Background: SMPng Challenges

• Substantial redesign for new synchronization

  • Prior “implicit” synchronization relied on many undocumented assumptions
  • Was fault-prone even in 4.x w/o SMP

• Trade off “granularity” and “locking cost”

  • Don't unnecessarily invest in granularity that doesn't buy under-utilized parallelism

• Provide a migration path that allows for iterative kernel update to new approach