the matrices
and
to recover the message. The precise procedure for reis a function of the particular construction.
covering
As noted above, each node in the network is associated to a
encoding vector . In the regeneration process,
distinct
of length , that
we will need to call upon a related vector
contains a subset of the components of . To regenerate a failed
node , the node replacing the failed node connects to an arbiof storage nodes which we will refer
trary subset
to as the helper nodes. Each helper node passes on the inner
product of the symbols stored in it with , to the replacement
passes
node: the helper node

The replacement node thus obtains the product matrix

where

III. COMMON PRODUCT-MATRIX FRAMEWORK

The constructions described in this paper follow a common
product-matrix framework. Under this framework, each codeword in the distributed storage code can be represented by an
code matrix whose th row contains the symbols
stored by the th node. Each code matrix is the product
(9)
encoding matrix and an
message maof an
trix . The entries of the matrix are fixed a priori and are
independent of the message symbols. The message matrix
contains the message symbols, with some symbols possibly
of as the encoding
repeated. We will refer to the th row
vector of node as it is this vector that is used to encode the
message into the form in which it is stored within the th node
(10)
where the superscript is used to denote the transpose of a matrix. Throughout this paper, we consider all symbols to belong
of size .
to a finite field

is the submatrix of consisting of the rows

. From this it turns out, as will be shown subsequently, that one can recover the desired symbols. Here again,
the precise procedure is dependent on the particular construction.
Remark 1: An important feature of the product-matrix conparticistruction presented here, is that each of the nodes
pating in the regeneration of node , needs only have knowledge
of the encoding vector of the failed node and not the identity
of the other nodes participating in the regeneration. This significantly simplifies the operation of the system.
Systematic Codes: The following theorem shows that any
linear exact-regenerating code can be converted to a systematic
form via a linear remapping of the symbols. The proof of the
theorem may be found in Appendix B.
Theorem 1: Any linear exact-regenerating code can be converted to a systematic form via a linear remapping of the message symbols. Furthermore, the resulting code is also linear and
possesses the data-reconstruction and exact-regeneration properties of the original code.
Thus, all codes provided in the present paper can be converted to a systematic form via a linear remapping of the message symbols. Specific details on the product-matrix MBR and

RASHMI et al.: OPTIMAL EXACT-REGENERATING CODES FOR DISTRIBUTED STORAGE AT THE MSR AND MBR POINTS

MSR codes in systematic form are provided in the respective

sections, Sections IV and V.
IV. PRODUCT-MATRIX MBR CODE CONSTRUCTION
In this section, we identify the specific make-up of the enthat results in an
coding matrix and the message matrix
MBR code with
. A notable feature of the con,
struction is that it is applicable to all feasible values of
. Since the code is
i.e., all , , satisfying
required to be an MBR code with
, it must possess the
data-reconstruction and exact-regeneration properties required
of a regenerating code, and in addition, have parameters
that satisfy (7) and (8). Equation (8) can be rewritten in the form

Thus the parameter set of the desired

MBR code is

Let be a
matrix constructed so that the
entries
in the upper-triangular half of the matrix are filled up by
distinct message symbols drawn from the set
. The
entries in the strictly lower-triangular portion of the matrix are
then chosen so as to make the matrix a symmetric matrix. The
message symbols are used to fill up a second
remaining
matrix . The message matrix
is then defined
as the
symmetric matrix given by
(11)
The symmetry of the matrix will be found to be instrumental
when enabling node repair. Next, define the encoding matrix
to be any
matrix of the form

where and are

and
matrices respectively, chosen in such a way that:
1) any rows of are linearly independent;
2) any rows of are linearly independent.
The above requirements can be met, for example, by choosing
to be either a Cauchy [20] or else a Vandermonde matrix.1 The
only constraint on the field size comes from the above required
properties of the encoding matrix . For instance, when is
chosen as a Vandermonde matrix, any field of size or higher
suffices.
As per the product-matrix framework, the code matrix is then
. The two theorems below establish that the
given by
code presented is an
MBR code by establishing respectively, the exact-regeneration and data-reconstruction properties
of the code.
Theorem 2 (MBR Exact-Regeneration): In the code presented, exact-regeneration of any failed node can be achieved

1Over a large finite field, a randomly chosen matrix

will suffice with high
probability. The present paper does not elaborate on the same, since the focus
is on providing explicit, deterministic code constructions.

