1. Distributed File Systems
Cullen Eason, Jordan Messec, Jason Propp, Richard Briglio
CS455: Distributed Systems
2015 Spring Semester

2. Why is the Problem Important?
We desire to find answers to complex and relevant problems
Parsing and aggregation of massive data-sets in a timely manner
Effective solutions bring efficiency, reliability, profitability and client trust

3. Problem Characterization
Data storage needs are increasing
Accessibility
Reliability
Consistency
Google processes more than twenty petabytes of data daily
Millions of pages added to the internet daily
Hundreds of petabytes in the future?
Upgrading within data centers
Needs simplicity in expansion/replacement while keeping services available
Issues with standard file systems
Files exist on separate servers
Must be accessed through that server directly
Location matters
Load balancing
Repeated access to a file should spread across multiple servers

4. Trade-Off Space for Solutions in this Area
Reliability vs Increased Storage Requirements
Increased replication gives more reliability, results in higher storage and meta data
Ease of Access vs Single Point of Failure
Centralizing system knowledge eases use, increases dependence on Master Node
Network Efficiency vs Reliability
Locality between replicas reduces network traffic but decreases reliability
Load balancing on Master Node reduces network traffic, reduces reliability by increasing stress on Master Node
CAP: Consistency, Availability, Partition Tolerance
Only two can be present at one time

5. Dominant Approaches to the Problem
GFS
Only available to Google
Proprietary permissions system
Oldest mapreduce based DFS
Poorer namenode splitting
HDFS
Open source and broadly used
POSIX permissions system
No physical security
Lustre
Open source
Specific file system clients
Multiple simultaneous control servers
Ceph
Does not need object lists
Traffic goes directly between object storage clusters and clients

6. Dominant Approaches to the Problem
Virtual Resource Distance Management
Addresses data and resource locality
Addresses slow resources
Especially relevant with multiple data centers
Load Rebalancing Algorithms
Necessitated by hardware failures, upgrades, or usage changes
Classify chunk servers as overloaded or underloaded
Underloaded servers seek out overloaded servers to take work from
Power Usage
Data centers use large amounts of power with electricity bills reaching the billions
Power usage levels can be split into “gears” for more efficiency with low usage
RABBIT is better than PARAID for read operations, but worse for write operations

7. Insights Gleaned Issues facing DFS Solutions Trade-offs Data Locality
Single point bottlenecks and failures
Power consumption
Solutions
hardware and architecture
software and algorithms for solutions
Trade-offs
Cost benefit balance
Needs of specific developers and the DFS’ application

8. What the Problem Space in the Future Will Look Like
Vertical and Horizontal Expansion
Current fields producing data will produce more
Web is expanding rapidly, new weather centers come online
New fields entering data collection
Companies collect new fields of information, e.g. length of customer phone calls, time of day of sales
Higher volume of queries
User field will expand from large companies to individuals
Big data analysis will be the new web search
Range of companies making data requests will expand geographically and across industries
Volatile Nodes
Further integration of cell phones, laptops, glasses, watches, and other mobile devices; begin to be utilized as storage and analysis nodes in DFSs
Possibly even human augmentations or implants

9. Trade-off Space and Solutions in the Future
Future of physical fabric for data transfer
Fibre Channel’s multiple-pipes vs. Ethernet’s single-pipe
Multi-colored LEDs for optical fiber cables?
Separation of traffic
Colors define routes for data
Geographic data locality
Master data center updates other data centers
Maintaining consistent data across geographically different locations
Reduced latency for data access
Cheaper memory modules
One billion files on HDFS = 300Gb memory needed on namenode
150 bytes per file, folder, block
Hybrid storage/memory
Fast access of RAM applied to storage
Separated RAM no longer necessary