Skip to content

Latest commit

 

History

History
91 lines (71 loc) · 2.92 KB

lecture03-building-cmu-cloud.md

File metadata and controls

91 lines (71 loc) · 2.92 KB
Raw

Lecture 3 Building a Carnegie Mellon Cloud

CMU's Dilemma

  • Multiple computing resources available at CMU
  • Problem: pretty much used only during deadlines
  • Monetizing spare cycles
    • Option 1: rent out entire machines
      • Problem: machines are heterogeneous, and users have different demands
    • Option 2: rent out portions of machines
      • A time, space sharing approach

The Initial CMU Cloud

the_initial_cmu_cloud

The Provisioner

  • Basic
    • Compute cores, MBs memory, GBs storage
    • Network, special hardware (GPUs?)
    • AWS creates bins of these (e.g., m1.large)
  • Optional features
    • Persistent storage
    • Failure domains and auto-scaling
    • elastic IP addresses
  • Assignment of users to machines
    • Matching: requested resource <-> available machines
    • Bin-packing problem: satisfy max possible user requests
  • Some degrees of freedom: migrating existing users
    • Cost-benefit problem: migration can increase job runtimes

The Scheduler

  • Purpose: choose which user jobs/processes to run
  • Goal 1: Prioritization
  • Goal 2: Oversubscription
    • Allocate more resources than available because most users under-utilize resources
  • Goal 3: Workload constraints
    • Gang scheduling: must co-schedule distributed software taht runs in lock-step

Encapsulation

  • Goal: encapsulation for compute, storage, networking, and data
  • We need to isolate co-located user processes
    • Use selected software/OS environment
    • Avoid issues due to environment/configuration changes
    • Avoid performance interference
    • Keep files and data private
  • Ensure that users obtain resources they paid for

Virtualization

  • Goal: multiplex single physical resource among multiple software-based resources
    • Hardware, OS -> Virtual machines
    • Networks -> VLANs
    • Disks -> Virtual disks
  • Does not solve performance interference issues

Fault Tolerance

  • For storage
    • Replication (within/across data centers)
    • Redundant Array of Independent Disks (RAID)
  • For critical infrastructure and services
    • State-machine replication, checkpointing, logging
    • Use state-free software design

CMU Cloud + Building Blocks

cmu_cloud_building_blocks

Building Blocks

  • Storage services: scalable, fault tolerant data stores
    • e.g., key-value stores, file systems, databases
  • Tools: Programming models and frameworks
    • e.g., analytics, HPC cluster, training AI models
  • Automation: Reactive systems and elastic scaling
    • e.g., monitoring and tracking, load balancers

Elastic Scaling

  • Traditional solution: provision for peak
  • The Elastic scaling approach
    • Cloud monitors load
    • Add replication instances as necessary
  • Cloud monitors users applications
    • Provides alerts when thresholds crossed
  • Cloud provides tracing libraries
    • Analyzes traces for problen root causes