5231

by downloading one symbol each from any of the

remaining nodes.
be the row of corresponding to the failed
Proof: Let
node . Thus the symbols stored in the failed node correspond
to the vector
(12)
The replacement for the failed node connects to an arbitrary
of helper nodes. Upon being contacted
set
by the replacement node, the helper node computes the inner
product

and passes on this value to the replacement node. Thus, in the

equals
itself. The represent construction, the vector
from
placement node thus obtains the symbols
the helper nodes, where

..
.
By construction, the
matrix
is invertible. Thus,
through multiplication on
the replacement node recovers
. Since
is symmetric
the left by
(13)
and this is precisely the data previously stored in the failed node.
Theorem 3 (MBR Data-Reconstruction): In the code premessage symbols can be recovered by consented, all the
necting to any nodes, i.e., the message symbols can be recovered through linear operations on the entries of any rows of
the matrix .
Proof: Let
(14)
submatrix of , corresponding to the rows of
be the
to which the data-collector connects. Thus, the data-collector
has access to the symbols
(15)
By construction,
is a nonsingular matrix. Hence, by mulon the left by
, one can recover
tiplying the matrix
first and subsequently, .
A. An Example for the Product-Matrix MBR Code
Let
,
,
. Then
and
. Let us
so we are operating over
. The matrices and
choose
are filled up by the 9 message symbols
as follows:
(16)

5232

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 8, AUGUST 2011

Fig. 1. Example for the MBR code construction: On failure of node 1, the replacement node downloads one symbol each from nodes 2, 4, 5 and 6, using which
node 1 is exactly regenerated. The notation h1; 1i indicates an inner product of the stored symbols with the vector [ 1 1 1 1 ] .

so that the message matrix

matrix2. Clearly the code is systematic. It can be verified that

the matrix has the properties listed just above Theorem 2.

is given by

(17)

We choose
given by

to be the

Vandermonde matrix over

(18)

Fig. 1 shows at the top, the

code matrix
with entries expressed as functions of the message symbols
. The rest of the figure explains how exact-regeneration
of failed node 1 takes place. To regenerate node 1, the helper
nodes (nodes 2, 4, 5, 6 in the example), pass on their respective
for
, 4, 5, 6. The
inner products
replacement node then recovers the data stored in the failed
where
node by multiplying by

V. THE PRODUCT-MATRIX MSR CODE CONSTRUCTION

In this section, we identify the specific make-up of the enthat results in an
coding matrix and the message matrix
MSR code with
. The construction applies to
.3 Since the code is required
all parameters
to be an MSR code with
, it must possess the data-reconstruction and exact-regeneration properties required of a rethat
generating code, and in addition, have parameters
satisfy (5) and (6). We begin by constructing an MSR code in
and will show
the product-matrix framework for
in Section V-C how this can be very naturally extended to yield
.
codes with
we have
At the MSR point with
(21)
and hence
(22)
Also
(23)

(19)

We define the

message matrix

as
(24)

as explained in the proof of Theorem 2 above.

B. Systematic Version of the Code
As pointed out in Section III, any exact-regenerating code
can be made systematic through a nonsingular transformation
of the message symbols. In the present case, there is a simpler
approach, in which the matrix can be chosen in such a way
that the code is automatically systematic. We simply make the
choice:
(20)

where
and
are
symmetric matrices constructed
such that the
entries in the upper-triangular part of each
of the two matrices are filled up by
distinct message
symbols. Thus, all the
message symbols are
and . The entries in the
contained in the two matrices
and
strictly lower-triangular portion of the two matrices
are chosen so as to make the matrices
and
symmetric.
Next, we define the encoding matrix to be the
matrix
given by
(25)

where
is the
zero matrix, and

identity matrix, 0 is a
are matrices of sizes
respectively, such that

and
is a Cauchy

2In general, any matrix, all of whose submatrices are of full rank, will suffice.
3As mentioned previously, it is impossible to construct linear MSR codes for
the case of d < 2k 0 3 when  = 1 (see [6], [14]).

