Benchmarking Ampere’s ARM CPU in Google Cloud Platform

While creating an instance today I noticed GCP offers ARM based CPUs made by Ampere, a company based in Santa Clara with a large office in Portland. The monthly cost runs about $30/mo for a single CPU with 4 GB RAM – a bit pricier than comparable N1, but slightly less than a comparable T2D, which is the ultra-fast AMD EPYC Milan platform.

Since I mostly run basic Debian packages and python scripts, CPU platform really wasn’t an issue, so I was curious to have a quick bake-off using a basic 16 thread sysbench test to mimic a light to moderate load. Here’s the results


These are based on Ampere Altra and cost $29/mo in us-central1

CPU speed:
    events per second:  3438.95

General statistics:
    total time:                          10.0024s
    total number of events:              34401

Latency (ms):
         min:                                    0.28
         avg:                                    4.63
         max:                                   80.31
         95th percentile:                       59.99
         sum:                               159394.13

Threads fairness:
    events (avg/stddev):           2150.0625/4.94
    execution time (avg/stddev):   9.9621/0.03


These are based on AMD Milan and cost $32/mo in us-central1

CPU speed:
    events per second:  3672.67

General statistics:
    total time:                          10.0027s
    total number of events:              36738

Latency (ms):
         min:                                    0.27
         avg:                                    4.34
         max:                                  100.28
         95th percentile:                       59.99
         sum:                               159498.26

Threads fairness:
    events (avg/stddev):           2296.1250/3.24
    execution time (avg/stddev):   9.9686/0.02


These are based on Intel Skylake and cost $25/mo in us-central1

Prime numbers limit: 10000

Initializing worker threads...

Threads started!

CPU speed:
    events per second:   913.60

General statistics:
    total time:                          10.0072s
    total number of events:              9144

Latency (ms):
         min:                                    1.08
         avg:                                   17.45
         max:                                   89.10
         95th percentile:                       61.08
         sum:                               159544.06

Threads fairness:
    events (avg/stddev):           571.5000/1.00
    execution time (avg/stddev):   9.9715/0.03


These are based on AMD EPYC and cost $44/mo in us-central1

CPU speed:
    events per second:  1623.41

General statistics:
    total time:                          10.0046s
    total number of events:              16243

Latency (ms):
         min:                                    0.89
         avg:                                    9.82
         max:                                   97.24
         95th percentile:                       29.19
         sum:                               159485.50

Threads fairness:
    events (avg/stddev):           1015.1875/3.13
    execution time (avg/stddev):   9.9678/0.02


These are based in Intel Cascade Lake and cost $50/mo in us-central1

CPU speed:
    events per second:  1942.56

General statistics:
    total time:                          10.0036s
    total number of events:              19435

Latency (ms):
         min:                                    1.01
         avg:                                    8.21
         max:                                   57.04
         95th percentile:                       29.19
         sum:                               159499.92

Threads fairness:
    events (avg/stddev):           1214.6875/8.62
    execution time (avg/stddev):   9.9687/0.02


These are based on availability and have 1-2 shared CPU cores and cost $25/mo in us-central1

CPU speed:
    events per second:  1620.67

General statistics:
    total time:                          10.0055s
    total number of events:              16217

Latency (ms):
         min:                                    0.85
         avg:                                    9.84
         max:                                   65.18
         95th percentile:                       29.19
         sum:                               159647.07

Threads fairness:
    events (avg/stddev):           1013.5625/3.43
    execution time (avg/stddev):   9.9779/0.02


Amphere’s ARM CPUs offered slightly lower performance against the latest goodies from AMD. It did however beat it in the bang for buck ratio thanks to costing $1 less per month to run.

But, the key take away is both platforms completely blow away the older CPU platforms from Intel. Here’s some nice little charts visualizing the numbers.


