Skip to content

EC2

Regional and Global AWS Architecture

aws-architecture

EC2 Purchase Options (Launch Types)

On-Demand

  • Instances are isolared but multiple customer instances run on shared hardware
  • Per-second billing while an instance is running
  • Default purchase option
  • No upfront cost and discount
  • Predictable pricing
  • Apps which cannot be interrupted

Spot

  • Cheapest to launch EC2 instances
  • SPOT pricing is AWS selling unused EC2 host capacity for upto 90% discount
    • the spot price is based on the spare capacity at a given time
  • As long as the customer price is lower than the spot price, instances are provisioned else they are terminated
  • Use for non time critical workloads which are stateless and cost sensitive.
    • workloads which can be re-run
  • Never use SPOT for workloads that can't tolerate interruptions

Reserved

  • Long-term consumption of EC2 instances - at a discounted price
  • Unused Reservation still billed

reserved-instances

Scheduled Reserved instances

  • Ideal for long tern usage which which doesn't run constantly
    • Batch processing daily for 5 hours starting at 23:00
    • Weekly data, sales analysis.. every Friday for 24 hours
    • 100 hours of EC2 per month
  • Doesn't support all isntance types or regions
  • 1200 hours per year & 1 year term minimum

Capacity Reservations Reserved purchases -> on-demand -> spot (in priority order)

  • Regional Reservation - provides a bililing discount for valid instances launched in any any AZ in that region
    • but don't reserve instances within an AZ
  • Zonal Reservation - only apply to one AZ providing billing discounts and capacity reservation in that AZ
  • On-demand capacity reservations can be booked to ensure you always have access to capacity in an AZ when you need it
    • full on-demand price
    • no term limits
    • you pay regardless of if you consume it
    • no savings

Dedicated Hosts

  • Pay for HOST
  • No instance charges
  • Use case - Licensing based on Sockets/Cores where host affinity is important
  • Host affinity links instances to hosts
    • can launch multiple instances with multiple instance types
    • maximum capacity to launch the no. of instances
    • customer still have to manage host capacity

dedicated-hosts

Dedicated Instances

  • Don't own the host but do not share the host - extra cost for the instances, but dedicated hardware
  • Requirement where hosts cannot be shared.

dedicated-instances

EC2 Savings Plan

  • A hourly commitment for a 1 or 3 year term
  • A reservation of general compute \$ amounts ($20 per hour for 3 years)
    • or a specific EC2 Savings Plan - flexibility on size and OS
  • Compute products (currently) - EC2, Fargate & Lambda
  • Products have an on-demand rate and a savings plan rate
  • Resource usage consumes savings plan commitment at the reduced savings plan rate
  • Beyond the commitment, the pricing is switched to on-demand rate

EC2 Networking

Elastic network interfaces - ENI - Elastic Network Interface - Primary ENI is created when instance is created and cannot be removed and deleted when instance is terminated - Additional ENIs (secondary ENIs) can be added in and removed from other subnets(multi homed) but NOT in other AZs

WHY multiple ENIs ?

  • Security groups is attached to the ENI
  • Multiple SGs to control access to EC2 instances - different protection rules
  • Different NACL for different subnets and secondary ENIs can be placed into different subnets from subnet where primary ENI is created
  • Each EC2 has 1 primary private IPv4 address - same for lifecycle of the EC2
    • primary IP is visible to OS
    • depending on the instance type, 1 or more no. of secondary private IPs can be allocated to the instance and those IPs are also visilble to the OS
    • secondary IPs can be attached and detached from the instances
  • Public IPv4 address assigned to EC2 instance is not static and not visible to the OS
    • elastic IP per private IPv4 (to make static) - associate with primary interface
  • ENI can have 1 or more IPv6 addresses (all publicly routable)
    • 1 MAC address
    • 1 or more Security Groups
  • ENI source/destination check
    • disable them for instance instance running services such as network address translation, routing, or firewalls

Use Cases for multiple ENIs

  • Management or isolated nteworks
  • Software Licensing (MAC)
  • Security or Network Appliances
  • Multi-homes instances with workloads/roles on specific subnets
  • Low Budget & Simple High-Availability solutions

Bootstrapping vs AMI

Bootstrapping or AMI Baking are two ways to influence the speed of EC2 deployments - and how flexible those deployments can be.

  • Software installation can be extracted from bootstrapping and move to AMI
    • app, tools and languages dependencies baked into AMI
    • quick and ready to deploy - launch many instances from the AMI
  • Bootstrap running executable (binary and containers)
  • Combination of AMI baking and bootstrap for maximum efficiency

EC2 Placement Group

HPC = High Performance Computing

Cluster Placement Groups

  • Pack instances close together
  • Best Practise - launch all the instances at the same time (use same instances type)
  • Performance focused - launched in single AZ - same rack and sometimes same host

  • All members have direct connetions to each other

    • low latency and node-to-node communications - typical of HPC applications
    • use instances with high performance as well as use enhanced networking

  • Tradeoff - No resiliency since all the instances are placed in the same AZ

    • AZ is locked when launching first instance
  • Can span VPC peers - but impacts performance
  • Requires supported instance type

Spread Placement Groups

  • Strictly places a small group of instances across distinct underlying hardware(distinct rack) accross multiple AZs
    • HARD LIMIT - 7 Instances per AZ
  • Provides infrastructure isolation
    • each racnk has its own network and power source
  • Not supported for Dedicated Instances or Hosts
  • Use Case - small no. of critical instances that need to be kept sepeated from each other

Partition - groups of instances spread apart

  • Maximum of 7 partitions per AZ
    • each partition has it own rack - no sharing between partitions
  • Instances can be placed in a specific partition OR auto placed
  • Large distributed and replciated workloads such as Hadoop, Cassandra and Kafka
  • Contain the impact of failure to specific part of an application