RASHMI et al.: OPTIMAL EXACT-REGENERATING CODES FOR DISTRIBUTED STORAGE AT THE MSR AND MBR POINTS

5233

where is an
matrix and is an
diagonal
matrix. The elements of are chosen such that the following
conditions are satisfied:
1) any rows of are linearly independent;
2) any rows of are linearly independent;
3) the diagonal elements of are distinct.
The above requirements can be met, for example, by choosing
to be a Vandermonde matrix with elements chosen carefully
to satisfy the third condition. In this case, let the th row of
(for
) be
, which gives
. In order to satisfy the third property,
to be any field of size
or higher,
one may choose
, where is the generator of the multiplicative
with
. Note that as in the MBR code, the
group of the finite field
only constraint on the field size in this construction arises from
the above required properties of the encoding matrix .
Then under our code-construction framework, the th row of
product matrix
, contains the code
the
symbols stored by the th node. The two theorems below establish that the code presented is an
MSR code by establishing respectively, the exact-regeneration and data-reconstruction properties of the code.

which is precisely the data previously stored in the failed node.

Theorem 4 (MSR Exact-Regeneration): In the code presented, exact-regeneration of any failed node can be achieved
of the
by downloading one symbol each from any
nodes.
remaining
be the row of corresponding
Proof: Let
to the failed node. Thus the symbols stored in the failed node
were

(30)

(26)
The replacement for the failed node connects to an arbitrary
of helper nodes. Upon being contacted
set
by the replacement node, the helper node computes the inner
product
and passes on this value to the replacement
node. Thus, in the present construction, the vector
equals .
The replacement node thus obtains the symbols
from the helper nodes, where

Theorem 5 (MSR Data-Reconstruction): In the code premessage symbols can be recovered by

sented, all the
connecting to any nodes, i.e., the message symbols can be
recovered through linear operations on the entries of any rows
of the code matrix .
Proof: Let
(28)
be the
submatrix of , containing the rows of which
correspond to the nodes to which the data-collector connects.
Hence, the data-collector obtains the symbols

(29)
The data-collector can post-multiply this term with
tain

Next, let the matrices

and

to ob-

be defined as
(31)
(32)

As
and
are symmetric, the same is true of the matrices
and . In terms of and , the data-collector has access to the
symbols of the matrix
(33)
The

th,

, element of this matrix is

(34)

while the

th element is given by

(35)
..
.

By construction, the
matrix
the replacement node now has access to

is invertible. Thus

where (35) follows from the symmetry of and . By construction, all the
are distinct and hence using (34) and (35),
,
for all
the data-collector can solve for the values of
.
be given by
Consider first the matrix . Let
..
.

(36)

As
and
are symmetric matrices, the replacement node
and
.
has thus acquired through transposition, both
Using this, it can obtain

All the nondiagonal elements of are known. The elements in

the th row (excluding the diagonal element) are given by

(27)

(37)

5234

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 8, AUGUST 2011

Fig. 2. Example for the MSR code construction: On failure of node 1, the replacement node downloads one symbol each from nodes 2, 4, 5, and 6, using which
1 indicates an inner product of the stored symbols with the vector [ 1 1 ] .
node 1 is exactly regenerated. The notation

< :; >

However, the matrix to the right is nonsingular by construction

and hence the data-collector can obtain

Hence the
are

matrix

and the

diagonal matrix

(38)
Selecting the first

of these, the data-collector has access to

..
.

(43)

(39)

The matrix on the left is also nonsingular by construction and

hence the data-collector can recover . Similarly, using the
values of the nondiagonal elements of , the data-collector can
recover .
Remark 2: It is shown in [6], [14] that interference alignment is, in fact, a necessary ingredient of any minimum storage
regenerating code. Interference alignment is also present in the
product-matrix MSR code, and Appendix C brings out this connection.
A. An Example for the Product-Matrix MSR Code

code matrix
with
Fig. 2 shows at the top, the
. The
entries expressed as functions of the message symbols
rest of the figure explains how exact-regeneration of failed node
1 takes place. To regenerate node 1, the helper nodes (nodes 2,
4, 5, 6 in the example), pass on their respective inner products
for
, 4, 5, 6. The replacement node multiplies
, where
the symbols it receives with
(44)

and decodes

,
,
. Then
and
. Let us choose
, so we are operating over
. The matrices
and
are filled up by the six message
symbols
as follows:

and

Let

(40)
so that the message matrix

is given by

(41)

We choose
given by

to be the

Vandermonde matrix over

(42)

(45)
and
to obtain the data stored in
Finally, it processes
the failed node as explained in the proof of Theorem 4 above.
B. Systematic Version of the Code
It was pointed out in Section III, that every exact-regenerating
code has a systematic version and further, that the code could be
made systematic through a process of message-symbol remapping. In the following, we make this more explicit in the context
of the product-matrix MSR code.
be the
submatrix of , containing the rows
Let
of corresponding to the nodes which are chosen to be made
symbols stored in these nodes are
systematic. The set of
matrix
. Let be a
given by the elements of the
matrix containing the
source symbols. We map
(46)
and solve for the entries of
in terms of the symbols in .
This is precisely the data-reconstruction process that takes place