Migrating Terraform State Files to Workspaces in an AWS S3 Bucket

Just as I did with GCP a few weeks ago, I needed to circle back and migrate my state files to a cloud storage bucket. This done mainly to centralize the storage location automatically and thus lower the chance of a state file loss or corruption.

Previously, I’d been separating the state files using the -state parameter. I then use a different input file and state file for each environment like this:

terraform apply -var-file=env1.tfvars -state=env1.tfstate
terraform apply -var-file=env2.tfvars -state=env2.tfstate
terraform apply -var-file=env3.tfvars -state=env3.tfstate

To instead store the state files in an AWS S3 bucket, create a file with this content:

terraform {
  backend "s3" {
    bucket               = "my-bucket-name"
    workspace_key_prefix = "tf-state"
    key                  = "terraform.tfstate"
    region               = "us-west-1"

This will use a bucket named ‘my-bucket-name’ in AWS region us-west-1. Each workspace will store its state file in tfstate/<WORKSPACE_NAME>/terraform.tfstate

Note: if workspace_key_prefix is not specified, the directory ‘env:‘ will be created and used.

Since the backend has changed, I have to run this:

terraform init -reconfigure

I then have to copy the local state files to the correct location that the workspace will be using. This is easiest done with the AWS CLI tool, which will automatically create the sub-directory if it doesn’t exist.

aws s3 cp env1.tfstate s3://my-bucket-name/tf-state/env1/terraform.tfstate
aws s3 cp env2.tfstate s3://my-bucket-name/tf-state/env2/terraform.tfstate
aws s3 cp env3.tfstate s3://my-bucket-name/tf-state/env3/terraform.tfstate

Now I’m ready to run an terraform plan/apply and verify the new state file location is being used.

$ terraform workspace new env1
Created and switched to workspace "env1"!

$ terraform apply -var-file=env1.tfvars

No changes. Your infrastructure matches the configuration.

Terraform has compared your real infrastructure against your configuration and found no differences, so no changes are needed.

Apply complete! Resources: 0 added, 0 changed, 0 destroyed.

$ terraform workspace new env2
Created and switched to workspace "env2"!

$ terraform apply -var-file=env2.tfvars

No changes. Your infrastructure matches the configuration.

Terraform has compared your real infrastructure against your configuration and found no differences, so no changes are needed.

Apply complete! Resources: 0 added, 0 changed, 0 destroyed.

An alternate way to do this migration is enable workspaces first, then migrate the backend to S3.

$ terraform workspace new env1
Created and switched to workspace "env1"!

$ mv env1.tfstate terraform.tfstate.d/env1/terraform.tfstate

$ terraform apply -var-file=env1.tfvars

No changes. Your infrastructure matches the configuration.

Terraform has compared your real infrastructure against your configuration and found no differences, so no changes are needed.

Apply complete! Resources: 0 added, 0 changed, 0 destroyed.

Then create the file and run terraform init -reconfigure. You’ll then be prompted to move the state files to S3:

$ terraform init -reconfigure
Initializing modules...

Initializing the backend...
Do you want to migrate all workspaces to "s3"?

Enter a value: yes

$ terraform apply -var-file=env1.tfvars

No changes. Your infrastructure matches the configuration.

Terraform has compared your real infrastructure against your configuration and found no differences, so no changes are needed.

Apply complete! Resources: 0 added, 0 changed, 0 destroyed.

A weird, ugly Error message when running

[Expert@cp-member-a:0]# $FWDIR/scripts/
GCP HA TESTER: started
GCP HA TESTER: checking access scopes...

