We have our own GitLab-CI runner, but with only one we can't run many
jobs in parallel so it takes a very long time before all the tests have
completed. In addition, if that runner goes down, we have nothing at
all.
GitLab offers the following machine types for hosted runners on Linux
x86-64 [1]. The default is the "small" runner. Using a larger runner
increases the "Cost factor" [2].
For opensource projects, the Cost factor is reduced to 0.5 (1 minute per
2 minutes of job time) whatever the runner type.
Runner Tag vCPUs Memory Storage Cost factor (OSS)
saas-linux-small-amd64 (default) 2 8 GB 30 GB 1 (0.5)
saas-linux-medium-amd64 4 16 GB 50 GB 2 (0.5)
saas-linux-large-amd64 (Premium and Ultimate only) 8 32 GB 100 GB 3 (0.5)
saas-linux-xlarge-amd64 (Premium and Ultimate only) 16 64 GB 200 GB 6 (0.5)
saas-linux-2xlarge-amd64 (Premium and Ultimate only) 32 128 GB 200 GB 12 (0.5)
Compute minutes consumed by a job is calculated by:
Job duration / 60 * Cost factor
(Job duration: The time, in seconds, that a job took to run, not
including time spent in the created or pending statuses.)
Thanks to the GitLab OSS program [3], Buildroot benefits from a free
Ultimate subscription and can use GitLab shared runners tagged with
saas-linux-{large, xlarge, 2xlarge}-amd64. In addition, we receive
50,000 free runner minutes per month.
In order to use one of those tags in Buildroot GitLab-CI jobs, we have
to classify all tests by resource requirement, to make sure the job
doesn't fail because it times out or has insufficient memory or disk
space. While we usually shouldn't use the largest runner for
everything, we can use larger runner without cost penalty thanks to
the cost factor reduced to 0.5 for opensource projects. This will
reduce the CI minutes consumed by a CPU intensive job.
First we introduce some new templates used to add the corresponding
runner tag to a runtime test job (reusing the GitLab terminology).
.runner-{small,medium,large,xlarge,2xlarge}
Most of our tests are fast (checkpackage, test_external_bootlin...), so
saas-linux-small-amd64 runner tag is enough. Default to this tag if
nothing else is specified.
Add a comment next to the test class to provide the runner tag.
This runner tag is retrieved when generating the
generated-gitlab-ci.yml file used to create the child pipeline where
the runtime test jobs are executed.
We use the list of runtime tests returned by node2:
"tests.boot.test_edk2.TestEdk2.test_run"
We convert each element of this list to get the path to the test source
file and the name of the test:
"support/testing/tests/boot/test_edk2.py"
TestEdk2
With that, we can grep into the test source file to retrieve the runner
tag placed one line above the test class:
# GitLab-runner: large
class TestEdk2(infra.basetest.BRTest):
Once the runner tag is retrieved, it's used to use the corresponding
runner template to the runtime test job:
tests.boot.test_edk2.TestEdk2.test_run: { extends: [ .runtime_test_base, .runner-large ]}
GitLab runners hosted by the Buildroot project should be able to run
any jobs, so they should be tagged with Gitlab runner tags
(saas-linux-{small,medium,large,xlarge,2xlarge}-amd64).
A specific runner tag "buildroot-runner" can be used to allow running
a job only on such runners.
If a test can't be executed by any shared GitLab-CI runners, we have
to use a runner owned by the Buildroot project. In this case we have
to use a specific template ".runner-buildroot-runner-only" in order to
add the specific runner tag "buildroot-runner" to the job running the
test. There is no such runtime test at the moment.
The proposed classification is based on a previous pipeline analysis
[5]:
- Tests lasting more than 3 hours will use 2xlarge runners.
- Tests lasting more than 2 hours will use xlarge runners.
- Tests lasting more than 1 hours will use large runners.
- Tests building a kernel or a toolchain will use medium runners.
- All other tests will use small runners when possible.