RASHMI et al.: OPTIMAL EXACT-REGENERATING CODES FOR DISTRIBUTED STORAGE AT THE MSR AND MBR POINTS

when a data-collector connects to the chosen nodes. Thus,

can be obtained by following the
the value of the entries in
to obtain
procedure outlined in Theorem 5. Then, use this
the code
. Clearly, in this representation, the chosen
nodes store the source symbols in uncoded form.
C. Explicit MSR Product-Matrix Codes for
In this section, we show how an MSR code for
can be used to obtain MSR codes for all
. Our starting
point is the following theorem.
Theorem 6: An explicit
exact-regenerating code that achieves the cut-set bound
at the MSR point can be used to construct an explicit
exact-regenerating code that also achieves the cut-set bound
at the MSR point. Furthermore if
in code ,
in code . If is linear, so is .
Proof: If both codes operate at the MSR point, then the
number of message symbols , in the two cases must satisfy

respectively, so that

We begin by constructing an MSR-point-optimal

exact-regenerating code in systematic form with the first
rows containing the
message symbols. Let
be the subcode of consisting of all code matrices in whose top row
message symbols
is the all-zero row (i.e., the first of the
are all zero). Clearly, the subcode
is of size
.
also possesses the same exact-regeneration and
Note that
data-reconstruction properties as does the parent code .
Let the code now be formed from subcode by puncturing
(i.e., deleting) the first row in each code matrix of . Clearly,
exact-recode is also of size . We claim that is an
generating code. The data-reconstruction requirement requires
that the underlying message symbols be recoverable from the
contents of any rows of a code matrix in . But this follows since, by augmenting the matrices of code by placing at
the top an additional all-zero row, we obtain a code matrix in
and code
has the property that the data can be recovered
from any
rows of each code matrix in . A similar argument shows that code also possesses the exact-regeneration
property. Clearly if is linear, so is code . Finally, we have

By iterating the procedure in the proof of Theorem 6 above

times we obtain:
Corollary 7: An explicit
exact-regenerating code that achieves the cut-set bound
at the MSR point can be used to construct an explicit
exact-regenerating code that also achieves the cut-set bound

in code
at the MSR point. Furthermore if
in code . If is linear, so is .
The corollary below follows from Corollary 7 above.

5235

Corollary 8: An MSR-point optimal exact-regenerating code

with parameters
for any
can
be constructed from an MSR-point optimal exact-regenerating
code with
and
. If is linear, so is .
VI. ANALYSIS AND ADVANTAGES OF THE CODES
In this section, we detail the system-implementation advantages of the two code constructions presented in the paper.
A. Reduced Overhead
In the product-matrix based constructions provided, the data
stored in the th storage node in the system is completely deterof length . This is in
mined by the single encoding vector
contrast to a
generator matrix in a general code, comprising of the encoding vectors of length as its columns,
each associated to a different symbol stored in the node. The
encoding vector suffices for the encoding, data-reconstruction,
and regeneration purposes. The short length of the encoding
vector reduces the overhead associated with the need for nodes
to communicate their encoding vectors to the data-collector
during data-reconstruction, and to the replacement node during
regeneration of a failed node.
Also, in both MBR and MSR code constructions, during regeneration of a failed node, the information passed on to the replacement node by a helper node is only a function of the index
of the failed node. Thus, it is independent of the identity of the
other nodes that are participating in the regeneration.
Once again, this reduces the communication overhead by requiring less information to be disseminated.
B. Applicability to Arbitrary
In any real-world distributed storage application such as
peer-to-peer storage, cloud storage, etc, it is natural that the
number of nodes may go up or down: in due course of time,
new nodes may be added to the system, or multiple nodes may
fail or exit the system. For example, in peer-to-peer systems,
individual nodes are free to come and go at will. The existing,
explicit constructions of exact-regenerating codes [6], [7],
. On the other
[11][13] restrict the value of to be
hand, the codes presented in this paper are applicable for all
values of , and independent of the values of the parameters
and . This gives a practical appeal to the code constructions
presented here.
C. Complexity
1) Linearity and Field Size: The codes are linear over a
, i.e., the source symbols are from this
chosen finite field
finite field, and any stored symbol is a linear combination of
. As mentioned previously, to arrive at
these symbols over
the product-matrix MBR code, any field of size or higher suffices, and for the product-matrix MSR code, any field of size
or higher suffices. By cleverly choosing the matrix
that meets the conditions governing the respective codes, it
may often be possible to reduce the field size even further.

