1. DuraStore – Achieving Highly Durable Data Centers
Lead faculty:
Yong Chen
Students:
Wei Xie
Jiang Zhou (postdoctoral researcher)
CAC faculty:
Cloud and Autonomic Computing Center &
Data-Intensive Scalable Computing Laboratory
Presenter:
Wei Xie

2. Data-Intensive Scalable Computing Laboratory
Faculty/staff members
Dr. Yong Chen
Dr. Yu Zhuang
Dr. Dong Dai
Dr. Jiang Zhou
Ph.D. student members
Mr. Ghazanfar Ali
Ms. Elham Hojati
Mr. John Leidel
Ms. Neda Tavakoli
Mr. Xi Wang
Mr. Wei Xie
Mr. Wei Zhang
Masters/UG student members
Ms. Priyanka Kumari, Mr. Ali Nosrati, Mr. Yuan Cui, Mr. Frank Conlon, Ms. Yan Mu, Mr. Zachary Hansen, Mr. Yong Wu, Mr. Alex Weaver
Website:
Mission statement
Broad research interests in parallel and distributed computing, high-performance computing, cloud computing with a focus on building scalable computing systems for data-intensive applications
Projects (over $2M external grants in past 5 years, from NSF/DOE/DOD)
OpenSoC HPC: Open Source, Extensible High Performance Computing Platform (DOD/LBNL)
Compute on Data Path: Combating Data Movement in HPC (NSF)
Unistore: A Unified Storage Architecture (NSF-IUCRC/Nimboxx)
Development of a Data-Intensive Scalable Computing Instrument (NSF)
Decoupled Execution Paradigm for Data-Intensive High-End Computing (NSF)
Active Object Storage for Big Data Applications (DOE/ANL)
in high-performance scientific computing/high-end enterprise computin

3. Outline Project Goals, Motivations, Challenges Project Overview
Project team members
Background and Related Research
Overview of project tasks
Activities and outcomes
Deliverables and benefits
LIFE form input

4. Project Goals, Motivation, Challenges
Design a data center storage system that is significantly more durable under correlated failure events.
Motivations
Correlated failure events (e.g. site-wide power outage)
Random replication used in mainstream storage system has high data loss probability
It takes high cost to recover data loss resulted from correlated failures
Both industry and research groups reported this problem
Challenges
Recently emerged durability aware data replication scheme improves durability but sacrifices load balance, scalability, and many other features of storage systems

5. Project Overview
Core component: durability aware data replication/erasure coding
Use copyset and combination theory to minimize the correlated failure data loss probability
Compatible with existing data replication/erasure coding scheme with plug-in replacement
Data store software and storage infrastructure should be modified to be compatible
Durable distributed/parallel data store
(HDFS, GFS, RAMCloud, Ceph)
Durability aware
data replication/erasure coding
Durable storage cluster

6. Project Team Members Faculty Students Expertise Dr. Yong Chen
Assistant Professor of CS, Texas Tech University
Dr. Jiang Zhou
Postdoctoral Researcher of CS, Texas Tech University
Students
Wei Xie
CS Ph.D. Candidate, Texas Tech University
Expertise
Storage systems
Parallel/distributed file systems
Cloud software stack
Data models and data management

7. Background and Related Research
Correlated failures in Data Centers
In a cluster-wide failure event such as power outage occurs, about 0.5-1% of the nodes fail to reboot (from report of Yahoo! and LinkedIn)
Long fixed time to recover lost data (locating lost data)
Copyset Data Replication Scheme
A copyset is a set of servers that a data chunk is replicated to
The total number of unique copysets in the system determines the data loss probability under correlated failure
Minimize the total number of unique copysets could significantly reduce the data loss probability (100% to 0.1%)
Geo-replication
Replication on remote site to achieve better durability

8. Overview of Project Tasks
Task 1: Analysis and modeling of mainstream data stores in terms of data durability
Task2: Design a new data replication or erasure coding scheme
Task 3: Apply the proposed data replication or erasure coding scheme on the targeted system
Task 4: Test the durability and conduct experiments to evaluate the load balance, scalability, and overhead of the new scheme

9. Task 1: Analysis and Modeling
Task 1: Analysis and modeling of mainstream data stores in terms of data durability
Data redundancy scheme (replication or erasure coding)
Correlated failure durability model

10. Task 2: Design
Task2: Design a new data replication or erasure coding scheme
Enhance data durability without sacrificing load balance and scalability
For load balance and scalability, a consistent hashing based replication is a good candidate
It is challenging to improve the durability of consistent hashing based replication
Solutions will be developed based upon extensive prior R&D

11. Task 3: Implementation
Task 3: Apply the proposed data replication or erasure coding scheme on the targeted system
Target systems: Sheepdog, Ceph, and HDFS
A prototype system with the newly proposed durable data replication or erasure coding scheme is planned
The code developed will be free to use

12. Task 4: Evaluation
Task 4: Test the durability and conduct experiments to evaluate the load balance, scalability, and overhead of the new scheme
Small-scale test will be conducted
Large-scale simulation to complement the test

13. Activities and outcomes
Preliminary study and prior R&D in this space
Modeling of mainstream data store systems
Outcomes
Knowledge and findings from a systematic study on the correlated failure durability
Mathematical model and a simulator for simulating the data loss in data centers under correlated failures
New data replication algorithm or erasure coding scheme for DuraStore and prototype implementation source codes

14. Deliverables and benefits
One or more correlated-failure durability aware data replication or erasure coding schemes for data center storage
Evaluation results from the simulation and experiment of the implementation of the proposed schemes
Report/paper presenting the design and evaluation results
Benefits
Access to the DuraStore design and prototype and free use of IP generated
Enhanced productivity and utilization of cloud storage systems
Less operational cost to data centers
Less maintenance and trouble-shooting manpower and resources for data centers
Simpler fault-tolerance design of data centers
Collaboration with faculty/staff and graduate students
Access to reports and papers
Recruitment and spin-off opportunities

15. Preliminary Results and Publications
W. Xie and Y. Chen. Elastic Consistent Hashing for Distributed Storage Systems, IPDPS’17
J. Zhou, W. Xie, D. Dai and Y. Chen. Pattern-Directed Replication Scheme for Heterogeneous Object-based Storage, CCGrid’17
J. Zhou, W. Xie, J. Noble, K. Echo and Y. Chen. SUORA: A Scalable and Uniform Data Distribution Algorithm for Heterogeneous Storage Systems, NAS’16
W. Xie, J. Zhou, M. Reyes, J. Noble and Y. Chen. Two-Mode Data Distribution Scheme for Heterogeneous Storage in Data Centers, BigData’15
Sheepdog with Elastic Consistent Hashing (IPDPS'17 paper):
Copyset consistent hashing on lib-ch-placement:

16. LIFE Form Input Please take a moment to fill out your L.I.F.E. forms.
Select “Cloud and Autonomic Computing Center”
then select “IAB” role.
What do you like about this project?
What would you change?
(Please include all relevant feedback.)