Expecting value: line 1 column 1 (char 0)

Got this message when trying to test a CheckPoint R81.10 cluster build in a new environment. Obviously, this error message is not at all helpful in determining what the problem is. So I wrote a little debug script to try and isolate the issue:

import traceback
import gcp as _gcp 

global api
api = _gcp.GCP('IAM', max_time=20)
metadata = api.metadata()[0]

project = metadata['project']['projectId']
zone = metadata['instance']['zone'].split('/')[-1]
name = metadata['instance']['name']

print("Got metadata: project = {}, zone = {}, name = {}\n".format(project, zone, name))
path = "/projects/{}/zones/{}/instances/{}".format(project, zone, name)

    head, res ="GET",path,query=None, body=None,aggregate=False)
except Exception as e:

Running the script, I now see an exception when trying to make the initial API call:

[Expert@cp-cluster-member-a:0]# cd $FWDIR/scripts
[Expert@cp-cluster-member-a:0]# python3 ./

Got metadata: project = myproject, zone = us-central1-b, name = cp-member-a

Traceback (most recent call last):
  File "", line 18, in <module>
    head, res =,path,query=None,body=None,aggregate=False)
  File "/opt/CPsuite-R81.10/fw1/scripts/", line 327, in rest
    max_time=self.max_time, proxy=self.proxy)
  File "/opt/CPsuite-R81.10/fw1/scripts/", line 139, in http
    headers['_code']), headers, repr(response))
gcp.HTTPException: Unexpected HTTP code: 403

This at least indicates the connection to the API is OK and it’s some type of permissions issue with the account.

The CheckPoints have always been really tough to troubleshoot in this aspect, so to keep it simple, I deploy them with the default service account for the project. It’s not explicitly called out

I was able to re-enabled Editor permissions for the default service account with this Terraform code:

# Set Project ID via input variable
variable "project_id" {
  description = "GCP Project ID"
  type = string
# Get the default service account info for this project
data "google_compute_default_service_account" "default" {
  project = var.project_id
# Enable editor role for this service account
resource "google_project_iam_member" "default_service_account_editor" {
  project = var.project_id
  member  = "serviceAccount:${}"
  role    = "roles/editor"

Migrating Terraform to Workspaces & Storage Buckets

As I started using Terraform more, I quickly realized it’s beneficial to use separate state files for difference groups of resources. It goes without saying multiple environments should be in different state files, as should MSP scenarios where there’s multiple customer deployments running off the same Terraform code. The main benefit is to reduce blast radius if something goes wrong, but the additional benefit is limiting dependencies and improving performance.

So when running Terraform, I’d end up doing these steps:

git pull
terraform init
terraform plan -var-file="env1.tfvars" -state="env1.tfstate"
terraform apply -var-file="env1.tfvars" -state="env1.tfstate"
terraform plan -var-file="env2.tfvars" -state="env2.tfstate"
terraform apply -var-file="env2.tfvars" -state="env2.tfstate"
git add *.tfstate *.tfstate.backup
git commit -m "updated state files"
git push

This works OK, but isn’t ideal for a couple reasons. First, the state file can’t be checked out and updated by two users at the same time – git would try and merge the two files, which would likely result in corruption. Also, state files can contain sensitive information like passwords, and really shouldn’t be stored in the repo at all.

So the better solution is store in a Cloud Storage bucket, such as AWS S3 or Google Cloud Storage. This is usually configured by a file that specifies the bucket name and directory prefix for storing state files and looks something like this:

terraform {
  backend "gcs" {
    bucket = "my-gcs-bucket-name"
    prefix = "terraform"

After creating this file, we must run terraform init to initialize the new backend:

terraform init
Initializing modules...

Initializing the backend...

Successfully configured the backend "gcs"! Terraform will automatically
use this backend unless the backend configuration changes.

But now if we run terraform with the -state parameter, it will look for the state file in the bucket, not find it, and determine it needs to re-create everything, which is incorrect.

The solution to this problem is use a different workspace for each state file.

terraform workspace list
* default

terraform workspace new env1
Created and switched to workspace "env1"!

You're now on a new, empty workspace. Workspaces isolate their state,
so if you run "terraform plan" Terraform will not see any existing state
for this configuration.

