numam-spdk/docker/README.md

100 lines
3.7 KiB
Markdown
Raw Normal View History

docker: Add docker-compose for building basic SPDK containers This suite can be used to deploy containers with the following functionality (more details in README.md): - storage-target - proxy-container - traffic-generator This will run simple fio test as per fio.conf against nvmf controller provided by initiator-container. Similar task can be performed directly from initiator-container as well. Each container includes SPDK installation with most common tools, e.g. rpc.py, available under $PATH. This allows for something like: docker-compose exec storage-target rpc.py nvmf_get_subsystems Note that SPDK environment heavily depends on a running kernel hence all the containers need to be privileged. That said, to make sure containers are not affecting the host too much, some tasks must be done prior running them. This includes: - loading proper kernel modules (like nvme-fabrics, etc.) - allocating hugepages and having at least one hugetlbfs mount available under /dev/hugepages base_build is created as docker multi-stage build. This is done in order to decrease the size of the final image. The SPDK RPMs are built inside a base image and then copied over to the main image (+ fio binary) - this leaves all the dependencies inside the intermediate image instead of the final one. The resulted difference in size may look similar to the following (it may differ depending on the docker version etc.): no multi-stage build: spdk_base == 1.04GB multi-stage build: spdk_base == 261MB Signed-off-by: Michal Berger <michalx.berger@intel.com> Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com> Change-Id: I825bd0d0bb4071bd9d44b6a0749c033894899ae0 Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9055 Reviewed-by: Jim Harris <james.r.harris@intel.com> Reviewed-by: Monica Kenguva <monica.kenguva@intel.com> Reviewed-by: Xiaodong Liu <xiaodong.liu@intel.com> Reviewed-by: Changpeng Liu <changpeng.liu@intel.com> Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
2021-08-03 07:16:49 +00:00
# SPDK Docker suite
This suite is meant to serve as an example of how SPDK can be encapsulated
into docker container images. The example containers consist of SPDK NVMe-oF
target sharing devices to another SPDK NVMe-oF application. Which serves
as both initiator and target. Finally a traffic generator based on FIO
issues I/O to the connected devices.
## Prerequisites
docker: We recommend version 20.10 and above because it supports cgroups v2 for
customization of host resources like CPUs, memory, and block I/O.
docker-compose: We recommend using 1.29.2 version or newer.
kernel: Hugepages must be allocated prior running the containers and hugetlbfs
mount must be available under /dev/hugepages. Also, tmpfs should be mounted
under /dev/shm. Depending on the use-case, some kernel modules should be also
loaded into the kernel prior running the containers.
proxy: If you are working behind firewall make sure dockerd is aware of the
proxy. Please refer to:
[docker-proxy](https://docs.docker.com/config/daemon/systemd/#httphttps-proxy)
To pass `$http_proxy` to docker-compose build use:
~~~{.sh}
docker-compose build --build-arg PROXY=$http_proxy
~~~
## How-To
`docker-compose.yaml` shows an example deployment of the storage containers based on SPDK.
