Remote direct memory access: Difference between revisions

Revision as of 10:07, 9 January 2007

Revision as of 22:04, 24 May 2007

Line 3:

RDMA supports [[zero-copy]] networking by enabling the [[network adapter]] to transfer data directly to or from application memory, eliminating the need to copy data between application memory and the data buffers in the operating system. Such transfers require no work to be done by [[Central processing unit|CPUs]], [[cache|caches]], or [[context switch|context switches]], and transfers continue in parallel with other system operations. When an application performs an RDMA Read or Write request, the application data is delivered directly to the network, reducing latency and enabling fast message transfer.

This strategy presents several problems related to the fact that the target node is not notified of the completion of the request (1-sided communications). The common way to notify it is to change a memory byte when the data has been delivered, but it requires the target to poll on this byte. Not only does this polling consume CPU cycles, but also the memory footprint and the latency increases linearly with the number of possible other nodes. These issues and the fact that the RDMA programming model is very different from the one that is generally used in the High-Performance Computing world ([[MPI]], the [[Message Passing Interface]]) explains why the RDMA success is very limited in HPC. The Send/Recv model used by other [[zero-copy]] HPC interconnects such as [[Myrinet]] or [[Quadrics]] does not have any of these problems and presents as good performance since their native programming interface is very similar to MPI.

RDMA reduces the need for [[Protocol (computing)|protocol]] overhead, which can squeeze out the capacity to move data across a network, reducing [[Network performance|performance]], limiting how fast an application can get the data it needs, and restricting the size and scalability of a [[Computer cluster|cluster]].