RDMA promises to remove speed barriers
 
|
|||
 | |||
|
Sign up to receive this and other networking newsletters in your inbox.
As LANs increase in speed, technologists continue to try to remove some of the barriers to reaching high speeds that exist on network end points.
One of the barriers that is expected to become more of a limiting factor is memory bandwidth. While CPU performance in servers or workstations increases along the well-known curve of Moore's Law, computers haven't made similar advances in copying data to memory. Multiprocessor machines and clusters are most affected by the shortcoming.
As 10 Gigabit Ethernet grows in prominence, vendors are looking at ways to get around the memory bottleneck. If they don't get around it, end stations would be hard pressed to ever use the full power of 10 Gig.
One way to do this would be through remote direct memory access, or RDMA. RDMA is a way of moving data from the memory of one computer over a network and into the memory of a second computer, all with minimal involvement from the CPUs of either machine. This saves on memory bandwidth as well as CPU cycles.
In May, a group of vendors formed the Remote Direct Memory Access Consortium and began work on a method for providing RDMA over TCP/IP. RDMA has been used in proprietary ways, but a standard is really necessary for it to take off. So, the group intends to take its work to the Internet Engineering Task Force for standardization.
The founding members are Adaptec, Broadcom, Hewlett-Packard, IBM, Intel, Microsoft and Network Appliance.
RELATED LINKS
RDMA over IP, the problem statement
In addition to writing this newsletter, Jeff Caruso edits Network World's e-mail newsletters from his office on New York's Long Island. If you would like to make suggestions about newsletter format or content, or even just express your opinion on today's topic, you can reach Jeff at jcaruso@nww.com.
High Speed LANs archive
Past newsletters.