CI minute cost estimate:
tests.package.test_clang.TestClangCompilerRT.test_run lasts 4h25 on the
Buildroot runner. If we this duration for 2xlarge runners, the CI
minute consumed would be:
(15900 / 60) * 0.5 = 133
With 6 jobs using a 2xlarge runners we used ~795 CI minutes.
tests.package.test_kmscube.TestKmsCube.test_run list 2h04 on the
Buildroot runner. If we this duration for xlarge runners, the CI
minute consumed would be:
(7440 / 60) * 0.5 = 62
With 4 jobs using a xlarge runners we used ~248 CI minutes.
tests.package.test_weston.TestWeston.test_run last 1h15 on th
Buildroot runner. If we this duration for large runners, the CI
minute consumed would be:
(4500 / 60) * 0.5 = 37.5
With 28 jobs using a large runners we used ~1050 CI minutes.
tests.package.test_gstreamer1.TestGstreamer1.test_run last 46min on the
Buildroot runner. If we this duration for large runners, the CI
minute consumed would be:
(2760 / 60) * 0.5 = 23
With 31 jobs using a medium runners we used ~713 CI minutes.
tests.package.test_python.TestPython3Py.test_run last 13min on the
Buildroot runner. If we this duration for large runners, the CI
minute consumed would be:
(780 / 60) * 0.5 = 6.5
With 691 jobs using a medium runners we used ~4491 CI minutes.
In total, one pipeline for the runtime tests cost ~7297 CI minutes.
After a first try [6], we are actually using 8000 CI minutes per
pipeline.
We run such pipeline once a week (on Monday), one for each Buildroot
releases every 3 month, one for each stable and LTS release per month,
and one for each release candidate (3).
Worst case (release month):
(4 weeks + 1 release + 1 stable + 1 LTS + 3 release candidate) * 8000 CI
minutes: 80000 CI minutes / 50000.
So we would spend the minutes very quickly in the worst case scenario.
We have to keep one pipeline under 5000 CI minutes.
[1] https://docs.gitlab.com/ci/runners/hosted_runners/linux/#machine-types-available-for-linux---x86-64
[2] https://docs.gitlab.com/ci/pipelines/compute_minutes/#cost-factorshttps://docs.gitlab.com/ci/pipelines/compute_minutes/#compute-usage-calculationhttps://docs.gitlab.com/ci/pipelines/compute_minutes/#cost-factors-of-hosted-runners-for-gitlabcom
[3] https://gitlab.com/buildroot.org/gitlab-oss
[4] https://docs.gitlab.com/ci/runners/hosted_runners/#gitlabcom-hosted-runner-workflow
[5] https://gitlab.com/buildroot.org/buildroot/-/pipelines/2416603721
[6] https://gitlab.com/buildroot.org/buildroot/-/pipelines/2562421098
Signed-off-by: Romain Naour <romain.naour@smile.fr>
[Arnout:
- simplify parsing of test_file and test_name;
- match the entire test_name instead of substring;
- assume "small" by default;
- remove the "small" tags;
- use "gitlab-runner" instead of "Gitlab-runner".
]
Signed-off-by: Arnout Vandecappelle <arnout@rnout.be>
Signed-off-by: Arnout Vandecappelle <arnout@rnout.be>
We have our own GitLab-CI runner, but with only one we can't run many
jobs in parallel so it takes a very long time before all the tests have
completed. In addition, if that runner goes down, we have nothing at
all.
GitLab offers the following machine types for hosted runners on Linux
x86-64 [1]. The default is the "small" runner. Using a larger runner
increases the "Cost factor" [2]:
Runner Tag vCPUs Memory Storage Cost factor
saas-linux-small-amd64 (default) 2 8 GB 30 GB 1
saas-linux-medium-amd64 4 16 GB 50 GB 2
saas-linux-large-amd64 (Premium and Ultimate only) 8 32 GB 100 GB 3
saas-linux-xlarge-amd64 (Premium and Ultimate only) 16 64 GB 200 GB 6
saas-linux-2xlarge-amd64 (Premium and Ultimate only) 32 128 GB 200 GB 12
Compute minutes consumed by a job is calculated by:
Job duration / 60 * Cost factor
(Job duration: The time, in seconds, that a job took to run, not
including time spent in the created or pending statuses.)
Thanks to the GitLab OSS program [3], Buildroot benefits from a free
Ultimate subscription and can use GitLab shared runners tagged with
saas-linux-{large, xlarge, 2xlarge}-amd64. In addition, we receive
50,000 free runner minutes per month.
For opensource projects, the Cost factor is reduced by 0,5 (1 minute per
2 minutes of job time).