5236

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 8, AUGUST 2011

2) Striping: The codes presented here divide the entire mes. Since each
sage into stripes of sizes corresponding to
stripe is of minimal size, the complexity of encoding, data-reconstruction and regeneration operations, are considerably lowered, and so are the buffer sizes required at data-collectors and
replacement nodes. Furthermore, the operations that need to be
performed on each stripe are identical and independent, and
hence can be performed in parallel efficiently by a GPU/FPGA/
multi-core processor.
3) Choice of the Encoding Matrix : The encoding matrix
, for both the codes described, can be chosen as a Vandermonde matrix. Then each encoding vector can be described by
just a scalar. Moreover with this choice, the encoding, datareconstruction, and regeneration operations are, for the most
part, identical to encoding or decoding of conventional ReedSolomon codes.

At the MSR point, with

(5) and (6) that

, we have from
(47)
(48)

Let be a
message symbols
matrix4 given by

matrix whose entries are precisely the

and let
be the
message

(49)
Next, let be a
chosen such that

Cauchy matrix over

and

a scalar

(50)
Let

be the

encoding matrix given by

VII. CONCLUSIONS
In this paper, an explicit MBR code for all values of the
system parameters
, and an explicit MSR code for all
are presented. Both
parameters satisfying
constructions are based on a common product-matrix framework introduced in this paper, and possess attributes that make
them attractive from an implementation standpoint. To the best
of our knowledge, these are the first explicit constructions of
exact-regenerating codes that allow to take any value independent of the other parameters; this results in a host of desirable properties such as the ability to optimally handle multiple
simultaneous node failures as well as the ability of allowing the
total number of storage nodes in the system to vary with time.
Our results also prove that the MBR point on the storage-repair
bandwidth tradeoff is achievable under the additional constraint
of exact-regeneration for all values of the system parameters,
and that the MSR point is achievable under exact-regeneration
.
for all
APPENDIX A
DESCRIPTION OF A PREVIOUSLY CONSTRUCTED MSR CODE IN
THE PRODUCT-MATRIX FRAMEWORK
An explicit code that performs data-reconstruction, and
exact-regeneration of the systematic nodes is provided in [6],
.
for the MSR point with parameters
Subsequently, it was shown in [7] that for this set of parameters,
the code introduced in [6] for exact-regeneration of only the
systematic nodes can also be used for exact-regeneration of
the nonsystematic (parity) nodes, provided repair construction
schemes are appropriately designed. Such an explicit repair
scheme is indeed designed and presented in [7]. In this section,
we provide a simpler description of this code in the product-matrix framework.
, since the
As in [6], [7], we begin with the case
code as well as both data-reconstruction and exact-regeneration
algorithms can be extended to larger values of by making use
of Corollary 8.

(51)
The code constructed in [6], [7] can be verified to have an
alternate description as the collection of code matrices of the
form
(52)
Note that the first nodes store the message symbols in uncoded
form and hence correspond to the systematic nodes. A simple
description of the exact-regeneration and data-reconstruction
properties of the code is presented below.
Theorem 9 (Exact-Regeneration): In the code presented,
exact-regeneration of any failed node can be achieved by downnodes.
loading one symbol each from the remaining
used in the exactProof: In this construction, the vector
regeneration of a failed node is composed of the first
symbols of
.
1) Exact-Regeneration of Systematic Nodes: Consider
regeneration of the th systematic node. The symbols thus
. The replacement
desired by the replacement node are
node obtains the following
symbols from the remaining
nodes:
(53)
matrix which is the identity matrix
where is a
with th row removed. Since is full rank by construction, the
replacement node has access to

(54)

4Note that the constructions presented in Sections IV and V employ a (d )

matrix M as the message matrix, whereas the dimension of M in the present
construction is ((d + 1) ).

