Google Compute: Optimising Costs and Performance in Google Cloud Platform

Sustained use discounts

Automatic discounts for running VM instances for significant portion of the billing month

• Example: If you use N1, N2 machine types for more than 25% of a month you get 20% to 50% discount on very incremental minute.
• Discount increses with usage.
• No action required on your part!

Applicable for instances created by Google Kubernetes Engine and Compute. Engine

RESTRCITION

• Does not apply on certain machine types (ex: E2 and A2)
• Does not apply to VMs created by App Engine flexible and Dataflow

Committed use discounts

• For workloads with predictable resource needs
• Commit for 1 year or 3 years
• Up to 70% discount based on machine type and GPUs
• Applicable for instances created by Google Kubernetes Engine and Compute Engine
• Does not apply to VMs create by App Engine flexible and Dataflow

Preemptible VM

Short-lived cheaper ( upto 80%) compute instances

• Can be stopped by GCP any time (preempted) within 24 hours
• Instances get 30 second warning (to save anything they want to save)

Use Preempt VM’s if:

• Your applications are fault tolerant
• You can very cost senstive
• Your workload is not immediate
• Example: Non immediate batch processing jobs

RESTRICTIONS:

• NOT always available
• No SLA and CANNOT be migrated to regular VMs
• No Automatic Restrats
• Free Tier credits not applicable

Spot VMs

• Spot VMs: Latest version of preemptible VMs
• Key Difference: Does not have a maximum runtime
  • Compared to traditional preemptible VMs which have a maximum runtime of 24 hours
• Other features similar to traditional preemptible VMs

Google Compute Engine – Billing

• Your are billed by the second (After a minimum of 1 minute)
• Your are NOT billed for compute when a compute instance is stopped
  • However, you will be billed for nay storage attached with in!
• (RECOMMENDATION) Always create Budget alerts and make use of Budget export to stay on top of billing!
• What are ways you can save money?
  • Choose the right machine type and image for your workload
  • Be aware of discount available
    • Sustained use discount
    • Committed use discount
    • Discount for preemptible VM instance

Compute Engine: Live Migration & Availability Policy

How do you keep your VM instances running when a host system needs too be updated (a software or a hardware update needs to be performed)?

Live Migration

• Your runing instance is migrated to another host in the some zone
• Does NOT change any attributes or properties of the VM
• SUPPORTED for instances with local SSDs
• Not supported for GPUs and preemptible instances

Important Configuration – Availability Policy:

• On host maintenance: What should happen during periodic infrastructure maintenance?
• Automatic restart – Restart VM instance if they are terminated. due to non-user initiated reasons (maintenance event, hardware failure etc.)

Compute Engine Features: Custom Machine Types

• What do you do when predefined VM options are not appropriate for your workload?
  • Create a machine type customised to your needs (a Custom Machine Type)
• Custom Machine Type: Adjust vCPUs, memory and GPUs
  • Choose between E2, N2, or N1 machine types
  • Support a wide variety of Operating Systems: CentOS, CoreOS, Debian, RedHat, Ubuntu, Window etc
• Billed per vCPUs, memory provisioned to each instance
  • Example Hoursly Price: $0.033174 / vCPU + $0.004446 /GB

Compute Engine Features: GPUs

How do you accelerate math intensive and graphics-intensive workloads for AI/ML etc?

• Add a GPU to your virtual machine:
  • High performance for math intensive and graphics-intensive workloads
  • Higher Cost
  • (REMEMBER) Use images with GPU libraries (Deep Learning) installed
    • OTHERWISE, GPU will not be used
  • GPU restrictions:
    • NOT supported on all machine types (For example, not supported on shared-core or memory optimised machine types)
    • On host maintenance can only have the value “Terminate VM instance”
• Recommended availability policy for GPUs
  • Automatic restart – on

Virtual Machine Quick Review

• Associated with a project
• Machine type availability can vary from region to region
• You can only change the machine type (adjust the number of vCPUs and memory) of a stopped instance
  • You CANNOR change the machine type of a running instance
• VM’s can be filtered by various properties
  • Name, Zone, Machine Type, internal/External IP, Network, Labels etc
• Instances are Zonal (Run in a specific zone (in a specific region))
  • Images are global (you can provide access to other project – if needed)
  • Instances templates are global (Unless you use zonal resource in your templates)
• Automatic Basic Monitoring is enabled
  • Default Metrics: CPU utilisation, Network Bytes (in/out), Disk throughput/IOPS
  • For Memory Utilisation & Disk Utilisation – Cloud Monitoring agent is needed

Virtual Machine – Best Pratice

• Choose Zone and Region based on:
  • Cost, Regulation, availability Needs, Latency and specific Hardware needs
  • Distribute instance in multiple zone and regions for high availability
• Choose right machine type for you needs:
  • Play with them to find out the right machine type
  • Use GPUs for math and Graphic intensive applications
• Reserve for “committed use discounts” for constant workloads
• Use preemptible instances for fault-tolerant, NON time critical workloads
• Use labels to indicate environment, team, Business unit etc

Compute Engine Scenarios

1. You want dedicated hardware for your compliance, licensing, and management needs
   • Answer: Sole-tenant nodes
2. I have 1000s of VM and I want to automate OS patch management, OS inventory management and OS configuration management (manage software installed)
   • Answer: Use “VM Manager”
3. You want to login to your VM instance to install software
   • Answer: You can SSH into it
4. You do not want to expose a VM to internet
   • Answer: Do NOT assign an external IP address
5. You want to allow HTTP traffic to your VM
   • Configure Firewall Rules