In order to use one of those tags in Buildroot GitLab-CI jobs, we have
to classify all tests by resource requirement, to make sure the job
doesn't fail because it times out or has insufficient memory or disk
space. We also don't want to use the largest runner for everything
because that would spend the minutes very quickly while the resources
aren't used efficiently (mostly serialized).
First we introduce some new templates used to add the corresponding
runner tag to a runtime test job (reusing the GitLab terminology).
.runner-{small,medium,large,xlarge,2xlarge}
Most of our tests are fast (checkpackage, test_external_bootlin...), so
saas-linux-small-amd64 runner tag is enough. Default to this tag if
nothing else is specified.
Add a comment next to the test class to provide the runner tag.
This runner tag is retrieved when generating the
generated-gitlab-ci.yml file used to create the child pipeline where
the runtime test jobs are executed.
We use the list of runtime tests returned by node2:
"tests.boot.test_edk2.TestEdk2.test_run"
We convert each element of this list to get the path to the test source
file and the name of the test:
"support/testing/tests/boot/test_edk2.py"
TestEdk2
With that, we can grep into the test source file to retrieve the runner
tag placed one line above the test class:
# GitLab-runner: large
class TestEdk2(infra.basetest.BRTest):
Once the runner tag is retrieved, it's used to use the corresponding
runner template to the runtime test job:
tests.boot.test_edk2.TestEdk2.test_run: { extends: [ .runtime_test_base, .runner-large ]}
GitLab runners hosted by the Buildroot project should be able to run any
jobs, so a specific runner tag "buildroot-runner" is used to allow
running a job on such runners. The "buildroot-runner" tag is used in
addition to GitLab ones to run a job on the Buildroot runner or on a
shared GitLab-CI runner.
If a test can't be executed by any shared GitLab-CI runners, we have
to use a runner owned by the Buildroot project. In this case we have
to use a specific template ".runner-buildroot-runner-only". There is no
such runtime test at the moment.
The proposed classification is based on a previous pipeline analysis
[5]:
- Tests lasting more than 3 hours will use 2xlarge runners.
- Tests lasting more than 2 hours will use xlarge runners.
- Tests lasting more than 1 hours will use large runners.
- Tests building a kernel or a toolchain will use medium runners.
- All other tests will use small runners when possible.
CI minute cost estimate:
tests.package.test_clang.TestClangCompilerRT.test_run lasts 4h25 on the
Buildroot runner. If we this duration for 2xlarge runners, the CI
minute consumed would be:
(265 / 60) * 12 * 0.5 = 26,5
With 6 jobs using a 2xlarge runners we used ~160 CI minutes.
tests.package.test_kmscube.TestKmsCube.test_run list 2h04 on the
Buildroot runner. If we this duration for xlarge runners, the CI
minute consumed would be:
(124 / 60) * 6 * 0.5 = 6,2
With 4 jobs using a xlarge runners we used ~37 CI minutes.
tests.package.test_weston.TestWeston.test_run last 1h15 on th
Buildroot runner. If we this duration for large runners, the CI
minute consumed would be:
(75 / 60) * 3 * 0.5 = 1.87
With 28 jobs using a large runners we used ~53 CI minutes.
tests.package.test_gstreamer1.TestGstreamer1.test_run last 46min on the
Buildroot runner. If we this duration for large runners, the CI
minute consumed would be:
(46 / 60) * 2 * 0.5 = 0.76
With 31 jobs using a medium runners we used ~24 CI minutes.
tests.package.test_python.TestPython3Py.test_run last 13min on the
Buildroot runner. If we this duration for large runners, the CI
minute consumed would be:
(13 / 60) * 1 * 0.5 = 0.1
With 691 jobs using a medium runners we used ~69 CI minutes.
In total, one pipeline for the runtime tests cost ~350 CI minutes.
We run such pipeline once a week (on Monday), one for each Buildroot
releases every 3 month, one for each stable and LTS release per month,
and one for each release candidate (3).
Worst case (realse month):
(4 weeks + 1 release + 1 stable + 1 LTS + 3 release candidate) * 350 CI
minutes: 3500 CI minutes / 50000.
So we should be able to run our pipeline without exhausting our CI
minutes credit.