RASHMI et al.: OPTIMAL EXACT-REGENERATING CODES FOR DISTRIBUTED STORAGE AT THE MSR AND MBR POINTS

5237

From (53) and (54), we see that the replacement node has access
to

indexed through are also known, the data-collector has thus

access to the product

(55)

(60)

, the
matrix on the left is nonsingular.
Since
,
This allows the replacement node to recover the symbols
desired.
which are precisely the set of symbols
2) Exact-Regeneration of Non-Systematic Nodes: Let
be the row of corresponding to the failed node. Then the
symbols stored in the failed node are
. The replacesymbols
ment node requests and obtains the following
from the remaining nodes:

Now as
is nonsingular, being a
subma.
trix of a Cauchy matrix, the data-collector can recover
In this way, the data-collector has recovered all the entries in
the rows of indexed by , as well as all the entries in the
columns of indexed by . Clearly, the same statement holds
when is replaced by . Thus the data-collector has access to
the product:
(61)

(57)

Again,
is nonsingular, and this enables the data-collector
. It is easy to see that since
to recover
, from the diagonal elements of this matrix, all the diagonal
can be obtained. The nondiagonal elements
elements of
and
for
,
are however of the form
,
. Again since
, all the nondiagonal elements
can also be decoded. In this way, the data-collector
of
has recovered all the entries of .

(58)

APPENDIX B
EQUIVALENT CODES AND CONVERSION OF NONSYSTEMATIC
CODES TO SYSTEMATIC

(56)
where
is the submatrix of containing the
rows
corresponding to the remaining nonsystematic nodes. This gives
and therefore to
the replacement node access to

Hence the replacement node has access to

The matrix on the left is easily verified to be nonsingular and

and
individually
thus the replacement node acquires
from which it can derive the desired vector
.
Theorem 10 (Data-Reconstruction): In the code presented,
all the message symbols can be recovered by connecting to
any nodes, i.e., the message symbols can be recovered through
linear operations on the entries of any rows of the matrix .
Proof: We first introduce the following notation to denote
matrix and ,
submatrices of a matrix. If is an
are arbitrary subsets of
and
respecto denote the submatrix of contively, we will use
taining only the rows and columns, respectively, specified by the
indices in and . For the cases when either
or
, we will simply indicate this as all.
and
be the
Let
systematic and nonsystematic nodes respectively to which
, i.e., the
the data-collector connects. Let
systematic nodes to which the data-collector does not connect.
Then the data-collector is able to access the
symbols

In this section, we define the notion of equivalent codes,

and show that any exact-regenerating code is equivalent to a
systematic exact-regenerating code.
Given any linear exact-regenerating code, one can express
symbols stored in the nodes as a linear comeach of the
bination of the message symbols
. Let
denote the th symbol stored in the th node.
Thus, we have the relation:

(62)
block generator matrix
where the
is composed of the component generator submatrices

each of size
, and associated to a distinct node.5 Let
denote the column-space of . A little thought will show that
a distributed storage code is an exact-regenerating code iff:
,
1) for every subset of nodes

(59)
Thus, the data-collector has access to the rows of indexed
by the entries of and consequently, has access to the correas well.
sponding columns of
Consider the columns of
indexed by .
are known, the dataSince the entries of these columns in
. Now since the rows of
collector has access to

and
2) for every subset of
the subspaces

nodes
contain a vector

,
such that

5In the terminology of network coding, the (B

1) column vector
termed the j th global kernel associated to the ith node.

is

5238

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 8, AUGUST 2011

We can thus define two exact-regenerating codes to be equivaare identical. It is also

lent if the associated subspaces
clear that two codes are equivalent if one can be obtained from
the other through a nonsingular transformation of the message
symbols and the symbols stored within the nodes. With these
two observations, it follows that two codes with generator matrices having the following relation are equivalent:

..

where the
pre-multiplication matrix , and the
post-multiplication block diagonal matrix comprising of
matrices
, are nonsingular. Clearly, equivthe
alent codes have identical data-reconstruction and regeneration
properties.
Systematic Version of Exact-Regenerating Codes: It also
follows that any exact-regenerating code is equivalent to a systematic, exact-regenerating code. To see this, suppose the set of
nodes to be systematic are the first nodes. Let

be the set of helper nodes. Recall that [from (3)],

. Further, since all the
at the MSR point we have
message symbols should be recoverable from any subset of
nodes, it must be that any subset of nodes does not store
, be an -length
any redundant information. Let ,
vector denoting the symbols stored in node . Then, from
the above argument, it is clear that any symbol in the system
symbols in
can be written as a linear combination of the
.
, denote the symbol passed by node to
Let ,
assist in the repair of node . Then we can write
(64)
for some vectors
and
each of length . The symbols
in
have no redundancy among themselves.
are undeThus, the components comprising of
sired and hence are termed as interference components, and the
component comprising of is termed the desired component.
It is shown in [6], [14] that for any MSR code, it must be that
, the set of vectors
for every
(65)

denote a set of linearly independent column vectors drawn

.
from the generator matrices of the first nodes
That such a subset is guaranteed to exist follows from the databe the
reconstruction property of a regenerating code. Let
invertible matrix

are aligned (i.e., are scalar multiples of each other).

