When a process update enters the second part of commit phase a server level lock is taken. On our system with around ~125,000 metadata objects and attached to a SAN the operation takes 3 minutes 22 seconds. Further detailed information on our process update can be found at http://blog.calvett.co.uk/2013/02/21/the-anatomy-of-a-process-update-part-2/Our tests show that the commit is single threaded and that each delete is taking ~2ms and the only way we can improve the situation is to use SSD's to move delete performance into the nanosecond range but this is not viable.We need to reduce the duration of the server lock so that our users do not experience what they consider to be a server outage.
Operating System Language
Steps to Reproduce