[1] https://docs.gitlab.com/ci/runners/hosted_runners/linux/#machine-types-available-for-linux---x86-64
[2] https://docs.gitlab.com/ci/pipelines/compute_minutes/#cost-factorshttps://docs.gitlab.com/ci/pipelines/compute_minutes/#compute-usage-calculationhttps://docs.gitlab.com/ci/pipelines/compute_minutes/#cost-factors-of-hosted-runners-for-gitlabcom
[3] https://gitlab.com/buildroot.org/gitlab-oss
[4] https://docs.gitlab.com/ci/runners/hosted_runners/#gitlabcom-hosted-runner-workflow
[5] https://gitlab.com/buildroot.org/buildroot/-/pipelines/2416603721
Signed-off-by: Romain Naour <romain.naour@smile.fr>
[Arnout:
- simplify parsing of test_file and test_name;
- match the entire test_name instead of substring;
- assume "small" by default;
- remove the "small" tags;
- use "gitlab-runner" instead of "Gitlab-runner".
]
Signed-off-by: Arnout Vandecappelle <arnout@rnout.be>
This commit bump the kernel version of this docker test (5.4->6.18), and thus
fixes several issues preventing the Docker Compose runtime test from
succeeding due to this bump:
Kernel Infrastructure:
The Docker daemon failed to start because the 'nat' table could not
be initialized. This was due to missing legacy Netfilter support.
- Enable CONFIG_NETFILTER_XTABLES_LEGACY to support iptables nat table.
- Enable CONFIG_NAMESPACES and CONFIG_USER_NS for container isolation.
- Enable CONFIG_TUN and CONFIG_VETH for container networking.
- Enable CONFIG_IKCONFIG and CONFIG_IKCONFIG_PROC for easier debugging.
Build Fixes:
While building the kernel tools, the libelf header was missing because
objtool was not handled correctly for this kernel version. This resulted
in a fatal error: "gelf.h: No such file or directory".
- Select BR2_LINUX_KERNEL_NEEDS_HOST_LIBELF=y to build host-libelf.
Resources and Stability:
The previous 512MB of RAM was tight for the stack (Kernel +
Docker + Containerd + Python), potentially leading to Out-Of-Memory kills.
Idem for the disk size for Docker images/layers storage.
- Increase QEMU RAM to 1024MB (-m 1024M).
- Increase rootfs size to 1024MB.
- Switch to EXT4 (BR2_TARGET_ROOTFS_EXT2_4=y) for better stability
and modern feature support required by Docker's storage drivers.
Test Script Adjustments:
- Update rootfs path in the test script to point to rootfs.ext4.
Signed-off-by: Dowan Gullient <dowan.gullient@smile.fr>
Signed-off-by: Julien Olivain <ju.o@free.fr>
Since commit 947dbc92a2 the Bootlin
stable toolchain uses headers 5.4, which can lead to build outputs
incompatible with older kernels. Update to the latest 5.4 kernel to
avoid this. The kernel config is updated according to the requirements
of current Docker, runc, and crun.
Additionally switch the CPU of the test VM to Haswell, to avoid
dockerd failing to start with a "This program can only be run on AMD64
processors with v3 microarchitecture support" error.
Signed-off-by: Fiona Klute (WIWA) <fiona.klute@gmx.de>
Signed-off-by: Romain Naour <romain.naour@smile.fr>
Via the docker-compose runtime test, we can exercise the full suite of
the docker-related packages: docker-compose, of course, but also
docker-engine, which in turn allows exercising containerd. The latter
by defualt uses runc as the container runtime, but can alternatively use
crun.
Extend the docker-compose runtime test with a variant that enables crun.
Signed-off-by: Yann E. MORIN <yann.morin.1998@free.fr>
Signed-off-by: Julien Olivain <ju.o@free.fr>
Python-docker needs a working docker setup to do anything useful, so add it
to the existing docker_compose (which tests docker and docker-compose)
rather than adding a completely new test.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Reviewed-by: Christian Stewart <christian@aperture.us>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Docker compose up outputs terminal control characters intended for
interactive output viewing.
Wget similarly can use the -q option to produce quieter logs.
Signed-off-by: Christian Stewart <christian@paral.in>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
To avoid spending some time to build the x86_64 toolchain (~20min),
switch to corei7 cpu (Nahalem) and use the prebuilt Bootlin toolchain.
