Skip to content

Commit

Permalink
end-to-end precompiled gpu driver testing
Browse files Browse the repository at this point in the history
Signed-off-by: shiva kumar <[email protected]>
  • Loading branch information
shivakunv committed Aug 21, 2024
1 parent f8c3a2b commit 9211cf6
Show file tree
Hide file tree
Showing 11 changed files with 190 additions and 9 deletions.
62 changes: 60 additions & 2 deletions .github/workflows/ci.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,19 @@ on:
- completed
branches:
- main
repository_dispatch:
types: [trigger_ci]
branches:
- precompilede2e
pull_request:
types:
- opened
- synchronize
branches:
- precompilede2e
push:
branches:
- precompilede2e

jobs:
e2e-tests-nvidiadriver:
Expand All @@ -36,6 +49,7 @@ jobs:
AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
AWS_SSH_KEY: ${{ secrets.AWS_SSH_KEY }}
AWS_SESSION_TOKEN: ${{ secrets.AWS_SESSION_TOKEN }}
with:
aws_access_key_id: ${{ secrets.AWS_ACCESS_KEY_ID }}
aws_secret_access_key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
Expand All @@ -52,20 +66,23 @@ jobs:
run: |
echo "instance_hostname=ubuntu@${{ steps.get_public_dns_name.outputs.result }}" >> $GITHUB_ENV
echo "private_key=${{ github.workspace }}/key.pem" >> $GITHUB_ENV
echo "COMMIT_SHORT_SHA=${GITHUB_SHA:0:8}" >> $GITHUB_ENV
# SHIVA
# echo "COMMIT_SHORT_SHA=${GITHUB_SHA:0:8}" >> $GITHUB_ENV
echo "COMMIT_SHORT_SHA=shivaku" >> $GITHUB_ENV
DRIVER_VERSIONS=$(grep '^DRIVER_VERSIONS ?=' versions.mk | awk -F' ?= ' '{print $2}')
echo "DRIVER_VERSIONS=$DRIVER_VERSIONS" >> $GITHUB_ENV
- name: Validate gpu driver
env:
TEST_CASE: "./tests/cases/nvidia-driver.sh"
SSH_RETRY: "0"
run: |
sudo chmod 644 ${{ github.workspace }}/.cache/key
echo "${{ secrets.AWS_SSH_KEY }}" > ${private_key} && chmod 400 ${private_key}
rc=0
for driver_version in ${DRIVER_VERSIONS}; do
echo "Running e2e for DRIVER_VERSION=$driver_version"
./tests/ci-run-e2e.sh ${TEST_CASE} ${COMMIT_SHORT_SHA}-${driver_version} || status=$?
./tests/ci-run-e2e.sh ${TEST_CASE} ${COMMIT_SHORT_SHA}-${driver_version} ${SSH_RETRY} || status=$?
if [ $status -ne 0 ]; then
echo "e2e validation failed for driver version $driver_version with status $status"
rc=$status
Expand All @@ -82,3 +99,44 @@ jobs:
name: nvidiadriver-e2e-test-logs
path: ./logs/
retention-days: 15

- name: Precompiled e2e test- upgrade kernel and Validate gpu driver
env:
TEST_CASE_KERNEL_UPGRADE: "./tests/cases/nvidia-kernel-upgrade.sh"
TEST_CASE: "./tests/cases/nvidia-driver.sh"
run: |
sudo chmod 644 ${{ github.workspace }}/.cache/key
echo "${{ secrets.AWS_SSH_KEY }}" > ${private_key} && chmod 400 ${private_key}
rc=0
for driver_version in ${DRIVER_VERSIONS}; do
echo "Running e2e for DRIVER_VERSION=$driver_version"
export SSH_RETRY="0"
./tests/ci-run-e2e.sh ${TEST_CASE_KERNEL_UPGRADE} ${driver_version} ${SSH_RETRY} || status=$?
if [ $status -ne 0 ]; then
echo "Kernel upgrade failed"
rc=$status
else
# system rebooted enable ssh retry
export SSH_RETRY="1"
DRIVER_BRANCH=$(echo "${TARGET_DRIVER_VERSION}" | cut -d '.' -f 1)
source ./tests/scripts/.definitions.sh
source "${SCRIPT_DIR}"/kernel_version.txt
DRIVER_VERSION="${DRIVER_BRANCH}-${KERNEL_VERSION}-ubuntu22.04"
./tests/ci-run-e2e.sh ${TEST_CASE} ${DRIVER_VERSION} ${SSH_RETRY} || status=$?
if [ $status -ne 0 ]; then
echo "e2e validation failed for driver version $DRIVER_VERSION with status $status"
rc=$status
fi
fi
done
source ./tests/scripts/.definitions.sh
./tests/scripts/pull.sh ${LOG_DIR} logs
exit $rc
- name: Archive test logs
if: ${{ failure() }}
uses: actions/upload-artifact@v4
with:
name: nvidiadriver-Precompiled-e2e-test-logs
path: ./logs/
retention-days: 15
17 changes: 13 additions & 4 deletions .github/workflows/image.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -22,11 +22,9 @@ on:
- synchronize
branches:
- main
- release-*
push:
branches:
- main
- release-*