The following lemma considers the repair scenario discussed
above to illustrate how interference alignment arises in the
product-matrix MSR code presented in Section V.
Lemma 11: For every helper node ,
exist scalars
and an
such that

Then we have the relation

, there
-length vector

(63)
is the corresponding set of code symbols. It
where
follows that if we wish to encode in such a way that the code
, the input
is systematic with respect to code symbols
to be fed to the generator matrix is

(66)
Proof: Rewriting the symbols passed by the helper node
(67)
(68)

APPENDIX C
INTERFERENCE ALIGNMENT IN THE
PRODUCT-MATRIX MSR CODE
The concept of interference alignment was introduced in [21]
and [22] in the context of wireless communication. This concept
was subsequently used to construct regenerating codes in [6],
[7], [11], [14]. Furthermore, [6], [14] showed that interference
alignment is in fact, a necessary ingredient of any linear MSR
code. Since the product-matrix MSR construction provided in
the present paper does not explicitly use the concept of interference alignment, a natural question that arises is how does interference alignment manifest itself in this code. We answer this
question in the present section.
Consider repair of a failed node (say, node ) in a distributed storage system employing an MSR code, and let nodes

(69)
(70)
where (68) follows from the symmetry of matrices
By construction, the values of the scalars
distinct, which allows us to write

and

.
are

(71)
-length vectors
Also, since the
are linearly independent by construction, for
such that
exist scalars

, there

(72)

RASHMI et al.: OPTIMAL EXACT-REGENERATING CODES FOR DISTRIBUTED STORAGE AT THE MSR AND MBR POINTS

From (70), (71), and (72), for any

, we can write
(73)
(74)

(75)
where (74) follows from (72), and (75) follows from (71).
REFERENCES
[1] D. A. Patterson, G. Gibson, and R. H. Katz, A case for redundant
arrays of inexpensive disks (RAID), in Proc. ACM SIGMOD Int. Conf.
Management of Data, Chicago, IL, Jun. 1988, pp. 109116.
[2] S. Rhea, P. Eaton, D. Geels, H. Weatherspoon, B. Zhao, and J. Kubiatowicz, Pond: The OceanStore prototype, in Proc. 2nd USENIX
Conf. File and Storage Technologies (FAST), 2003, pp. 114.
[3] R. Bhagwan, K. Tati, Y. C. Cheng, S. Savage, and G. M. Voelker,
Total recall: System support for automated availability management,
in Proc. 1st Conf. Networked Systems Design and Implementation
(NSDI), 2004.
[4] A. G. Dimakis, P. B. Godfrey, M. Wainwright, and K. Ramchandran,
Network coding for distributed storage systems, in Proc. 26th IEEE
Int. Conf. Computer Communications (INFOCOM), Anchorage, AK,
May 2007, pp. 20002008.
[5] Y. Wu, A. G. Dimakis, and K. Ramchandran, Deterministic regenerating codes for distributed storage, in Proc. 45th Annu. Allerton
Conf. Control, Computing, and Communication, Urbana-Champaign,
IL, Sep. 2007.
[6] N. B. Shah, K. V. Rashmi, P. V. Kumar, and K. Ramchandran, Explicit codes minimizing repair bandwidth for distributed storage, in
Proc. IEEE Information Theory Workshop (ITW), Cairo, Egypt, Jan.
2010.
[7] C. Suh and K. Ramchandran, Exact-repair MDS codes for distributed
storage using interference alignment, in Proc. IEEE Int. Symp. Information Theory (ISIT), Austin, TX, Jun. 2010, pp. 161165.
[8] Y. Wu, Existence and construction of capacity-achieving network
codes for distributed storage, IEEE J. Select. Areas Commun., vol.
28, no. 2, pp. 277288, Feb. 2010.
[9] A. G. Dimakis, P. B. Godfrey, Y. Wu, M. Wainwright, and K. Ramchandran, Network coding for distributed storage systems, IEEE
Trans. Inf. Theory, vol. 56, no. 9, pp. 45394551, Sep. 2010.
[10] A. Duminuco and E. Biersack, A practical study of regenerating codes
for peer-to-peer backup systems, in Proc. 29th IEEE Int. Conf. Distributed Computing Systems (ICDCS), Jun. 2009, pp. 376384.
[11] Y. Wu and A. Dimakis, Reducing repair traffic for erasure
coding-based storage via interference alignment, in Proc. IEEE
Int. Symp. Information Theory (ISIT), Seoul, South Korea, Jul. 2009,
pp. 22762280.
[12] K. V. Rashmi, N. B. Shah, P. V. Kumar, and K. Ramchandran, Explicit construction of optimal exact regenerating codes for distributed
storage, in Proc. 47th Annu. Allerton Conf. Communication, Control,
and Computing, Urbana-Champaign, IL, Sep. 2009, pp. 12431249.

