constellation/.github/actions/e2e_kbench

K-Bench

Continuous Benchmarking

The K-Bench action runs K-Bench benchmarks on Constellation clusters. The benchmark suite records storage, network, and Kubernetes API benchmarks.

After testing, the action compares the results of the benchmarks to previous results of Constellation on the same cloud provider. That way, it is possible to evaluate performance progression throughout the development.

The data of previous benchmarks is stored in the private S3 artifact store.

In order to support encrypted storage, the action deploys the Azure CSI and GCP CSI drivers. It uses a fork of VMware's K-Bench. The fork deploys volumes that use the encrypted-storage storage class. Also, it has support to authenticate against GCP which is required to update the stored records for GKE.

Displaying Performance Progression

The action creates a summary of the action and attaches it the workflow execution log.

The table compares the current benchmark results of Constellation on the selected cloud provider to the previous records (of Constellation on the cloud provider).

The hashes of the two commits that are the base for the comparison are prepended to the table.

Example table:

• Commit of current benchmark: 8eb0a6803bc431bcebc2f6766ab2c6376500e106
• Commit of previous benchmark: 8f733daaf5c5509f024745260220d89ef8e6e440

Benchmark suite	Current	Previous	Ratio
pod_create (ms)	135	198	0.682 ⬇️
pod_list (ms)	100	99	1.01 ⬆️
pod_get (ms)	98	98	1.0 ⬆️
pod_update (ms)	187	132	1.417 ⬆️
pod_delete (ms)	119	108	1.102 ⬆️
svc_create (ms)	156	149	1.047 ⬆️
svc_list (ms)	97	96	1.01 ⬆️
svc_get (ms)	97	96	1.01 ⬆️
svc_update (ms)	100	101	0.99 ⬇️
svc_delete (ms)	143	139	1.029 ⬆️
depl_create (ms)	201	218	0.922 ⬇️
depl_list (ms)	101	101	1.0 ⬆️
depl_update (ms)	111	110	1.009 ⬆️
depl_scale (ms)	391	391	1.0 ⬆️
depl_delete (ms)	401	402	0.998 ⬇️
net_internode_snd (Mbit/s)	953.0	964.0	1.01 ⬆️
net_intranode_snd (Mbit/s)	18500.0	18600.0	1.01 ⬆️
fio_root_async_R70W30_R (MiB/s)	0.45	0.45	1.0 ⬆️
fio_root_async_R70W30_W (MiB/s)	0.20	0.20	1.0 ⬆️
fio_root_async_R100W0_R (MiB/s)	0.59	0.59	1.0 ⬆️
fio_root_async_R0W100_W (MiB/s)	1.18	1.18	1.0 ⬆️

Drawing Performance Charts

The action also draws graphs as used in the Constellation docs. The graphs compare the performance of Constellation to the performance of managed Kubernetes clusters.

Graphs are created with every run of the benchmarking action. The action attaches them to the benchmark artifact of the workflow run.

Updating Stored Records

Managed Kubernetes

One must manually update the stored benchmark records of managed Kubernetes:

AKS

Follow the Azure documentation to create an AKS cluster of desired benchmarking settings (region, instance types). If comparing against Constellation clusters with CVM instances, make sure to select the matching CVM instance type on Azure as well.

Once the cluster is ready, set up managing access via kubectl and take the benchmark:

# Setup
git clone https://github.com/edgelesssys/k-bench.git
cd k-bench && git checkout feat/constellation
./install.sh

# Remove the Constellation encrypted storage class
# Remember to revert this change before running K-Bench on Constellation!
yq 'del(.spec.storageClassName)' config/dp_fio/fio_pvc.yaml
yq 'del(.spec.storageClassName)' config/dp_netperf_internode/netperf_pvc.yml
yq 'del(.spec.storageClassName)' config/dp_network_internode/netperf_pvc.yaml
yq 'del(.spec.storageClassName)' config/dp_network_intranode/netperf_pvc.yml

