Lecture Six - Data Centre Flashcards
Data Centre - Definition
A facility housing computer systems and components, including servers, storage systems, and networking equipment.
Data Centre - Purpose
Provides centralized data storage, processing, and dissemination services for organizations.
Data Centre - Importance
Essential for supporting web services, cloud computing, and large-scale computational tasks.
Data Centre - Historical Context
Data centres have evolved from simple server rooms to complex facilities.
Data Centre - Key drivers
Demand for large-scale computing services.
Applications in web search, online social networks, and scientific computations.
Data Centre - Challenges
Increasing resource requirements.
Bottlenecks in inter-node communication bandwidth.
Majority of traffic is internal within the Data Centre Network (DCN).
Data Centre Networking Basics - Components
Servers: Transition from high-end enterprise servers to low-cost, large-scale server deployments.
Networking Infrastructure: Needs to support extensive inter-server communication.
Data Centre Networking Basics - Trends
Shift towards commoditization using large numbers of affordable servers.
Aggregate reliability mitigates individual server vulnerability.
Data Centre Networking Basics - Networking Challenges
Traditional enterprise solutions are insufficient for large-scale needs.
Movement towards scalable, commodity-based networking components.
Desirable Features for Data Centre Networks
Support for a large number of servers (100k+ hosts).
High capacity and bandwidth between servers.
Full access to resource pools with flexible network allocation.
High utilization and reliability.
Backward compatibility and energy efficiency.
Scalability and cost-effectiveness.
High Performance Computing (HPC)
Designed for applications requiring significant computational power.
Examples: Financial systems and scientific computing utilizing technologies like InfiniBand and Myrinet.
High Performance Computing (HPC) - Advantages
High bandwidth and low latency.
Simple setup and low maintenance complexity.
High Performance Computing (HPC) - Disadvantages
High costs due to non-commodity equipment.
Incompatibility with TCP/IP applications.
Networking Approach
Utilizes off-the-shelf networking equipment such as Ethernet networks.
Networking Approach - Advancements
Switches with more ports and higher bandwidth capabilities.
Leverages ongoing technology improvements for cost-effective solutions.
Networking Approach - Pros
Cost-effective and backward-compatible.
Networking Approach - Cons
Challenges in building scalable and efficient Data Centre Networks (DCNs).
Conventional Data Centre Networks Architecture - Structure
Commodity servers and networking elements arranged in racks.
Conventional Data Centre Networks Architecture - Topologies
Top-of-Rack (ToR): Network switches located on top of server racks.
End-of-Row (EoR): Centralized switching for entire rows of racks.
Conventional Data Centre Networks Architecture - Architecture Levels
Core: High-level routing and backbone connectivity.
Aggregation: Intermediate layer connecting core and access levels.
Access/Edge: Connects directly to servers and devices.
Data Centre Networks Issues - Challenge in Conventional Architectures
Resource Fragmentation: Inefficient resource allocation across the network.
Server-to-Server Bandwidth: Limited by high oversubscription ratios (few:1 to hundreds:1).
Service Isolation: Difficulties in segregating services within shared environments.
TCP Incast Problem: Network congestion due to simultaneous data requests from multiple servers.
Fat Tree Concept - Core Idea
Balances uplink and downlink bandwidths, creating a non-blocking and non-interfering network
Fat Tree Concept - Architecture
Hierarchical network design resembling a tree structure, expanding bandwidth as it ascends.
Fat Tree Concept - Features
Equal uplink and downlink bandwidth.
Scalable and flexible design, suitable for large-scale data centers.
Fat Tree Topology Examples
Structure:
Pods: Groups of interconnected switches (e.g., Pod 0, Pod 1, etc.).
Layers: Composed of edge, aggregation, and core layers.
Illustration: A fat tree with 𝑘 = 4 k=4 demonstrating scalability using k-port switches.
Advantages: Offers numerous equal-cost paths, enhancing redundancy and load balancing.
Fat Tree Properties - Configurations
Utilizes k-port switches in a 3-layer structure (edge, aggregation, core).
Fat Tree Properties - Maximum Hosts
Max Hosts = (k^3)/4 where k is the number of ports
Fat Tree Properties - Example Set Up
48-port switches in 48 pods.
Each pod has 24 edge and 24 aggregation switches.
24 servers per edge switch, forming 1,152 subnets.
Fat Tree Properties - Path Redundancy
Multiple equal-cost paths between nodes in different pods.
Fat Tree Properties - Routing Challenges
OSPF (Open Shortest Path First) uses one shortest path; ECMP (Equal Cost Multi-Path) enhances scalability but poses availability issues.
Fat Tree Addressing - Addressing Scheme
Hierarchical IP addressing based on the network topology.
Fat Tree Addressing - Example
Format: 10.pod.switch.host
Subnet and Host Routing: Determines forwarding paths based on destination IP.
Fat Tree Addressing - Routing Strategy
Two-stage lookup using prefix and suffix for efficient packet delivery.
Fat Tree Addressing - Illustration
Shows routing process for packets with different destination IPs.
Fat Tree Routing - Routing Process
Stage 1: Prefix Lookup for downward direction (toward hosts).
Stage 2: Suffix Lookup for upward direction (toward core network).
Fat Tree Routing - Example Routing
Packet with destination 10.2.0.2 forwarded to port 0.
Packet with destination 10.3.0.3 forwarded to port 3.
Fat Tree Routing - Benefits
Efficient routing using hierarchical addressing and lookup tables.
Fat Tree - Advantages
Utilizes commodity network equipment.
Reduces oversubscription and resource fragmentation.
Provides load balancing capabilities across the network.
Fat Tree - Disadvantages
Requires custom addressing and routing solutions.
Complexity in cabling and network setup.
Challenges with load balancing in heterogeneous environments.
Virtual Layer-2 Network - Concept
Simulates a flat Layer-2 network for simplified inter-server communication.
Virtual Layer-2 Network - Design Features
Server and Location Separation: Distinguishes between application addresses (AAs) and location addresses (LAs).
Traffic Spreading: Uses ECMP and Valiant load balancing to distribute traffic evenly.
Virtual Layer-2 Network - Benefits
Addresses scalability and load balancing challenges in traditional Layer-2 networks.
Portland Concept - Dual MAC Addressing
AMAC: Actual MAC address for device identification.
PMAC: Pseudo MAC address encoding network topology.
Portland Concept - Fabric Manager
Manages address mappings and network agility.
Portland Concept - Routing Efficiency
PMAC routing reduces complexity in forwarding decisions.
Topology encoding aids in quick routing resolutions.
Portland Concept - Scalability
Supports large Layer-2 networks with simplified management.
Beyond Fat Trees - Alternative Approaches
Server-Centric DCNs: Use servers as relay nodes to enhance network expandability and reduce network diameter (e.g., BCube, DCell).
Optical/Electronic DCNs: Leverage optical communication for energy efficiency and reduced complexity (e.g., Helios, C-Through).
Beyond Fat Trees - Pros and Cons
Server-Centric: Expandability vs. increased wiring complexity.
Optical: Energy efficiency vs. high equipment costs.
