FaRM

Article.

• Super fast with DRAM.

• All within one data center. Doesn't solve geo replication anyhow.

• Bottleneck - CPU.

• Sharded by key between primary/backup pairs.

• All replicas are updated if there's change. You always read from primary. If one replica is available you are good.

• All must fit in the (combined) RAM.

• Tolerate power failures - non-volatile RAM scheme (NVRAM).

• RDMA - network interface cards that can read/write memory directly.

• Kernel Bypass - app code can access network cards directly, without kernel help.

• When the battery system detects main power failure - stop all processes, write all RAM to SSD. Then die.

• RDMA - remote direct memory access. Access memory remotely without involving destination CPU at all. Append messages to a queue.

• API:

  • txCreate()

  • o = txRead(oid)

  • o.f += 1

  • txWrite(oid, o)

  • ok = txCommit()