Introduction

• memory system to exploit RDMA
• expose memory of many machines as shared address space
• transactions to allocate/read/write/free
• fast message passing primitive
• local memory is still 23x faster

FaRM

• one-sided rdma reads to access data
• rdma writes for the message passing primitive
  • circular buffer to implement unidirectional channels
  • sender writes messages to buffer tail, and advanced the tail pointer
  • never write past copy of HEAD
• TCP latency ceiling much higher
• issues with RDMA: significant perf hit as more memory registered
  • PhyCo kernel driver to allocate physically contiguous+aligned memory regions
• share queue pairs per machine to reduce number of queues on the NIC

API etc

• all operations must alloc a new transaction
• tx[Alloc|Free|Read|Write|Commit]
• 2GB pages are the unit of recovery, registration, etc
• consistent hashing to map region identifier to machine storing the object
• optimisitic concurrency control w/ 2PC
  • tx ctx record version numbers
  • machine running txn is coordinator
    • send prepare to all participant (primaries and replicas of object in write set)
    • primary locks modified objects
    • all primary+replica logs message before sending reply
    • coordinator sends validate message to primaries in read set to check if the latest version was read -