|
Our users run a variety of application with different I/O characteristic. Some applications have well-formed large sequential I/O, while others can have small random I/O access patterns. Lustre is the principal scratch filesystem and there are six individual filesystems. Each user is assigned to one filesystem. This helps to distribute storage allocations and utilization across all filesystems. But, there are occasions when a single user’s code adversely impact the whole filesystem. This results in poor filesystem throughput effecting job runtimes and poor interactive response for all users.
Although TBF provides some of the QoS functionally, it doesn’t meet all our requirements. We would like these additional capablilies added to TBF NRS.
- Add user id tracking
- Default value for all users
- Allow of setting of token buckets for each user
- Add a control trigger for enforcement of tokens. (ie RPC queue size or RPC queue wait)
|