# Run K-Bench
mkdir -p ./out
kubectl create namespace kbench-pod-namespace --dry-run=client -o yaml | kubectl apply -f -
./run.sh -r "kbench-AKS" -t "default" -o "./out/"
kubectl delete namespace kbench-pod-namespace --wait=true || true
kubectl create namespace kbench-pod-namespace --dry-run=client -o yaml |  kubectl apply -f -
./run.sh -r "kbench-AKS" -t "dp_fio" -o "./out/"
kubectl delete namespace kbench-pod-namespace --wait=true || true
kubectl create namespace kbench-pod-namespace --dry-run=client -o yaml |  kubectl apply -f -
./run.sh -r "kbench-AKS" -t "dp_network_internode" -o "./out/"
kubectl delete namespace kbench-pod-namespace --wait=true || true
kubectl create namespace kbench-pod-namespace --dry-run=client -o yaml |  kubectl apply -f -
./run.sh -r "kbench-AKS" -t "dp_network_intranode" -o "./out/"

# Benchmarks done, do processing.
mkdir -p "./out/AKS"
mv ./out/results_kbench-AKS_*m/* "./out/kbench-AKS/"

# Parse
git clone https://github.com/edgelesssys/constellation.git
mkdir -p benchmarks
BDIR=benchmarks
EXT_NAME=AKS
KBENCH_RESULTS=k-bench/out/

python constellation/.github/actions/e2e_kbench/evaluate/parse.py

# Upload result to S3
S3_PATH=s3://edgeless-artifact-store/constellation/benchmarks
aws s3 cp benchmarks/AKS.json ${S3_PATH}/AKS.json

GKE

Create a GKE cluster of desired benchmarking settings (region, instance types). If comparing against Constellation clusters with CVM instances, make sure to select the matching CVM instance type on GCP and enable confidential VMs as well.

Once the cluster is ready, set up managing access via kubectl and take the benchmark:

# Setup
git clone https://github.com/edgelesssys/k-bench.git
cd k-bench && git checkout feat/constellation
./install.sh

# Remove the Constellation encrypted storage class
# Remember to revert this change before running K-Bench on Constellation!
yq 'del(.spec.storageClassName)' config/dp_fio/fio_pvc.yaml
yq 'del(.spec.storageClassName)' config/dp_netperf_internode/netperf_pvc.yml
yq 'del(.spec.storageClassName)' config/dp_network_internode/netperf_pvc.yaml
yq 'del(.spec.storageClassName)' config/dp_network_intranode/netperf_pvc.yml

# Run K-Bench
mkdir -p ./out
kubectl create namespace kbench-pod-namespace --dry-run=client -o yaml | kubectl apply -f -
./run.sh -r "kbench-GKE" -t "default" -o "./out/"
kubectl delete namespace kbench-pod-namespace --wait=true || true
kubectl create namespace kbench-pod-namespace --dry-run=client -o yaml |  kubectl apply -f -
./run.sh -r "kbench-GKE" -t "dp_fio" -o "./out/"
kubectl delete namespace kbench-pod-namespace --wait=true || true
kubectl create namespace kbench-pod-namespace --dry-run=client -o yaml |  kubectl apply -f -
./run.sh -r "kbench-GKE" -t "dp_network_internode" -o "./out/"
kubectl delete namespace kbench-pod-namespace --wait=true || true
kubectl create namespace kbench-pod-namespace --dry-run=client -o yaml |  kubectl apply -f -
./run.sh -r "kbench-GKE" -t "dp_network_intranode" -o "./out/"

# Benchmarks done, do processing.
mkdir -p "./out/GKE"
mv ./out/results_kbench-GKE_*m/* "./out/kbench-GKE/"

# Parse
git clone https://github.com/edgelesssys/constellation.git
mkdir -p benchmarks
BDIR=benchmarks
EXT_NAME=GKE
KBENCH_RESULTS=k-bench/out/

python constellation/.github/actions/e2e_kbench/evaluate/parse.py

# Upload result to S3
S3_PATH=s3://edgeless-artifact-store/constellation/benchmarks
aws s3 cp benchmarks/GKE.json ${S3_PATH}/GKE.json

Constellation

The action updates the stored Constellation records for the selected cloud provider when running on the main branch.