Network emulation tools in IP networks: test system infrastructure before deployment - Connectivity

Computer Technology Review, August, 2002 by Garth Morrison

Storage solutions based on Internet protocol (IP) network technology offer new options to solve the storage conundrum, but an undercurrent of concern exists around the impact of IP networks on system performance. How can if professionals be sure that these new technologies work in the real world before they deploy them into their infrastructures? How can the integrator and VAR community add value to end-user customers to lift the veil of fear, uncertainty, and doubt surrounding these new IP storage products and increase their adoption rate?

Understanding the drivers of storage performance and the limitations imposed by IP networks will help IT managers make smart decisions when designing their IF storage systems. But the only way to really assess the impacts of IP networks on storage performance is to test them using network emulation tools that allow users to recreate lifelike network conditions without the cost and hassle of procuring the network up-front.

IP Storage Applications

Internet Protocol (IP) storage is a reality. The wide acceptance of networked attached storage (NAS) demonstrates the viability of storage delivery over IP networks. Today, new protocols are enabling next-generation storage applications. For example, iFCP (Internet Fibre Channel protocol) and FCLP (Fibre Channel over LP) protocols are allowing companies with established Fibre Channel infrastructures to build bridges between their SAN islands using IP wide-area networks. These technologies have become increasingly important in today's environment as companies scramble to put reliable, affordable off-site backup and disaster recovery strategies in place. In addition, the imminent ratification of the iSCSI specification further broadens the horizon for IP storage to move beyond file-service and wide-area-networking applications into the core of networked storage systems with block-level storage services over

These new IP storage capabilities present if managers with a compelling business proposition: Leverage the TCP/IP skill sets of the current employee base and take advantage of the significant cost savings offered by mass-produced IP networking gear. Not only is there an order of magnitude difference in capital cost between Fibre Channel and IP technologies (I recently priced 1Gbit Fibre Channel HBAs in the $800 range while 1Ghit NIC cards cost less than $80 per unit) but the availability of management tools for IP networks far outstrip that of Fibre Channel, further enhancing the TCO of IP based storage networks.

Performance Drivers

Performance is a crucial issue for storage applications. Storage systems performance depends primarily on two system parameters: response time and throughput. Response time is the time between a host's data request and arrival of the first response bit to that host. Throughput is the steady rate of data flow between the storage system and the host.

Different applications have different performance requirements. For example, a database application performing numerous read/writes of small data blocks is more sensitive to response time, while less sensitive to throughput. Since the amount of data per request is relatively small, the server's maximum performance is limited by the amount of time it spends waiting for the data to return from the storage system.

Conversely, video-on-demand applications are less sensitive to response time, but more sensitive to throughput performance. For video-on-demand, the user may not care if it takes a second or two for the video to begin playing. However, once the video begins to play, any degradation of throughput would make the video jump and skip--severely reducing the viewers quality of experience.

Since requirements for different applications vary so widely, it's impossible to give a one-size-fits-all rule of thumb to help IT managers decide what route to take. Some applications have very stringent system performance requirements, while others can afford to use lower performance--and significantly lower cost systems.

Real-World IP Networks

IP networks can be nasty environments for storage applications. The structure of IP networks impact response time and throughput capabilities. IP is a best-effort delivery protocol; packets may be dropped, rerouted, reordered, duplicated, delayed, and even corrupted by bit errors. Other network effects, such as router congestion and bandwidth restriction further impact performance.

All of these network characteristics affect the response time and throughput of the system. Response time is dominated by latency, or delay in the network caused by distance (the speed of light), queuing delays, and dropped packets. The further apart the host and the server and the greater number of hops in the network, the more latency is introduced into the system, and the slower the system's response time.

Real-world network conditions also affect throughput. Throughput depends greatly on the transport protocol (layer 4) chosen to control traffic flow. The transmission control protocol (TCP) is a connection-oriented transport protocol. It establishes a connection dialogue between devices to insure that datagrams are received before continuing to transmit additional data. The dialogue consists of data packets flowing in one direction, and acknowledgement packets flowing back. As the TCP stack determines that the connection is good, it ratchets up the number of packets it sends before expecting an acknowledgement--thus increasing the throughput capability of the link. However, when the IP network drops packets, due to router congestion or other conditions, the receiving system doesn't send an acknowledgment. A time-out condition triggers the sender to retransmit the packet, the TCP stack slows down the rate at which it sends packets between acknowledgments and throughput falls off dramatically.