1. Dr. Zhijie Huang and Prof. Hong Jiang University of Texas at Arlington
Efficient MDS Array Codes for Enhancing Reliability of Large-scale Storage Systems
Dr. Zhijie Huang and Prof. Hong Jiang
University of Texas at Arlington

2. The Problem, Need and Industrial Relevance
Disk/Node failures are inevitable in storage systems (IBM/NetApp/EMC/Google/Amazon)
Replication (multi-copies) v.s. Erasure coding
Using erasure coding rather than replication to protect data is much more economic
Array codes : a class of efficient erasure codes that use only XOR and Cyclic Shift operations in encoding/decoding
Our another research topic is about erasure coding. As is known to everybody, disk and node failures are inevitable in storage systems. Hence, in order to protect data, either replication or erasure coding should be employed. Since using erasure coding requires much less storage overhead, it is increasingly popular in recent years. We are particularly interested in a class of erasure codes called array codes, since their encoding and decoding procedures are quite efficient in that they use only XOR and cyclic shift operations.

3. The Problem, Need and Industrial Relevance
codeword: two-dimensional array, each column corresponds to a disk/node while each element corresponds to a sector/block
codeword is composed of data elements and coding/parity elements
each coding element is computed from certain data elements through XOR operations
Quite suitable for countering Disk/Node failures
Our another research topic is about erasure coding. As is known to everybody, disk and node failures are inevitable in storage systems. Hence, in order to protect data, either replication or erasure coding should be employed. Since using erasure coding requires much less storage overhead, it is increasingly popular in recent years. We are particularly interested in a class of erasure codes called array codes, since their encoding and decoding procedures are quite efficient in that they use only XOR and cyclic shift operations.

4. Project Goals and Objectives
Performance Metrics of Array Codes:
encoding complexity: number of XORs required for computing each coding element
decoding complexity: number of XORs required for recovering each erased element
update complexity: number of coding elements that need to be updated whenever a data element is changed
I/O cost during decoding: number of elements that need to be read during the decoding procedure
When being applied to storage systems, here are several performance metrics that are usually concerned by array codes’ researchers, i.e., encoding complexity, decoding complexity, update complexity, and I/O cost during decoding. The encoding and decoding complexities determine the computational overhead of storage systems that employ array codes, while the update complexity and I/O cost during decoding affect the I/O and communication cost of the system.

5. Project Goals and Objectives
HOW TO:
Construct highly fault-tolerant array codes
Minimize the update complexity
Minimize the encoding/decoding complexity
Minimize the I/O cost and network traffic during decoding
According to the demand of storage systems, there are several important issues of array codes need to be solved, such as how to construct efficient and highly fault-tolerant array codes, how to construct array codes without code length limit, how to minimize the encoding, decoding and update complexities, and how to minimize the I/O cost during decoding. These issues are significant and challenging, and our long-term research goal is to completely solve them.

6. Research Methods and Novelty of Approach
Two typical column sizes of array codes: p-1 and (p-1)/2, p is a prime number
Extend the approach that we used to derive RΛ-Code from XI-Code
Column size
Instances
Distance
Provenance
P-1
XI-Code 4
T-COM 2016
(p-1)/2
RΛ-Code
ISIT 2017
Blaum-Roth
Any positive
T-IT 1993

7. Expected Outcomes and Deliverables
Low density MDS array codes that can correct any given number of erasures
Efficient encoding and decoding algorithms for these new codes
An erasure coding library implementing the proposed codes
Research papers published in top conferences and top journals

8. Project Timeframe and Budget
Example 1 (the longer / more expensive the project, the less competitive it might be)
2 year project duration
Year 1: Code construction and poof of correctness
Year 2: Develop of encoding/decoding algorithm
Task demonstration and evaluation
Estimated cost
$15K materials/equipment; $70K labor (PI, PostDoc)
TOTAL: $85K

9. Impact of work proposed or already performed
Lowest density MDS array codes of distance 3 (IEEE Communications Letters, 2015)
An improved decoding algorithm for generalized RDP codes (IEEE Communications Letters, 2016)
XI-Code: A practical family of lowest density MDS array codes of distance 4 (IEEE Transactions on Communications, 2016)
Efficient lowest density MDS array codes of distance 4 (IEEE International Symp. on Info. Theory, 2017)
We have got involved in this field for several years, and have obtained some achievements. In particular, we have constructed several classes of lowest density MDS array codes with optimal properties, and designed an improved decoding algorithm for the generalized RDP codes. Most of the results have been published in top journals. Since the presentation time is limited, we cannot show the technical details of these results. Nevertheless, we have prepared a poster regarded to our recent work, so please take a look at our poster later if you are interested in our research.

10. Hong.jiang@uta.edu Zhijie.huang@uta.edu
Closing remarks - benefits and contact information