We have to use the "stable" Bootlin toolchain to use the same kernel version
for the toolchain kernel headers and the running kernel.
With the "bleeding-edge" toolchain we have the "kernel too old" issue
(running kernel 4.19 vs kernel headers 5.4)
Runtime tested locally.
Signed-off-by: Romain Naour <romain.naour@gmail.com>
Cc: Peter Korsgaard <peter@korsgaard.com>
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
gcc 10.x is now used by default but the kernel 4.19 used by
test_docker_compose doesn't build with it.
Bump the kernel to 4.19.204 release that contains a lot of
fixes for newer gcc.
Signed-off-by: Romain Naour <romain.naour@gmail.com>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
The BRTest() class implements an assertRunOk() method that does the
very common work of running a command inside the emulator, and
checking that it is successful.
This commit changes all locations where this .assertRunOk() method can
be used, instead of open-coding the same logic.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Signed-off-by: Yann E. MORIN <yann.morin.1998@free.fr>
Extend docker_compose_test() to expose /bin on the host to the container
through a volume mount and verify that /bin/busybox can be downloaded and
contains the right data.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Extend docker_test() to expose a random (8888) port to verify that doesn't
fail, and extend the docker-compose test to run the busybox httpd in the
background, expose that as port 80 and verify that /etc/resolv.conf could be
fetched by wget.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Since commit 4f8229653 (package/docker-engine: needs more runtime
dependencies), docker-engine now automatically pulls in cgroupfs-mount, so
drop the explicit handling of it in TestDockerCompose.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Fixes https://gitlab.com/buildroot.org/buildroot/-/jobs/332656041
The recent bump of a number of python packages broke docker-compose, as
docker-compose specifies both minimum and maximum versions for (most of) its
dependencies:
Dependencies of docker-compse 1.20.1 (! = unmet):
cached-property: < 2 (currently 1.51)
docopt: < 0.7 (currently 0.6.2)
! pyyaml: < 4.0, patched to < 4.3 (currently 5.1.2)
requests: < 2.19, patched to < 3 (currently 2.22.0)
! texttable: < 0.10 (currently 1.6.2)
websocket-client: < 1.0 (currently 0.56.0)
! docker: < 4.0 (currently 4.1.0)
dockerpty: < 0.5 (currently 0.4.1)
six: < 2 (currently 1.12.0)
jsonschema: < 3 (currently 2.5.1)
enum34: < 2 (currently 1.1.6)
backports.ssl-match-hostname: >= 3.5 (currently 3.7.0.1)
ipaddress: >= 1.0.16 (currently 1.0.23)
To fix this, bump docker-compose to the most recent release (1.24.1). This
is unfortunately not enough, as our docker, pyyaml, requests and texttable
packages are too new, so add 3 patches from upstream to relax the version
checks of dependencies. Notice that patch 0003 is from
https://github.com/docker/compose/pull/6623 and has not been merged yet.
Discussions around the problem of these maximum versions of the dependencies
and the fact that all downstream users have to patch it is ongoing here:
https://github.com/docker/compose/issues/6756
docker-compose 1.24.1 added a requirement for ssh support in python-docker in:
7b82b2e8c7
So add a dependency for python-paramiko and update the toolchain dependency
for C++ (from python-paramiko -> python-cryptography) and adjust the
toolchain configuration of the runtime test to match.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
runc (which is a reverse dependency of docker-engine) is about to gain a
!uclibc dependency, so move to a glibc toolchain instead.
There are currently no prebuilt x86_64 / core2 / glibc toolchains available,
so instead use the internal toolchain backend to build one.
While we are at it, drop the infra.basetest.BASIC_TOOLCHAIN_CONFIG
reference, as that ARM toolchain configuration doesn't make any sense for
this x86-64 based test.
add docker / docker-compose tests
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Build for x86-64 as public containers in general are only available for
x86-64. Docker needs a number of kernel options enabled, so use a custom
kernel config based on the qemu one.
Docker needs entropy at startup, so enable the virtio-rng-pci device to
expose entropy to the guest. The default RAM amount (128M) is not enough to
run docker / docker-compose, so bump to 512MB.
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
Reviewed-by: Matthew Weber <matthew.weber@rockwellcollins.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
