archlitchi
commented
3 weeks ago yes, it seems ascend-for-volcano use a FIFO strategy with 'bind-time timestamp' instead of using nodelock, and works properly, we can try that as well
Open lengrongfu opened 3 weeks ago
archlitchi
commented
3 weeks ago yes, it seems ascend-for-volcano use a FIFO strategy with 'bind-time timestamp' instead of using nodelock, and works properly, we can try that as well
Currently, there are some problems with using nodelock that cannot be solved.
Current hami scheduler use nodelock to implement serial bind pod, and release the nodelock only after device-plugin allocate device succeeds, but