Terraform will now look in the bucket for terraform/env1.tfstate, but that file is still local. So we must manually copy it over:

gsutil copy env1.tfstate gs://my-gcs-bucket/terraform/

Repeat this process for all state files. Now, when we run terraform plan/apply, there is no need to specify the state file. It’s automatically known. And assuming we’ve made no changes, terraform should report no changes required.

terraform workspace select env1
terraform apply -var-file="env1.tfvars"

No changes. Your infrastructure matches the configuration.

Apply complete! Resources: 0 added, 0 changed, 0 destroyed.

terraform workspace select env2
terraform apply -var-file="env2.tfvars"

No changes. Your infrastructure matches the configuration.

Apply complete! Resources: 0 added, 0 changed, 0 destroyed.

And it’s all good

Making Async Calls to Google Cloud Storage

I have a script doing real-time log analysis, where about 25 log files are stored in a Google Cloud Storage bucket. The files are always small (1-5 MB each) but the script was taking over 10 seconds to run, resulting in slow page load times and poor user experience. Performance analysis showed that most of the time was spent on the storage calls, with high overhead of requesting individual files.

I started thinking the best way to improve performance was to make the storage calls in an async fashion so as to download the files in parallel. This would require a special library capable of making such calls; after lots of Googling and trial and error I found a StackOverFlow post which mentioned gcloud AIO Storage. This worked very well, and after implementation I’m seeing a 125% speed improvement!

Here’s a rundown of the steps I did to get async working with GCS.)

1) Install gcloud AIO Storage:

pip install gcloud-aio-storage

2) In the Python code, start with some imports

import asyncio
from gcloud.aio.auth import Token
from import Storage

3) Create a function to read multiples from the same bucket:

async def IngestLogs(bucket_name, file_names, key_file = None):

    SCOPES = [""]
    token = Token(service_file=key_file, scopes=SCOPES)
    async with Storage(token=token) as client:
        tasks = (, _) for _ in file_names)
        blobs = await asyncio.gather(*tasks)
    await token.close()
    return blobs

It’s important to note that ‘blobs’ will be a list, with each element representing a binary version of the file.

4) Create some code to call the async function. The decode() function will convert each blob to a string.

def main():

    bucket_name = "my-gcs-bucket"
    file_names = {
       'file1': "path1/",
       'file2': "path2/file2.def",
       'file3': "path3/file3.ghi",
    key = "myproject-123456-mykey.json" 

    blobs =, file_names.values(), key_file=key))

    for blob in blobs:
        # Print the first line from each blob

I track the load times via NewRelic synthetics, and it showed a 300% performance improvement!

Using GCP Python SDK for Network Tasks

Last week, I finally got around to hitting the GCP API directly using Python. It’s pretty easy to do in hindsight. Steps are below

If not done already, install PIP. On Debian 10, the command is this:

sudo apt install python3-pip

Then of course install the Python packages for GCP:

sudo pip3 install google-api-python-client google-cloud-storage

Now you’re ready to write some Python code. Start with a couple imports:

#!/usr/bin/env python3 

from googleapiclient import discovery
from google.oauth2 import service_account

By default, the default compute service account for the VM or AppEngine will be used for authentication. Alternately, a service account can be specific with the key’s JSON file:

KEY_FILE = '../mykey.json'
creds = service_account.Credentials.from_service_account_file(KEY_FILE)

Connecting to the Compute API will look like this. If using the default service account, the ‘credentials’ argument is not required.

resource_object ='compute', 'v1', credentials=creds)

All API calls require the project ID (not name) be provided as a parameter. I will set it like this:

PROJECT_ID = "myproject-1234"

With the connection to the API established, you can now run some commands. The resource object will have several methods, and in each there will typically be a list() method to list the items in the project. The execute() at the end is required to actually execute the call.

