Vault 2017

Infiniband network block device (IBNBD) uses RDMA to transfer block IO bypassing SCSI sublayer. The driver implements client-side server memory management and requires only two infiniband messages per read or write IO on the wire. We take advantage of the infiniband "immediate" field to inform a receiver where the transferred data can be found. The per-cpu infiniband connections on client side allow for an IO-response to be processed on the same cpu, the IO was submitted on. The block MQ shared tags feature is used to fairly share the resource among multiple mapped devices. From the user perspective only remote device path and ib port gid of the server are required on client side for mapping, while no configuration on server is necessary. In the presentation we describe the simplistic design of the driver, its application in cloud infrastructure and compare it with existing solutions.