Running `docker-compose build` creates 5 docker images:
docker: Add docker-compose for building basic SPDK containers This suite can be used to deploy containers with the following functionality (more details in README.md): - storage-target - proxy-container - traffic-generator This will run simple fio test as per fio.conf against nvmf controller provided by initiator-container. Similar task can be performed directly from initiator-container as well. Each container includes SPDK installation with most common tools, e.g. rpc.py, available under $PATH. This allows for something like: docker-compose exec storage-target rpc.py nvmf_get_subsystems Note that SPDK environment heavily depends on a running kernel hence all the containers need to be privileged. That said, to make sure containers are not affecting the host too much, some tasks must be done prior running them. This includes: - loading proper kernel modules (like nvme-fabrics, etc.) - allocating hugepages and having at least one hugetlbfs mount available under /dev/hugepages base_build is created as docker multi-stage build. This is done in order to decrease the size of the final image. The SPDK RPMs are built inside a base image and then copied over to the main image (+ fio binary) - this leaves all the dependencies inside the intermediate image instead of the final one. The resulted difference in size may look similar to the following (it may differ depending on the docker version etc.): no multi-stage build: spdk_base == 1.04GB multi-stage build: spdk_base == 261MB Signed-off-by: Michal Berger <michalx.berger@intel.com> Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com> Change-Id: I825bd0d0bb4071bd9d44b6a0749c033894899ae0 Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9055 Reviewed-by: Jim Harris <james.r.harris@intel.com> Reviewed-by: Monica Kenguva <monica.kenguva@intel.com> Reviewed-by: Xiaodong Liu <xiaodong.liu@intel.com> Reviewed-by: Changpeng Liu <changpeng.liu@intel.com> Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
2021-08-03 07:16:49 +00:00
- build_base
- storage-target
- proxy-container
- traffic-generator-nvme
- traffic-generator-virtio
docker: Add docker-compose for building basic SPDK containers This suite can be used to deploy containers with the following functionality (more details in README.md): - storage-target - proxy-container - traffic-generator This will run simple fio test as per fio.conf against nvmf controller provided by initiator-container. Similar task can be performed directly from initiator-container as well. Each container includes SPDK installation with most common tools, e.g. rpc.py, available under $PATH. This allows for something like: docker-compose exec storage-target rpc.py nvmf_get_subsystems Note that SPDK environment heavily depends on a running kernel hence all the containers need to be privileged. That said, to make sure containers are not affecting the host too much, some tasks must be done prior running them. This includes: - loading proper kernel modules (like nvme-fabrics, etc.) - allocating hugepages and having at least one hugetlbfs mount available under /dev/hugepages base_build is created as docker multi-stage build. This is done in order to decrease the size of the final image. The SPDK RPMs are built inside a base image and then copied over to the main image (+ fio binary) - this leaves all the dependencies inside the intermediate image instead of the final one. The resulted difference in size may look similar to the following (it may differ depending on the docker version etc.): no multi-stage build: spdk_base == 1.04GB multi-stage build: spdk_base == 261MB Signed-off-by: Michal Berger <michalx.berger@intel.com> Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com> Change-Id: I825bd0d0bb4071bd9d44b6a0749c033894899ae0 Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9055 Reviewed-by: Jim Harris <james.r.harris@intel.com> Reviewed-by: Monica Kenguva <monica.kenguva@intel.com> Reviewed-by: Xiaodong Liu <xiaodong.liu@intel.com> Reviewed-by: Changpeng Liu <changpeng.liu@intel.com> Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
2021-08-03 07:16:49 +00:00
The `build_base` image provides the core components required to containerize SPDK
applications. The fedora:33 image from the Fedora Container Registry is used and then SPDK is installed. SPDK is installed out of `build_base/spdk.tar.gz` provided.
See `build_base` folder for details on what's included in the final image.
Running `docker-compose up` creates 3 docker containers:
-- storage-target: Contains SPDK NVMe-oF target exposing single subsystem to
`proxy-container` based on malloc bdev.
-- proxy-container: Contains SPDK NVMe-oF target connecting to `storage-target`
and then exposing the same devices to `traffic-generator-nvme` using NVMe-oF and
to `traffic-generator-virtio` using Virtio.
-- traffic-generator-nvme: Contains FIO using SPDK plugin to connect to `proxy-container`
and runs a sample workload.