jobs:
image:
Expand Down Expand Up @@ -62,7 +60,9 @@ jobs:
- name: Calculate build vars
id: vars
run: |
echo "COMMIT_SHORT_SHA=${GITHUB_SHA:0:8}" >> $GITHUB_ENV
# SHIVA
# echo "COMMIT_SHORT_SHA=${GITHUB_SHA:0:8}" >> $GITHUB_ENV
echo "COMMIT_SHORT_SHA=shivaku" >> $GITHUB_ENV
echo "LOWERCASE_REPO_OWNER=$(echo "${GITHUB_REPOSITORY_OWNER}" | awk '{print tolower($0)}')" >> $GITHUB_ENV
REPO_FULL_NAME="${{ github.event.pull_request.head.repo.full_name }}"
echo "${REPO_FULL_NAME}"
Expand Down Expand Up @@ -125,7 +125,9 @@ jobs:
- name: Calculate build vars
id: vars
run: |
echo "COMMIT_SHORT_SHA=${GITHUB_SHA:0:8}" >> $GITHUB_ENV
# SHIVA
# echo "COMMIT_SHORT_SHA=${GITHUB_SHA:0:8}" >> $GITHUB_ENV
echo "COMMIT_SHORT_SHA=shivaku" >> $GITHUB_ENV
echo "LOWERCASE_REPO_OWNER=$(echo "${GITHUB_REPOSITORY_OWNER}" | awk '{print tolower($0)}')" >> $GITHUB_ENV
REPO_FULL_NAME="${{ github.event.pull_request.head.repo.full_name }}"
echo "${REPO_FULL_NAME}"
Expand Down Expand Up @@ -167,12 +169,19 @@ jobs:
docker cp base-${BASE_TARGET}-${{ matrix.flavor }}:/var/kernel_version.txt kernel_version.txt && break
sleep 10
done
echo "KUMAR"
echo "cat kernel_version.txt"
echo "=============="
- name: Build image
env:
IMAGE_NAME: ghcr.io/nvidia/driver
VERSION: ${COMMIT_SHORT_SHA}
PRECOMPILED: "true"
DIST: signed_ubuntu22.04
run: |
echo "SHIVA"
echo "SHIVA============= ${VERSION}"
echo "cat kernel_version.txt"
echo "=============="
source kernel_version.txt && \
make DRIVER_VERSIONS=${DRIVER_VERSIONS} DRIVER_BRANCH=${{ matrix.driver }} build-${DIST}-${DRIVER_VERSION}
8 changes: 8 additions & 0 deletions tests/cases/nvidia-kernel-upgrade.sh
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
#! /bin/bash
# This test case runs the operator installation / test case with the default options.

SCRIPTS_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )"/../scripts && pwd )"
source "${SCRIPTS_DIR}"/.definitions.sh

# Run an end-to-end test cycle
"${SCRIPTS_DIR}"/nvidia-kernel-upgrade-aws.sh
6 changes: 3 additions & 3 deletions tests/ci-run-e2e.sh
Original file line number Diff line number Diff line change
Expand Up @@ -2,14 +2,14 @@

set -xe

if [[ $# -ne 2 ]]; then
echo "TEST_CASE TARGET_DRIVER_VERSION are required"
if [[ $# -ne 3 ]]; then
echo "TEST_CASE TARGET_DRIVER_VERSION SSH_RETRY are required"
exit 1
fi

export TEST_CASE=${1}
export TARGET_DRIVER_VERSION=${2}

export SSH_RETRY=${3}

TEST_DIR="$(pwd)/tests"

Expand Down
4 changes: 4 additions & 0 deletions tests/local.sh
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,10 @@ SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )"/scripts && pwd )"
source ${SCRIPT_DIR}/.definitions.sh
source ${SCRIPT_DIR}/.local.sh

if [ "${SSH_RETRY}" == "1" ]; then
remote_retry
fi

# Sync the project folder to the remote
${SCRIPT_DIR}/push.sh

Expand Down
6 changes: 6 additions & 0 deletions tests/scripts/.definitions.sh
Original file line number Diff line number Diff line change
Expand Up @@ -24,3 +24,9 @@ CASES_DIR="$( cd "${TEST_DIR}/cases" && pwd )"
: ${POD_STATUS_TIME_OUT:="2m"}

