Next gen Broadcom Gig Ethernet controller arrives

RIGHT BEFORE heading to E3 in Los Angeles, I made a short stop in Las Vegas for Networld + Interop 2004 show. One of the hottest products being shown off was one of Broadcom's latest product offerings. Only introduced on May 10th, the NetXtreme II, a next generation Gigabit Ethernet controller, which includes an integrated transmission control protocol/Internet protocol (TCP/IP) offload engine, iSCSI host bus adapter functionality, remote direct memory access (RDMA) technology, and remote system management.

This is one of the first controllers in the industry to simultaneously perform storage networking, high-performance clustering, accelerated data networking and remote system management pass-through functions. The model number for the NetXtreme II is the BCM5706. The $35 chipset price is comparable to existing Gigabit Ethernet controllers. Additionally the BCM5706 is pin- and layout-compatible with earlier generations of the Broadcom's NetXtreme product line. The 400-pin PBGA package (21 x 21 mm) is currently sampling to OEMs. Volume production and shipping is expected to commence in the fourth quarter of this year.

The emergence of converged network interface controllers, which is commonly abbreviated as C-NIC, will provide accelerated client/server, clustering, and storage networking environment enabling the use of unified TCP/IP Ethernet communications. The current focus of the current generation of C-NIC products is to integrate each of the main hardware and software components of the IP protocol suite offload into a unified whole on-chip.

The NetXtreme II has TCP protocol offload supporting Chimney specification, RDMA over TCP offload support, the Microsoft iSCSI offload, and the capability of embedded server management. While it is commonly known and recognized that the server network I/O bottleneck is the greatest factor in impacting data center application performance and scalability. The major problem with current technologies is that network bandwidth and client/server traffic loads have outpaced and will continue to consistently outpace CPU performance increases.

The standard solution around this problem has been to use a range of different networking technologies optimized for specific traffic types. Ethernet has been used for client/server communications and file-based storage, while Fibre Channel for block-based storage. Other special purpose low-latency protocols have been created and used for server clustering. The problem with using separate technologies has been the implementation, and cost. This problem is highly visible in blade servers and virtualized server systems.

TOE support is one of the six key features on the NetXtreme II. TOE refers to the TCP/IP protocol stack being offloaded to a dedicated controller in order to reduce TCP/IP processing overhead in servers. Although TOE technology has been the focus of significant vendor engineering investment, a number of obstacles remain for broad-based TOE deployment.

The first generation of TOE products suffered from two related drawbacks. One is the lack of standard interfaces within server operating systems for TOE integration. The other obstacle is an all-inclusive approach to architecting TCP offload as a parallel server network protocol stack. The recent announcement of Microsoft's Chimney Offload Architecture allows for TOE integration within Windows server operating systems. This will allow for the development of TOE products that are directly supported by the operating system providing optimum server performance gains while lowering the cost of ownership.

The idea behind RDMA is a technology that allows the network card, under the control of the application, to place data directly into and out of application memory. This removes the need for data copying and adds support for low-latency communications. Current RDMA networking has until now only been used in the implementation of specialized interconnects such as InfiniBand.

For those who have not been keeping up for storage networking over TCP/IP, iSCSI is designed to enable end-to-end solution over TCP/IP networks. iSCSI, is a transport protocol for SCSI that operates on top of TCP through encapsulation of SCSI commands in a TCP data stream. As an alternative to parallel SCSI or Fibre Channel within the data center provides I/O transport for a range of applications. The two available options for supporting iSCSI initiators are through a software-only initiator, with a standard Gigabit Ethernet NIC or an iSCSI host bus adapter. iSCSI adapters typically use dedicated controllers to fully offload iSCSI and TCP/IP protocols from server host CPU.

The most important benefit of the NetXtreme II is its improved operational efficiency. By implementing C-NICs into a production environment, the interfaces to each server and rack are simplified. This integration provides for fewer connection points, fewer cables, fewer adapter cards, and easier upgrades to existing networks. All the changes are applied to the C-NIC, which may reduce acquisition and ongoing maintenance and management expenses.

C-NICs will provide a simplified upgrade path to 10 GbE (10 Gb Ethernet), with the addition of a single module to each server. The benefits of increased server and network performance, compared to existing Gigabit NICs make their debut a valuable tool for network administrators. µ