NAME : ADI ASIJA

REG. No. : 19BEC0247

SUBJECT : DIGITAL SIGNAL PROCESSING
SLOT : C1
TOPIC : Report on high dimensional coded ¿

What is the problem?

Distributed computing is at the heart of the big data thanks to the unprecedented increase in storage
and computing power.Distributed systems can be characterized by various sources of individual node
delays and latencies related to queuing, computation and communicating, these can not be tracker or
controlled at a fine scale.
The unpredictable individual component latencies result in a new kind of noise called as system noise
that we needs to be addressed in order to provide delay guarantees in large scale distributed
computing systems.

What is coded computation?

Coded computation is a principled framework to address the redundancies in a distributed computing
system in order to make them robust towards slower computing nodes or stranglers.

What does this report discuss?

This report would focus on a proposed method called as irregular product coded computation.
The target problem is as follows, a matrix multiplication over N distributed workers is considered I.e.
we aim to compute AT . B for for A∈ ℝ s x d and B ∈ ℝ s x t . The size of the input matrices are
considered to be fairly large this required to be spilt into small chunks in order avoid computations
that are physically not feasible in a single machine. Input matrices are split into K blocks as
A=[A1,A2……..Ak] and
B=[B1,…….,Bk] where Ai and Bi are of the dimensions Ai ∈ ℝ s x d / k and Bi ∈ ℝ s x t / k . The sub matrices
A1,A2……..Ak . The goal is to find a proper design for mapping function from the block of input matrices A1,
A2, . . . , Ak and Bˆ 1 , Bˆ 2 , . . . , Bˆ N , respectively. One possible direction is a design for minimizing
the total execution time which is a function of the computation and decoding time. Inspired by the
decoding efficiency of the product code and its potential to improve the computation time, we specifically
consider the irregular-product-coded computation. Our main goal is to analyze the latency of such
irregular-product-coded computation

What is the novel contribution of the paper reviewed?

A novel coded multiplication scheme is proposed based on irregular product codes. The said scheme
is then analysed on the basis of expected computation time.The results show that in a certain
regimem the irregular product coded communication have a better performance in terms of the total
execution time, defines as a function of the computation time and decoding time, compared to the
existing other schemes.

What is the definition of a irregular product code and how is it computed ?

For positive integers m, n and given sequences {Ai}i ∈[m] and {bj}j ∈[n], the {(n, ai)}i ∈[m]×{(m,
bj)}j∈[n] irregular product code C is a length m*n code that consists of m row codes (n, ai) i ∈[m]
and n column codes (m, bj) j∈[n]. That is, the codewords of are represented by the array of
m rows and n columns in which row i is a codeword in he (n, ai) code and column j is a codeword in
the (m, bj) code.

The row and column codes of the irregular product code can be any codes; however, the study is
confine to the MDS constituent (row and column) codes in this paper. Then, the following theorem
provides the highest possible dimension and a corresponding achievable scheme.

1
Now considering {(n,ai)}I E [n] X {(n,bj)}j E [n] irregular product coded computation. Assume mn
Workers each of which is denoted by w(i,j) meaning that its assigned subtask is encoded using the
(n,ai) and (m,bj) MDS codes, I.e the I-th row and j-th column codes.

Encoding Algorithm:

The procedure of encoding is described in the figure

alongside. A matrix C∈ ℝ m x nas a data collection
matrix is defined Cij:= 1 if the worker w(i,j) is already
assigned its sub-matrices otherwise Cij:= 0

Fig-1 shows the encoding and decoding of the irregualr

product matrix for given parameters n=m=4 , kc = 6 ,
|a1, a2 , a3, a4|=[2,3,3,3] and |b1, b2 , b3, b4|=[2,2,3,3]

2
3