volcano
volcano copied to clipboard
volcano-sh
volcano性能验证
volcano号称是若是10000节点，会有 2000pods/s的调度，我想请问的是针对这个数据，你们是怎么得出来的？我现在想要验证它的话，需要怎么做？有没有验证教程？ https://github.com/volcano-sh/volcano/blob/master/docs/Volcano%20-%20A%20Kubernetes%20Native%20Batch%20System.pdf
I have a plan to write a mock-scheduler on volcano, to do scheduling performance test for large scale nodes and pods.
I have a plan to write a mock-scheduler on volcano, to do scheduling performance test for large scale nodes and pods.
how to do？
I have a plan to write a mock-scheduler on volcano, to do scheduling performance test for large scale nodes and pods.
how to do？
Just use fake client to replace apiserver when binding to nodes. Other scheduling logic is same as origin.
I have a plan to write a mock-scheduler on volcano, to do scheduling performance test for large scale nodes and pods.
Do you have more detailed information about this plan? I'd like to see if I can help.Because recently we are also planning to test the performance of Volcano.
Do you have more detailed information about this plan? I'd like to see if I can help.Because recently we are also planning to test the performance of Volcano.
We can use fake instance to simulate those interface in scheduler cache and leave the whole scheduing logic same as volcano scheduler. To simulating the scheduler, those steps are need:
- make sure all clients are use interface: please help to merge https://github.com/volcano-sh/volcano/pull/2866 first. @william-wang
- then define a MockCache similar to Cache in volcano-scheduler
- deveop a client(cli, http api or sdk) which support CRUD operations to manage fake resource such as pod/job/podgroup/node, etc. This step may take some time.
- creating large number of those resource and wait mock-scheduler to scheduling those pods, statistic those metrics data.
There is also anothter simple way: use https://github.com/kubernetes-sigs/kwok to replace real k8s nodes. Then create large number os fake-nodes and submit large numbers of pods to those fake-nodes and waiting volcano scheduling.
Do you have more detailed information about this plan? I'd like to see if I can help.Because recently we are also planning to test the performance of Volcano.
We can use fake instance to simulate those interface in scheduler cache and leave the whole scheduing logic same as volcano scheduler. To simulating the scheduler, those steps are need:
- make sure all clients are use interface: please help to merge fix: events miss when there is a pod bind failed #2866 first. @william-wang
- then define a MockCache similar to Cache in volcano-scheduler
- deveop a client(cli, http api or sdk) which support CRUD operations to manage fake resource such as pod/job/podgroup/node, etc. This step may take some time.
- creating large number of those resource and wait mock-scheduler to scheduling those pods, statistic those metrics data.
There is also anothter simple way: use https://github.com/kubernetes-sigs/kwok to replace real k8s nodes. Then create large number os fake-nodes and submit large numbers of pods to those fake-nodes and waiting volcano scheduling.
ok，I am currently working on it.
volcano号称是若是10000节点，会有 2000pods/s的调度，我想请问的是针对这个数据，你们是怎么得出来的？我现在想要验证它的话，需要怎么做？有没有验证教程？
Is there any background on that requirement? For 2k pods/s, it not only about scheduling, but also about kube-apiserver enhancement which is not open source right now.
Is this topic still active? I would also like to discuss if there is any information available in this regard. I'm wondering if we can use KWOK here approach to simulate and validate scheduling performance.