5239

[13] D. Cullina, A. G. Dimakis, and T. Ho, Searching for minimum storage

regenerating codes, in Proc. 47th Annu. Allerton Conf. Communication, Control, and Computing, Urbana-Champaign, IL, Sep. 2009.
[14] N. B. Shah, K. V. Rashmi, P. V. Kumar, and K. Ramchandran, Interference alignment in regenerating codes for distributed storage: Necessity and code constructions, IEEE Trans. Inf. Theory, submitted for
publication.
[15] Y. Wu, A construction of systematic MDS codes with minimum repair
bandwidth, IEEE Trans. Inf. Theory, submitted for publication.
[16] V. R. Cadambe, S. A. Jafar, and H. Maleki, Distributed Data Storage
with Minimum Storage Regenerating CodesExact and Functional
Repair are Asymptotically Equally Efficient [Online]. Available:
arXiv:1004.4299 [cs.IT]
[17] C. Suh and K. Ramchandran, On the Existence of Optimal Exact-Repair MDS Codes for Distributed Storage [Online]. Available:
arXiv:1004.4663 [cs.IT]
[18] N. B. Shah, K. V. Rashmi, and P. V. Kumar, A flexible class of regenerating codes for distributed storage, in Proc. IEEE Int. Symp. Information Theory (ISIT), Austin, TX, Jun. 2010, pp. 19431947.
[19] N. B. Shah, K. V. Rashmi, P. V. Kumar, and K. Ramchandran, Distributed storage codes with repair-by-transfer and non-achievability of
interior points on the storage-bandwidth tradeoff, IEEE Trans. Inf.
Theory, submitted for publication.
[20] D. S. Bernstein, Matrix Mathematics: Theory, Facts, and Formulas
With Application to Linear Systems Theory. Princeton, NJ: Princeton
University Press, 2005.
[21] M. Maddah-Ali, A. Motahari, and A. Khandani, Communication
over MIMO X channels: Interference alignment, decomposition, and
performance analysis, IEEE Trans. Inf. Theory, vol. 54, no. 8, pp.
34573470, Aug. 2008.
[22] V. Cadambe and S. Jafar, Interference alignment and spatial degrees
of freedom for the k user interference channel, IEEE Trans. Inf.
Theory, vol. 54, no. 8, pp. 34253441, Aug. 2008.

K. V. Rashmi received the M.E. degree from the Indian Institute of Science
(IISc), Bangalore, in 2010.
Her research interests include coding theory, information theory, networks,
communications and signal processing, with a current focus on coding for data
storage networks and network coding.

Nihar B. Shah received the M.E. degree from the Indian Institute of Science
(IISc), Bangalore, in 2010.
His research interests include coding and information theory, algorithms, and
statistical inference.
Mr. Shah is a recipient of the Prof. S.V.C. Aiya Medal for the best master-ofengineering student in the ECE Department at IISc, 2010.

P. Vijay Kumar (S80M82SM01F02) received the B.Tech. and M.Tech.

degrees from the Indian Institutes of Technology (Kharagpur and Kanpur) ,and
the Ph.D. degree from the University of Southern California (USC) in 1983, all
in electrical engineering.
From 1983 to 2003, he was on the faculty of the EE-Systems Department at
USC. Since 2003, he has been on the faculty of the Indian Institute of Science,
Bangalore, and also holds the position of adjunct research professor at USC.
His current research interests include codes for distributed storage, distributed
function computation, sensor networks and space-time codes for MIMO and
cooperative communication networks.
Dr. Kumar is a an ISI highly-cited author. He is co-recipient of the 1995 IEEE
Information Theory Society prize paper award as well as of a best paper award
at the DCOSS 2008 conference on sensor networks.