JLab RDBbased Archiver Matt Bickley work by Anthony
JLab RDB-based Archiver Matt Bickley (work by Anthony Bavuso and Chris Slominski)
What? Another Archiver? • Scalability Requirement to support monitoring of 1 M channels • Manageability Requiring no interaction by administrators • Design issues – Data latency (10 -20 minutes from Czar) – Independent deadbanding (not using ADEL) – Data loss EPICS Collaboration Meeting, October 2007
Data loss from Czar • Users saw transients in Strip. Tool that were not recorded by Czar – IOC CPU loading less than 30% – Czar system loading less than 20% • Confirmed data loss with a test setup Analyzer Archive Czar Test IOC FFFF FFFE FFFC. . CA Server FFFF FFFE FFFD FFFC. . EPICS Collaboration Meeting, October 2007 Database
MYA Design User Copycat Thread Watchdog Thread Commands Work Queue Main Thread Work Queue Scribe Thread My. SQL Database EPICS Thread Work Control System Queue Scribe Thread EPICS Collaboration Meeting, October 2007
MYA realtime thread scheduling • Prioritization, highest to lowest – Channel Access communication – Database scribe threads – Client communication, watchdog, etc. • No loss of data • If work queue > 2 M events then database insertions are prioritized – Peak queue length 160, 000 events – Average queue length 5, 000 events EPICS Collaboration Meeting, October 2007
System Hardware • Dell Power. Edge 2850 – Dual quad-core CPUs – 16 GB memory – ~$8 K in 2006 ($4 K for the memory) • Eon. Store RAID box – 16 300 GB SCSI disks (~2 TB of storage) – RAID 0+1 – ~$21 K in 2006 ($17 K for the disks) EPICS Collaboration Meeting, October 2007
Performance • • 50, 000 channels monitored 25, 000 updates/second 4, 000 database insertions/second History data retrieval rate of 160, 000/second (Get one day of a 1 HZ channel’s data in 0. 5 seconds) • Store at least 1 year of data online EPICS Collaboration Meeting, October 2007
What’s Next… • Cut over to MYA for operations • “Sandbox” archiver – Users specify archive requests – Limited data retention (up to 1 month) – Limited archive request lifetime OR • Archive everything – 300, 000+ channels – Keep the data for 1 week (? ) – Needs more memory, much less disk space EPICS Collaboration Meeting, October 2007
- Slides: 8