Fibre Channel SANs vs. iSCSI.

Much has been said about the emergence (and supposed future dominance) of Internet Protocol (IP)-based Storage Area Networks (iSCSI). The hype regarding iSCSI in the storage industry has at times reached fever pitch, with articles in trade publications calling for the "Death of Fibre Channel" as we know it. Although there are certain environments where IP-based SANs may be advantageous, there are also disadvantages. With the majority of the world's SANs (including large banks, credit card companies, large enterprises, etc.) using the Fibre Channel protocol because of its performance and robust reliability, it is hard to call for the end of Fibre Channel anytime soon.

When comparing Fibre Channel and iSCSI, there are many questions about cost, speed, performance, interoperability and reliability that have not been answered. In this column (the first of two), we will cover speed, performance (specifically CPU utilization) and reliability. In the second column, we will cover cost differences, TOEs vs. HBAs and usability.

In 2004, the iSCSI market reached a milestone on the road to end-user adoption--standard ratification. After more than three years of development, only a handful of companies deployed SANs based on iSCSI. However, these were not production SANs in the usual sense, but were mostly smaller, test scenarios.

Differences Between FC SANs and iSCSI: Speed

Fibre Channel SANs currently run at 2Gbps with 4Gbps and 10Gb/sec (equivalent to 12Gb/sec) ratified by the industry and coming soon, while iSCSI SANs run at 1Gbps. There is a Fibre Channel roadmap to 8Gbps. Performance differences have just as much to do with silicon, host adapters, accelerators, and software stacks as they do with wire speed. Fibre Channel is handled in hardware and is capable of realizing full band-width performance, even during full-duplex operations. Any properly configured 2Gbps Fibre Channel SAN is capable of sustaining in excess of 380MB/sec.

[ILLUSTRATION OMITTED]

With Gigabit Ethernet/iSCSI, most of the processing of the data (TCP and iSCSI) is handled in software and is much slower. Most Gigabit Ethernet networks cannot utilize the capacity of the link (230+MB/sec) and most of them do not achieve or sustain half-duplex rates of 115MB/sec. This is due to the fact that the data is processed in software layers, requiring CPU utilization and extra data transfers to retrieve the data. TCP Offload Engines (TOEs) and 10-Gigabit Ethernet may help to alleviate some or most of the performance issues in the future, but at additional hardware and/or infrastructure costs.

CPU Utilization

Fibre Channel now runs virtually any I/O load through a Fibre Channel adapter with minimal impact on CPU utilization. With iSCSI and TCP, heavy I/O loads require host adapters with TCP/IP and iSCSI offload engines and accelerators, adding to cost. On the other hand, there are some low-performance applications that may run acceptably using iSCSI software drivers and inexpensive Ethernet NICs (network interface cards).

Fibre Channel was designed and architected to communicate directly at extremely high speeds. Those systems providing extended distance capability (over 20 km) use larger flow control buffers at increased cost. In addition to those systems with added buffering, long haul Fibre Channel solutions may utilize routers to push Fibre Channel across widely available networks, such as ATM, T3, and SONET/SDH. Additional solutions include FCIP and iFCP, which allow Fibre Channel to be transported across a TCP/IP infrastructure.

iSCSI and IP, however, are natively designed to run over long-haul and short-haul networks using standard IP interconnections. Such interconnections can be specifically tuned for the special requirements of storage applications, without incurring additional hardware costs.

Reliability

Reliability differences between Fibre Channel and iSCSI may drive users away from iSCSI. By default, Fibre Channel data is completely protected--it's designed to be 100% reliable. However, many Fibre Channel SANS are run on Class-3 connections, which are not acknowledged, thus leaving possibilities for error recovery to occur only at the SCSI ULP layer. Recovery can take one to two minutes while the SAN is potentially inoperable for one host or a number of hosts.

By comparison, iSCSI--or IP--is designed for expected failures and quick recovery times by utilizing TCP for acknowledgement. This is a fairly widespread method for data delivery used in many networking protocols (HTTP, FTP, NFS, SMB, etc.) However, in the iSCSI standard, cyclic redundancy checking (CRC), a data-protection method, is optional, leaving a risk of potential data corruption. Many vendors are choosing not to provide iSCSI CRC digest checking, so ultimate integrity is at risk.

Conclusions

iSCSI may be coming, but it is by no means going to be the death of Fibre Channel anytime soon. Those customers who have requirements for high storage performance, mission critical reliability and large complex connectivity (hospitals, financial institutions, brokerage firms, etc.) will undoubtedly continue to choose Fibre Channel, which has demonstrated these characteristics for years.

Part two of this article will appear in the April edition of CTR, discussing TOEs vs. HBAs, costs and usability.

Steve Klotz is founder, Technical Resources director, Finisar Medusa Labs Member, Fibre Channel Industry Association (San Francisco, CA)

www.fcia.org