_ = resource_object.firewalls().list(project=PROJECT_ID).execute()

It’s important to note the list().execute() returns a dictionary. The actual list of items can be found in key ‘items’. I’ll use the get() method to retrieve the values for the ‘items’ key, or use an empty list if ‘items’ doesn’t exist. Here’s an example

firewall_rules = _.get('items', [])
print(len(firewall_rules), "firewall rules in project", PROJECT_ID)
for firewall_rule in firewall_rules:
    print(" -", firewall_rule['name'])

The API reference guide has a complete list of everything that’s available. Here’s some examples:

firewalls() - List firewall rules
globalAddresses() - List all global addresses
healthChecks() - List load balancer health checks
subnetworks() - List subnets within a given region
vpnTunnels() - List configured VPN tunnels

Some calls will require the region name as a parameter. To get a list of all regions, this can be done:

_ = resource_object.regions().list(project=PROJECT_ID).execute()
regions = [region['name'] for region in _.get('items', [])]

Then iterate through each region. For example to list all subnets:

for region in regions:
    _ = resource_object.subnetworks().list(project=PROJECT_ID,region=region).execute()
    print("Reading subnets for region", region ,"...")
    subnets = _.get('items', [])
    for subnet in subnets:
        print(" -", subnet['name'], subnet['ipCidrRange'])

Deploying a Container to GCP Cloud Run

Copy a working container and review the Dockerfile

git clone
cd network-automation/docker/flask2-sqlchemy
cat Dockerfile

Verify logged in to gcloud and set to correct project:

gcloud projects list
gcloud config set project <PROJECT_ID>

Where <PROJECT_ID> is the Project ID

Then build the container and upload the image to Google Container Registry. Note this step is missing from the quickstart guide

gcloud builds submit --tag<PROJECT_ID>/flask2-sqlalchemy .

Now pick a region and deploy the container. In this example ‘flask2’ is the service name:

gcloud config set run/region us-central1
gcloud run deploy flask2 --image<PROJECT_ID>/flask2-sqlalchemy --allow-unauthenticated

VPNs to Google Cloud Platform (GCP) when FortiGate is behind a NAT gateway

Ran in to problems getting a VPN up and running between GCP and a FortiGate 60-E that was behind a NAT gateway (with ports udp/500 + udp/4500 forwarded). On the GCP side, these messages would show up in the logs:

remote host is behind NAT
generating IKE_AUTH response 1 [ N(AUTH_FAILED) ]
looking for peer configs matching GCP.VPN.GATEWAY.IP[%any]...[]

This error means that GCP connected to the Peer VPN gateway successfully, but it in the IKEv2 headers, it identified itself by the private IP rather than the expected public one. AWS is not picky about this, but with GCP, the Peer VPN gateway must identify itself by using the same external IP address of the NAT device.

Most vendors have long supported an option to manually override the IP address for such scenarios. In Cisco IOS or IOS-XE, this can be controlled in the IKEv2 profile with the identity local address option:

crypto ikev2 profile GCP_IKEV2_PROFILE
 match address local interface GigabitEthernet1
 identity local address MY.PUBLIC.IP.ADDRESS
 authentication remote pre-share
 authentication local pre-share
 keyring local GCP_KEYRING
 lifetime 36000
 dpd 20 5 on-demand

With Palo Alto, this is configured in the IKE Gateway, Local Identification field:

For the sake of argument, we’ll say that CheckPoint uses the “Statically NATed IP” field to influence Local ID, although this doesn’t actually work.

Fortigate does offer “Local ID” field in version 6.4.6 and higher, under the Phase 1 proposal:

Seems nice and straightforward, but even after changing this setting, the VPN tunnel still won’t establish. Logs on the GCP end change slightly and now show this:

looking for peer configs matching GCP.VPN.GATEWAY.IP[%any]...[]

The private IP is no longer showing, so it seems the issue should be solved. Instead, GCP reports a “Peer not responding” message. The Fortigate actually reports Phase 1 success, waits a few seconds, and then starts the negotiation all over. So not very helpful.

