Hello,
I am writing to ask if at Red Hat you would be interested to collaborate on container image standard improvements under Open Container Initiative ( https://opencontainers.org/).
At Huawei we identified that container image specification lacks compatibility requirements. For instance if you run a container that requires an NVIDIA GPU then you have to make sure that CUDA library in the container matches the CUDA driver on the host. The required version of the driver could be expressed in the container image. There are plenty of similar use cases when it comes to kernel configuration, boot args, modules or out-of-tree drivers.
More details could be found here: https://docs.google.com/document/d/1lzwh8DGMu5vXXHwJmnewYIMffkcOEvH8owX4UYjR...
I think OpenShift and its Machine Config operator could benefit from this by influencing worker node configuration based on the image specification, which would mean that customers would also benefit.
I am forming a working group under OCI and believe this initiative should not come only from one company since that will be a global standard.
I prepared a template for a working group that will be voted on by the Technical Oversight Board (https://github.com/opencontainers/tob). https://docs.google.com/document/d/1WZbr7xEpUohvyIiSvJEgCCxhAVMM9DmgXBsk1q1M...
The voting will take place once I create a PR. I think, after the working group is approved, we would need 2-3 months to come up with the first standard and a client tool to build and validate the specification.
If you are interested we can schedule a meeting to discuss this.
Regards, Marcin
Hi,
On Tue, Sep 5, 2023, at 9:26 AM, Marcin Franczyk wrote:
Hello,
I am writing to ask if at Red Hat you would be interested to collaborate on container image standard improvements under Open Container Initiative (https://opencontainers.org/).
At Huawei we identified that container image specification lacks compatibility requirements. For instance if you run a container that requires an NVIDIA GPU then you have to make sure that CUDA library in the container matches the CUDA driver on the host. The required version of the driver could be expressed in the container image. There are plenty of similar use cases when it comes to kernel configuration, boot args, modules or out-of-tree drivers.
More details could be found here: https://docs.google.com/document/d/1lzwh8DGMu5vXXHwJmnewYIMffkcOEvH8owX4UYjR...
I added a comment in https://docs.google.com/document/d/1lzwh8DGMu5vXXHwJmnewYIMffkcOEvH8owX4UYjR...
I think OpenShift and its Machine Config operator could benefit from this by influencing worker node configuration based on the image specification, which would mean that customers would also benefit.
Hmm; you're talking about changing the host OS depending on what gets scheduled to it? That'd make everything kind of loopy =) Normally one instead schedules workloads to compatible hosts, requiring the hosts to be prepared out of band.
I am forming a working group under OCI and believe this initiative should not come only from one company since that will be a global standard.
Right.
I prepared a template for a working group that will be voted on by the Technical Oversight Board (https://github.com/opencontainers/tob). https://docs.google.com/document/d/1WZbr7xEpUohvyIiSvJEgCCxhAVMM9DmgXBsk1q1M...
The voting will take place once I create a PR. I think, after the working group is approved, we would need 2-3 months to come up with the first standard and a client tool to build and validate the specification.
What's the expected time commitment from the working group?
Let me see if I can find others at Red Hat who may be interested in this too.
Hi,
Hmm; you're talking about changing the host OS depending on what gets
scheduled to it? That'd make everything kind of loopy =) Normally one instead schedules workloads to compatible hosts, requiring the hosts to be prepared out of band.
No. I am not saying to reconfigure a node every time you have a different container with a different set of requirements. What I am saying is that we will be working on a data source that can be used by different models to configure/provision worker nodes, and that could be used on a running cluster. It would be up to you to decide how and when you should feed Machine Config operator with this. It doesn't even have to be automatic. The data could be used to identify a specific group of applications with similar compatibility requirements (like enablement of SR-IOV) to configure a target group of nodes.
What's the expected time commitment from the working group?
Join the meeting once a week and express your ideas/concerns. How much time you would like to spend on tools implementation is up to you. For now the most important thing is to collect feedback/opinions/ideas from the various stakeholders.
Let me see if I can find others at Red Hat who may be interested in this
too.
Thanks.
Regards, Marcin
czw., 21 wrz 2023 o 14:52 Colin Walters walters@verbum.org napisał(a):
Hi,
On Tue, Sep 5, 2023, at 9:26 AM, Marcin Franczyk wrote:
Hello,
I am writing to ask if at Red Hat you would be interested to collaborate on container image standard improvements under Open Container Initiative (https://opencontainers.org/).
At Huawei we identified that container image specification lacks compatibility requirements. For instance if you run a container that requires an NVIDIA GPU then you have to make sure that CUDA library in the container matches the CUDA driver on the host. The required version of the driver could be expressed in the container image. There are plenty of similar use cases when it comes to kernel configuration, boot args, modules or out-of-tree drivers.