: ${LOG_DIR:="/tmp/logs"}

: ${SSH_RETRY:="0"}
: ${SYSTEM_ONLINE_CHECK_TIMEOUT:="900"}

: ${BASE_TARGET:="jammy"}
: ${KERNEL_FLAVOR:="generic"}
4 changes: 4 additions & 0 deletions tests/scripts/.local.sh
Original file line number Diff line number Diff line change
Expand Up @@ -3,3 +3,7 @@
function remote() {
${SCRIPT_DIR}/remote.sh "cd ${PROJECT} && "$@""
}

function remote_retry() {
${SCRIPT_DIR}/remote_retry.sh
}
17 changes: 17 additions & 0 deletions tests/scripts/findkernelversion.sh
Original file line number Diff line number Diff line change
@@ -0,0 +1,17 @@
#!/bin/bash

if [[ "${SKIP_INSTALL}" == "true" ]]; then
echo "Skipping install: SKIP_INSTALL=${SKIP_INSTALL}"
exit 0
fi

SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
source "${SCRIPT_DIR}"/.definitions.sh

export REGCTL_VERSION=v0.4.7
mkdir -p bin
curl -sSLo bin/regctl https://github.com/regclient/regclient/releases/download/${REGCTL_VERSION}/regctl-linux-amd64
chmod a+x bin/regctl
export PATH=$(pwd)/bin:${PATH}
DRIVER_BRANCH=$(echo "${TARGET_DRIVER_VERSION}" | cut -d '.' -f 1)
regctl image get-file ghcr.io/nvidia/driver:base-${BASE_TARGET}-${KERNEL_FLAVOR}-${DRIVER_BRANCH} /var/kernel_version.txt ${SCRIPT_DIR}/kernel_version.txt || true
5 changes: 5 additions & 0 deletions tests/scripts/install-operator.sh
Original file line number Diff line number Diff line change
Expand Up @@ -5,9 +5,14 @@ if [[ "${SKIP_INSTALL}" == "true" ]]; then
exit 0
fi

echo "Checking current kernel version..."
CURRENT_KERNEL=$(uname -r)
echo "Current kernel version: $CURRENT_KERNEL"

SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
source ${SCRIPT_DIR}/.definitions.sh


OPERATOR_OPTIONS="${OPERATOR_OPTIONS} --set driver.repository=${PRIVATE_REGISTRY}/nvidia --set driver.version=${TARGET_DRIVER_VERSION}"

# add helm driver repo
Expand Down
43 changes: 43 additions & 0 deletions tests/scripts/nvidia-kernel-upgrade-aws.sh
Original file line number Diff line number Diff line change
@@ -0,0 +1,43 @@
#!/bin/bash

if [[ "${SKIP_INSTALL}" == "true" ]]; then
echo "Skipping install: SKIP_INSTALL=${SKIP_INSTALL}"
exit 0
fi

SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
source "${SCRIPT_DIR}"/.definitions.sh

# finding kernel version
${SCRIPT_DIR}/findkernelversion.sh
source "${SCRIPT_DIR}"/kernel_version.txt

echo "Checking current kernel version..."
CURRENT_KERNEL=$(uname -r)
echo "Current kernel version: $CURRENT_KERNEL"

echo ""
echo ""
echo "--------------Starting the Precompiled kernel version ${KERNEL_VERSION} upgrade--------------"

sudo apt-get update -y
sudo apt-get install linux-image-${KERNEL_VERSION} -y
if [ $? -ne 0 ]; then
echo "Kernel upgrade failed."
exit 1
fi

echo "Checking the upgraded kernel version ${KERNEL_VERSION}..."
CURRENT_KERNEL=$(uname -r)
echo "Upgraded kernel version: $CURRENT_KERNEL"

echo "update grub ..."
sudo update-grub
echo "Rebooting ..."
# Run the reboot command with nohup to avoid abrupt SSH closure issues
nohup sudo reboot &

echo "--------------Installation of kernel completed --------------"

# Exit with a success code since the reboot command was issued successfully
exit 0
27 changes: 27 additions & 0 deletions tests/scripts/remote_retry.sh
Original file line number Diff line number Diff line change
@@ -0,0 +1,27 @@
#!/bin/bash

SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
source ${SCRIPT_DIR}/.definitions.sh
source ${SCRIPT_DIR}/.local.sh

try_ssh_connection() {
ssh -o ConnectTimeout=10 -i ${private_key} ${instance_hostname} "exit"
return $?
}

echo "Waiting for aws system to come back online..."
START_TIME=$(date +%s)
while true; do
sleep 60 # sleep before as system restarted earlier
try_ssh_connection
if [ $? -eq 0 ]; then
echo "Successfully connected to aws system after reboot."
break;
fi
ELAPSED_TIME=$(($(date +%s) - START_TIME))
if [ "$ELAPSED_TIME" -ge "$SYSTEM_ONLINE_CHECK_TIMEOUT" ]; then
echo "Failed to connect to aws within ${SYSTEM_ONLINE_CHECK_TIMEOUT} minutes after reboot."
exit 1
fi
echo "ssh retry again..."
done

0 comments on commit 9211cf6

Please sign in to comment.