I configured a test VPN between the FortiGate and a Palo Alto, which then gave a very specific and extremely useful error message:

IKE phase-1 negotiation is failed. When pre-shared key is used, peer-ID must be type IP address. Received type FQDN

Now this explains the problem! Even though the FortiGate is sending the correct IP address in the IKEv2 header, it’s being sent as the wrong identity type. The 5 identity types are listed in RFC 7815:

  • ID_IPV4_ADDR = 32 bit IPv4 address
  • ID_IPV6_ADDR = 128 bit IPv6 address
  • ID_FQDN = DNS hostname
  • ID_RFC822_ADDR = e-mail address
  • ID_KEY_ID = octet stream

If Fortigate were smart, it would either default to IPv4 address type or auto-determine this based on the text inputted in to the field. But it seems to simply default to FQDN. Oddly, there is a CLI option called “localid-type” under the Phase1-interface that clear is intended to provide this functionality:

FGT60E1234567890 # config vpn ipsec phase1-interface

FGT60E1234567890 (phase1-interface) # edit gcp

FGT60E1234567890 (gcp) # set localid-type 
auto         Select ID type automatically.
fqdn         Use fully qualified domain name.
user-fqdn    Use user fully qualified domain name.
keyid        Use key-id string.
address      Use local IP address.

But, similar to CheckPoint, it just doesn’t work, and can be considered a broken feature.

Since GCP does not support FQDN authentication, VPNs between GCP and FortiGates behind a NAT are not possible at this time.

Enabling Private Google Access in Google Cloud Platform

By default, calls to the various Google Cloud APIs will resolve to a random Google-owned IP, and require outbound internet access, either via external IP, Cloud NAT, or 3rd party network appliance.

If outbound Internet is not required for the application, or not desired for security reasons, enabling Private Google Access allows VM instances to connect an internally routed prefix.

  1. On the subnet, turn on Private Google Access via the radio button
  2. By default, all egress traffic is permitted. If egress traffic is being denied deliberately, create a rule allowing egress traffic to destination, tcp ports 80 and 443
  3. Create a Private DNS zone called, and apply it to any networks that will use Google Private Access.

In the DNS zone, create two entries:

An A record called ‘private’ that resolves to the following 4 IP addresses:


A wildcard record ‘*’ that points to hostname ‘private’

The zone will look like this once created.

Private Google Access should now be working. To test it, ping and it should resolve to one of those 4 IP addresses

PING ( 56(84) bytes of data.

Getting started with Flask and deploying Python apps to Google App Engine

Installing Flask on Linux or Mac

On Debian 10 or Ubuntu 20:

sudo pip3 install flask flask-cors

On Mac or FreeBSD:

sudo pip install flask flask-cors

Creating a basic flask app:

#!/usr/bin/env python3

from flask import Flask, request, jsonify

app = Flask(__name__)

@app.route("/", defaults = {'path': ""})

def index(path):
    req_info = {
        'path': "/" + path,
        'query_string': request.args,
        'remote_addr': request.environ.get('HTTP_X_REAL_IP', request.remote_addr),
        'user_agent': request.user_agent.string
    return jsonify(req_info)

if __name__ == '__main__':

Run the app

chmod u+x
Running on (Press CTRL+C to quit)

Do a test curl against it

$ curl -v "http://localhost:5000/oh/snap?x=1&x=2"

< HTTP/1.0 200 OK
< Content-Type: application/json
< Content-Length: 65
< Access-Control-Allow-Origin: *
< Server: Werkzeug/1.0.1 Python/3.7.8
< Date: Wed, 21 Apr 2021 17:07:58 GMT

Deploying to Google Cloud App Engine

Create a requirements.txt file:

echo "flask" > requirements.txt

Create an app.yaml file:

printf "runtime: python38\nenv: standard\n" > app.yaml 

Now deploy the app to Google using the gCloud command:

gcloud app deploy