More details could be found here:
https://docs.google.com/document/d/1lzwh8DGMu5vXXHwJmnewYIMffkcOEvH8owX4UYjR...
I added a comment in https://docs.google.com/document/d/1lzwh8DGMu5vXXHwJmnewYIMffkcOEvH8owX4UYjR...
I think OpenShift and its Machine Config operator could benefit from this by influencing worker node configuration based on the image specification, which would mean that customers would also benefit.
Hmm; you're talking about changing the host OS depending on what gets scheduled to it? That'd make everything kind of loopy =) Normally one instead schedules workloads to compatible hosts, requiring the hosts to be prepared out of band.
I am forming a working group under OCI and believe this initiative should not come only from one company since that will be a global standard.
Right.
I prepared a template for a working group that will be voted on by the Technical Oversight Board (https://github.com/opencontainers/tob).
https://docs.google.com/document/d/1WZbr7xEpUohvyIiSvJEgCCxhAVMM9DmgXBsk1q1M...
The voting will take place once I create a PR. I think, after the working group is approved, we would need 2-3 months to come up with the first standard and a client tool to build and validate the specification.
What's the expected time commitment from the working group?
Let me see if I can find others at Red Hat who may be interested in this too. _______________________________________________ CentOS-devel mailing list CentOS-devel@centos.org https://lists.centos.org/mailman/listinfo/centos-devel
I forgot to add one more info..
The PR for the working group has been already opened https://github.com/opencontainers/tob/pull/128. Going through the conversations and the document there will also help to get a better idea of the whole thing.
Regards, Marcin
czw., 21 wrz 2023 o 15:58 Marcin Franczyk marcin0franczyk@gmail.com napisał(a):
Hi,
Hmm; you're talking about changing the host OS depending on what gets
scheduled to it? That'd make everything kind of loopy =) Normally one instead schedules workloads to compatible hosts, requiring the hosts to be prepared out of band.
No. I am not saying to reconfigure a node every time you have a different container with a different set of requirements. What I am saying is that we will be working on a data source that can be used by different models to configure/provision worker nodes, and that could be used on a running cluster. It would be up to you to decide how and when you should feed Machine Config operator with this. It doesn't even have to be automatic. The data could be used to identify a specific group of applications with similar compatibility requirements (like enablement of SR-IOV) to configure a target group of nodes.
What's the expected time commitment from the working group?
Join the meeting once a week and express your ideas/concerns. How much time you would like to spend on tools implementation is up to you. For now the most important thing is to collect feedback/opinions/ideas from the various stakeholders.
Let me see if I can find others at Red Hat who may be interested in this
too.
Thanks.
Regards, Marcin
czw., 21 wrz 2023 o 14:52 Colin Walters walters@verbum.org napisał(a):
Hi,
On Tue, Sep 5, 2023, at 9:26 AM, Marcin Franczyk wrote:
Hello,
I am writing to ask if at Red Hat you would be interested to collaborate on container image standard improvements under Open Container Initiative (https://opencontainers.org/).
At Huawei we identified that container image specification lacks compatibility requirements. For instance if you run a container that requires an NVIDIA GPU then you have to make sure that CUDA library in the container matches the CUDA driver on the host. The required version of the driver could be expressed in the container image. There are plenty of similar use cases when it comes to kernel configuration, boot args, modules or out-of-tree drivers.
More details could be found here:
https://docs.google.com/document/d/1lzwh8DGMu5vXXHwJmnewYIMffkcOEvH8owX4UYjR...
I added a comment in https://docs.google.com/document/d/1lzwh8DGMu5vXXHwJmnewYIMffkcOEvH8owX4UYjR...
I think OpenShift and its Machine Config operator could benefit from this by influencing worker node configuration based on the image specification, which would mean that customers would also benefit.
Hmm; you're talking about changing the host OS depending on what gets scheduled to it? That'd make everything kind of loopy =) Normally one instead schedules workloads to compatible hosts, requiring the hosts to be prepared out of band.
I am forming a working group under OCI and believe this initiative should not come only from one company since that will be a global standard.
Right.
I prepared a template for a working group that will be voted on by the Technical Oversight Board (https://github.com/opencontainers/tob).
https://docs.google.com/document/d/1WZbr7xEpUohvyIiSvJEgCCxhAVMM9DmgXBsk1q1M...
The voting will take place once I create a PR. I think, after the working group is approved, we would need 2-3 months to come up with the first standard and a client tool to build and validate the specification.
What's the expected time commitment from the working group?
Let me see if I can find others at Red Hat who may be interested in this too. _______________________________________________ CentOS-devel mailing list CentOS-devel@centos.org https://lists.centos.org/mailman/listinfo/centos-devel
-
Colin Walters -
Marcin Franczyk