-- traffic-generator-virtio: Contains FIO using SPDK plugin to connect to `proxy-container`
docker: Add docker-compose for building basic SPDK containers This suite can be used to deploy containers with the following functionality (more details in README.md): - storage-target - proxy-container - traffic-generator This will run simple fio test as per fio.conf against nvmf controller provided by initiator-container. Similar task can be performed directly from initiator-container as well. Each container includes SPDK installation with most common tools, e.g. rpc.py, available under $PATH. This allows for something like: docker-compose exec storage-target rpc.py nvmf_get_subsystems Note that SPDK environment heavily depends on a running kernel hence all the containers need to be privileged. That said, to make sure containers are not affecting the host too much, some tasks must be done prior running them. This includes: - loading proper kernel modules (like nvme-fabrics, etc.) - allocating hugepages and having at least one hugetlbfs mount available under /dev/hugepages base_build is created as docker multi-stage build. This is done in order to decrease the size of the final image. The SPDK RPMs are built inside a base image and then copied over to the main image (+ fio binary) - this leaves all the dependencies inside the intermediate image instead of the final one. The resulted difference in size may look similar to the following (it may differ depending on the docker version etc.): no multi-stage build: spdk_base == 1.04GB multi-stage build: spdk_base == 261MB Signed-off-by: Michal Berger <michalx.berger@intel.com> Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com> Change-Id: I825bd0d0bb4071bd9d44b6a0749c033894899ae0 Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9055 Reviewed-by: Jim Harris <james.r.harris@intel.com> Reviewed-by: Monica Kenguva <monica.kenguva@intel.com> Reviewed-by: Xiaodong Liu <xiaodong.liu@intel.com> Reviewed-by: Changpeng Liu <changpeng.liu@intel.com> Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
2021-08-03 07:16:49 +00:00
and runs a sample workload.
Each container is connected to a separate "spdk" network which is created before
deploying the containers. See `docker-compose.yaml` for the network's detailed setup and ip assignment.
All the above boils down to:
~~~{.sh}
cd docker
tar -czf build_base/spdk.tar.gz --exclude='docker/*' -C .. .
docker-compose build
docker-compose up
~~~
The `storage-target` and `proxy-container` can be started as services.
Allowing for multiple traffic generator containers to connect.
docker: Add docker-compose for building basic SPDK containers This suite can be used to deploy containers with the following functionality (more details in README.md): - storage-target - proxy-container - traffic-generator This will run simple fio test as per fio.conf against nvmf controller provided by initiator-container. Similar task can be performed directly from initiator-container as well. Each container includes SPDK installation with most common tools, e.g. rpc.py, available under $PATH. This allows for something like: docker-compose exec storage-target rpc.py nvmf_get_subsystems Note that SPDK environment heavily depends on a running kernel hence all the containers need to be privileged. That said, to make sure containers are not affecting the host too much, some tasks must be done prior running them. This includes: - loading proper kernel modules (like nvme-fabrics, etc.) - allocating hugepages and having at least one hugetlbfs mount available under /dev/hugepages base_build is created as docker multi-stage build. This is done in order to decrease the size of the final image. The SPDK RPMs are built inside a base image and then copied over to the main image (+ fio binary) - this leaves all the dependencies inside the intermediate image instead of the final one. The resulted difference in size may look similar to the following (it may differ depending on the docker version etc.): no multi-stage build: spdk_base == 1.04GB multi-stage build: spdk_base == 261MB Signed-off-by: Michal Berger <michalx.berger@intel.com> Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com> Change-Id: I825bd0d0bb4071bd9d44b6a0749c033894899ae0 Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9055 Reviewed-by: Jim Harris <james.r.harris@intel.com> Reviewed-by: Monica Kenguva <monica.kenguva@intel.com> Reviewed-by: Xiaodong Liu <xiaodong.liu@intel.com> Reviewed-by: Changpeng Liu <changpeng.liu@intel.com> Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
2021-08-03 07:16:49 +00:00
~~~{.sh}
docker-compose up -d proxy-container
docker-compose run traffic-generator-nvme
docker-compose run traffic-generator-virtio
