Features - Enterprise Data Insights:

IS TCP/IP CLOGGING YOUR WINDOWS SERVERS?
By Joe Gervais, Director Of Marketing, Alacritech

In today's on-demand computing environments, the focus of enterprise systems has become data centric, and the timely delivery of information is critical for businesses. Applications continue to grow in size, average e-mail messages are larger, and storage needs are exploding at an alarming rate. TCP/IP is the universal language used for delivering all this data, with nearly a billion systems using the protocol. IT managers with constrained budgets are under increased pressure to accelerate data delivery using existing server and storage resources. As these demands on IT resources increase, enterprise managers are looking for new technologies that can increase the performance and efficiency of their existing systems to solve this information delivery crisis.

The key to understanding today's networking bottlenecks is to look at the protocols used for server communications. TCP/IP was developed over 25 years ago to allow reliable data transfer between systems over slow wide area network links. Today, TCP/IP runs over Ethernet Local Area Networks (LAN) for access to the Internet as well as system-to- system communications. These LANs typically use high-speed Ethernet switches to enable efficient hardware-based packet switching. On these LANs there is a large installed base of applications that utilize TCP/IP, including most Microsoft Windows networking applications.

What Causes The Bottleneck?

When a Windows application sends large amounts of information across a network, the system must perform TCP/IP and Ethernet processing to handle the data transfers. The Ethernet packet, or envelope that data is packaged within to move across a network, is very small. The last decade saw small objects being sent across the network, so performing TCP/IP processing in software was sufficient. With today's on-demand data consumption brought on by the advent of the Internet, TCP/IP is fed larger objects, which must be encapsulated into packets with a fixed maximum size for Ethernet transmission. Therefore, a single application transaction turns into many TCP/IP transactions. The software TCP/IP processing used to move data consumes significant server resources, which causes bottlenecks when moving today's larger quantities of data. As TCP/IP and Ethernet will remain dominant, IT managers are faced with finding a way to make their servers process TCP/IP more efficiently.

Efficient Data Delivery

Several companies have developed products that implement TCP protocol processing in hardware. The basic concept of these TCP/IP offload engine (TOE) based products is two-fold. Hardware, known as direct memory access (DMA) engines are used to move these large data objects, rather than the server CPU. DMA engines, used extensively in storage and graphics products, are much more efficient than a processor for copying or moving networked application data. Dedicated hardware circuits then disassemble and reassemble these large data objects, making the system oblivious to the tiny packet size used by Ethernet. Network transactions by applications no longer generate large numbers of server interactions to the network adapter. This drastic reduction of transactions by the TOE has significant impact on system performance; even with modest application I/O sizes making the system perform as if it has twice the number of processors.

Removing the Burden

TOE products that do hardware TCP protocol processing come in two forms-an iSCSI host bus adapter (HBA), which only handles block storage, and the more flexible TOE NIC (TNIC), designed to provide TCP/IP acceleration for both network and storage traffic, including iSCSI. The TNIC benefits all Microsoft Windows network applications or services, including file services, backup applications, and iSCSI block storage access by removing the network protocol overhead. By removing this burden, existing server farms increase both their network efficiency and performance.

Another significant, but often overlooked, burden on recurring operating costs in the data center is the power consumption of individual servers. The sheer number of enterprise servers used in a 24-7 operating environment makes overall electricity requirements quite substantial. Beyond just potential power savings from server consolidations, servers can also be made more energy efficient by leveraging TNICs for further efficiency. For a repetitive task, dedicated hardware is far more efficient than software on a general-purpose processor.

Extend Resources, Reduce Costs

Hardware designs, as opposed to software designs, provide the optimum efficiency for TCP/IP protocol processing. The common belief is one megahertz of processor speed is required for each megabit per second of software TCP/IP processing. To make full use of a Gigabit Ethernet link, a 100+ watt, 3 GHz Intel Xeon processor is required. A TNIC using TOE chip delivers better performance in a much smaller footprint while using less than 3 watts of power. Deployed across an enterprise's large number of Windows servers, TNICs significantly reduce electricity consumption and minimize subsequent environmental impact. A significant reduction in operating costs can be achieved simply by using TNIC to consolidate servers and extend resources.

With the advent of the World Wide Web, a fundamental shift from processing data to moving data has occurred in Windows enterprise computing. Given the current economic climate, budgets are more limited than they were at the end of the last decade. With TNICs, IT organizations can ensure efficient data delivery, while preserving investments in servers and networking equipment. With millions of Windows-based servers worldwide running TCP/IP over Ethernet, the hardware-based TCP/IP processing of new TNIC products provide the answer to achieving higher operating efficiency.

About Joe Gervais

Joe Gervais is the director of marketing at Alacritech Inc. His articles have appeared in LAN Times, Netware Connections, Network World and Novell AppNotes.