This article can also be found in the Premium Editorial Download "Storage magazine: Storage products of the year 2003."
Download it now to read this article plus other related content.
|Understanding iSCSI host connections|
Because iSCSI sits on top of TCP/IP, any standard network interface card (NIC) could be used to connect a host to an IP storage area network (SAN). However, special-purpose iSCSI cards are becoming available that can reduce CPU utilization and potentially provide greater levels of performance--albeit at a higher price--than a general purpose NIC (See "Network interface card options," below).
With a conventional NIC, all functionality above the GigE level is provided in software, including all SCSI, iSCSI and TCP/IP functions. The impact on the host depends on the CPU load and the I/O characteristics of the application. A lightly loaded CPU may have sufficient headroom to provide good iSCSI performance. A moderately busy CPU may even be able to support a light to moderate iSCSI load.
There are two alternatives for offloading the iSCSI processing overload: TOE NICs and iSCSI adapters. A TCP/IP Offload Engine (TOE) NIC offloads all TCP/IP operations. This type of interface isn't specific to iSCSI at all, and could be used as a replacement for a general-purpose NIC card in any application deemed appropriate. When using a TOE NIC for storage, SCSI and iSCSI processing would still be performed by the CPU.
An iSCSI adapter is directly analogous to a Fibre Channel host bus adapter. In this case, all iSCSI and TCP/IP operations are offloaded to the interface card. All of these options are currently available. Make sure you don't draw conclusions about performance without testing specific applications on specific platforms.
Network interface card options:
So, is 2004 the year to build an IP storage area network (SAN)? As with most storage adoption questions, the answer lies in understanding the current capabilities and limitations of the technology, and having a solid understanding of your storage networking needs. If your organization hasn't done so yet, it should begin asking--and answering--questions on how best to deploy an IP SAN. Among the areas that need to be evaluated are your application storage characteristics, network infrastructure and security requirements.
iSCSI design considerations
First and foremost, application considerations will impact design. Some questions to answer are: Is the application transaction-oriented or throughput-oriented? What are the I/O performance requirements of the application? What platforms do the application support? Ideal candidates for iSCSI today are low- to moderate-level I/O applications that run on Windows or Linux platforms. This typically includes applications such as Microsoft Exchange and many SQL Server-based applications.
The leading network consideration relates to the bandwidth required. An important decision to make is the method of host connection. Can a software initiator with a standard gigabit network interface cards (NICs) be used, or is a special-purpose adapter required? (See "Understanding iSCSI host connections" on this page.) There's no one right answer to this question. The major benefit of iSCSI host bus adapters (HBAs) and TCP/IP offload engine (TOE) NICs is the decrease in CPU utilization relative to a standard NIC, with performance improvement as a secondary benefit. If there's enough CPU headroom available and the application I/O requirements aren't extraordinary, it's worth considering foregoing a special-purpose card and using a standard NIC, given the potential cost savings.
Another network consideration is the segregation of traffic. It's generally recommended that iSCSI activity be isolated from general network traffic. For example, this means dedicated host connections and separate virtual LANs. Additional bandwidth considerations may be based on performance optimizations for routing storage traffic.
This was first published in December 2003