The VIA [INTEL], [INTEL2] interface model was developed based on the observation by researchers that the most significant overhead was the time required to communicate between processor, memory, and I/O subsystems that are directly connected to a network. In particular, communication time is not scaling to each individual component of the whole system and can possibly increase exponentially in a cluster of workstations.
This communication overhead is caused by the time accumulated when messages move through different layers of the Internet protocol suite of TCP/UDP/IP in the operating system. In the past, the overhead of end-to-end Internet protocols did not significantly contribute to the poor network performance since the latency equation was primarily dominated by the underlying network links. However, recent improvements in network technology and processor speeds has made the overhead of the Internet protocol stacks the dominant factor in the latency equation.
The VIA specification defines mechanisms to avoid this communication bottleneck by eliminating the intermediate copies of data. This effectively reduces latency and lowers the impact on bandwidth. The mechanisms also enable a process to disable interrupts under heavy workloads and enable interrupts only on wait-for-completion. This indirectly avoids context switch overhead since the mechanisms do not need to switch to the protocol stacks or to another process.
The VIA specification only requires control and setup to go through the OS kernel. Users (also known as VI Consumers) can transfer their data to/from network interfaces directly without operating system intervention via a pair of send and receive work queues. And, a process can own multiple work queues at any given time.
VIA is based on a standard software interface and a hardware interface model. The separation of hardware interface and software interface makes VI highly portable between computing platforms and network interface cards (NICs). The software interface is composed of the VI Provider library (VIPL) and the VI kernel agent. The hardware interface is the VI NIC which is media dependent. By providing a standard software interface to the network, VIA can achieve the network performance needed by communication intensive applications.
VIA supports send/receive and remote direct memory access (RDMA) read/write types of data movements. These operations describe the gather/scatter memory locations to the connected VI. To initiate these operations, a registered descriptor should be placed on the VI work queue. The current revision of the VIA specification defines the semantics of a DMA Read operation but does not require that the network interface support it.
The VI kernel agent provides synchronization by providing the scheduling semantics for blocking calls. As a privileged entity, it controls hardware interrupts from the VI architecture on a global, and per VI basis. The VI kernel agent also supports buffer registration and de-registration. The registration of buffers allows the enforcement of protection across process boundaries via page ownership. Privileged kernel processing is required to perform virtual-to-physical address translation and to wire the associated pages into memory.
