BeeGFS User Group (BUG), Atlanta
202411181959
Status: #idea
Tags: SC’24
BeeGFS User Group (BUG), Atlanta
Licensing changes
- Storage pools and mirroring moves to enterprise with BeeGFS v8
- Questions
- How we can get access to Hive
Development Updates
- Ubuntu 24.04 (Linux 6.8) supported
- New monitoring dashboards
- Will serve as the basis for tuning reccomendations
- Setup InfluxDB and Grafana. Makes it easier to get support as well
- BeeGFS Index
- Now, BeeGFS-specific metadata is stored in BeeGFS index
BeeGFS UseCases
xiRAID
- Questions
- How does xiRAID save a PCIe slot? Compared to a hardware card?
- How did we get 2 drive failures per OSS? Is it 2 per OST?
- How does re-striping work? Will data-parity ratio change for existing data?
- Booth #4603
BeeGFS v8
- Major version upgrade with a couple of incompatible changes.
- Adding support to offload data to other storage technologies (like S3)
- Expected in early next year
- Questions
- Is the TOML configuration file limited to the mgmtd only?
- What is BeeGFS watch?
- A pub-sub service for BeeGFS events. Can write our own subscribers with intelligence. Uses protobufs to transmit events
New Control Tool
- Questions
- Is there a better way to specify subnets that the services should use?
- Can put multiple interfaces in the files. Will use them as fallback
- Is there a machine readable format for beegfs bench?
- Not yet. Took as a feature request
- Changing the stripe count, does it effect the existing files?
- No
- Data inlining?
- Not yet. Internal discussions
- Is there a better way to specify subnets that the services should use?
BeeGFS Watch
- Replaces BeeGFS
event-listener - 1 watch per meta server
- Maintains a cursor in the event queue for each subscriber
- Additional metadata
- Last close on the file
- Currently, no flushing of watch data to disk. Will add later
BeeGFS Remote Storage Targets (Joe)
- Questions
- What happens if there is high amount of OST load? Are remote jobs de-prioritised?
- Right now treated the same other files. But we can create a user which only does sync jobs and BeeGFS will round robin between the users.
- What happens if there is high amount of OST load? Are remote jobs de-prioritised?
TODO
- intel’s pmem syscall-intercept for intercepting syscalls
- Follow it up by using BeegFS ioctl to actually set the file properties during the sync - Email Troy about a license for BeeGFS