TCP/IP protocol accelaration

TCP/UDP/IP and Ethernet have become undisputed leaders of network communication since last decade. TCP/IP stacks are usually implemented in operating system software and packets are handled by the main (host) processor due to which protocol processing of incoming and outgoing network traffic consumes processor cycle. Usually in the application of data transfer, there are several copies of data from user space, kernel space and then NIC which are one of the most CPU hungry operations and hog the CPU upto 80-85%, add latency, consume memory bus bandwidth, and require host processor (CPU) intervention. Zero-copy describes computer operations in which the CPU does not perform the task of copying data from one memory area to another. Using this concept, good data transfer speed can be achieved. If an application has the capability to directly build over the wire data, without kernel intervention, multiple copies could be avoided. The data transfer speed can be increased if kernel is bypassed. The kernel can be bypassed if the application itself can do the functions done by kernel. This can be achieved by providing a separate functionality in the application that performs the task of appending header and checksum. Thus the packet generated by this can directly be sent over the wire. Also every Ethernet card has multiple tx rings amongst which many are not used in most cases. Some of these unused tx rings can be exported to some special purpose application to reduce the copy overhead.