docker: Add docker-compose for building basic SPDK containers This suite can be used to deploy containers with the following functionality (more details in README.md): - storage-target - proxy-container - traffic-generator This will run simple fio test as per fio.conf against nvmf controller provided by initiator-container. Similar task can be performed directly from initiator-container as well. Each container includes SPDK installation with most common tools, e.g. rpc.py, available under $PATH. This allows for something like: docker-compose exec storage-target rpc.py nvmf_get_subsystems Note that SPDK environment heavily depends on a running kernel hence all the containers need to be privileged. That said, to make sure containers are not affecting the host too much, some tasks must be done prior running them. This includes: - loading proper kernel modules (like nvme-fabrics, etc.) - allocating hugepages and having at least one hugetlbfs mount available under /dev/hugepages base_build is created as docker multi-stage build. This is done in order to decrease the size of the final image. The SPDK RPMs are built inside a base image and then copied over to the main image (+ fio binary) - this leaves all the dependencies inside the intermediate image instead of the final one. The resulted difference in size may look similar to the following (it may differ depending on the docker version etc.): no multi-stage build: spdk_base == 1.04GB multi-stage build: spdk_base == 261MB Signed-off-by: Michal Berger <michalx.berger@intel.com> Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com> Change-Id: I825bd0d0bb4071bd9d44b6a0749c033894899ae0 Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9055 Reviewed-by: Jim Harris <james.r.harris@intel.com> Reviewed-by: Monica Kenguva <monica.kenguva@intel.com> Reviewed-by: Xiaodong Liu <xiaodong.liu@intel.com> Reviewed-by: Changpeng Liu <changpeng.liu@intel.com> Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
2021-08-03 07:16:49 +00:00
~~~
Enviroment variables to containers can be passed as shown in
[docs](https://docs.docker.com/compose/environment-variables/).
For example extra arguments to fio can be passed as so:
~~~{.sh}
docker-compose run -e FIO_ARGS="--minimal" traffic-generator-nvme
docker: Add docker-compose for building basic SPDK containers This suite can be used to deploy containers with the following functionality (more details in README.md): - storage-target - proxy-container - traffic-generator This will run simple fio test as per fio.conf against nvmf controller provided by initiator-container. Similar task can be performed directly from initiator-container as well. Each container includes SPDK installation with most common tools, e.g. rpc.py, available under $PATH. This allows for something like: docker-compose exec storage-target rpc.py nvmf_get_subsystems Note that SPDK environment heavily depends on a running kernel hence all the containers need to be privileged. That said, to make sure containers are not affecting the host too much, some tasks must be done prior running them. This includes: - loading proper kernel modules (like nvme-fabrics, etc.) - allocating hugepages and having at least one hugetlbfs mount available under /dev/hugepages base_build is created as docker multi-stage build. This is done in order to decrease the size of the final image. The SPDK RPMs are built inside a base image and then copied over to the main image (+ fio binary) - this leaves all the dependencies inside the intermediate image instead of the final one. The resulted difference in size may look similar to the following (it may differ depending on the docker version etc.): no multi-stage build: spdk_base == 1.04GB multi-stage build: spdk_base == 261MB Signed-off-by: Michal Berger <michalx.berger@intel.com> Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com> Change-Id: I825bd0d0bb4071bd9d44b6a0749c033894899ae0 Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9055 Reviewed-by: Jim Harris <james.r.harris@intel.com> Reviewed-by: Monica Kenguva <monica.kenguva@intel.com> Reviewed-by: Xiaodong Liu <xiaodong.liu@intel.com> Reviewed-by: Changpeng Liu <changpeng.liu@intel.com> Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
2021-08-03 07:16:49 +00:00
~~~
As each container includes SPDK installation it is possible to use rpc.py to
examine the final setup. E.g.:
~~~{.sh}
docker-compose exec storage-target rpc.py bdev_get_bdevs
docker-compose exec proxy-container rpc.py nvmf_get_subsystems
~~~
## Caveats
- If you run docker < 20.10 under distro which switched fully to cgroups2
(e.g. f33) make sure that /sys/fs/cgroup/systemd exists otherwise docker/build
will simply fail.
- Each SPDK app inside the containers is limited to single, separate CPU.