One of the hottest topics in the Storage Industry is the suitability of iSCSI for traditional storage applications. iSCSI is the encapsulation of SCSI commands inside TCP/IP. Some storage experts question whether iSCSI adds significant latency to storage access because of TCP/IP processing. Testing of the Cisco SN 5420, the industry's first iSCSI networking product, indicates that TCP/IP processing adds minimal additional latency to overall storage access. To measure the latency, Cisco developed a test environment that measured the number of read I/O operations per second (IOPS) using various block sizes that are widely used by database applications. Comparison tests were run with iSCSI attached and Fibre Channel attached storage. The test results show the SN 5420, which implements the iSCSI protocol, is suitable for traditional storage applications.
The term latency is used in numerous ways when describing computer system delays. In this document, latency refers to the time required to request a block of data from the disk drive and transfer it back to the application buffer in the host or server.
Latency in read operations to disk can be hidden by read-ahead, operating system (OS) buffering and the read cache RAM buffers located on the disk drives. The test code was specifically written to diminish these effects. To minimize the effect of read-ahead, both at an OS and disk drive level, the test program used a pseudo-random number generator to select a new seek position before each read operation. To remove any OS buffering, all tests were conducted on a Linux "raw" device, which bypasses the OS buffers. To help ensure that the influence of the cache on the disk drive was kept to a minimum, a 2 GB section of the drive was used that diminished the effect of any read caching done by the 4 MB of RAM cache on the drive.
Figure 1 iSCSI Test Setup
Fibre Channel Test Setup
The tests were run on a 933 MHz single-processor Intel Pentium III-based system. The PC motherboard was an Intel STL2 server board, which has two 64 bit/66 MHz PCI card slots and four 32 bit/33 MHz slots. The Fibre Channel tests used a Qlogic 2100 HBA and the iSCSI tests used a 3COM 3C985B-SX Gigabit Ethernet Network Interface Card (NIC). Both the Qlogic HBA and the 3COM Gigabit Ethernet NIC are based on similar serial gigabit technology with comparable raw throughput rates. Only one card at a time was plugged into the motherboard and the card was plugged into a 64 bit/66 MHz PCI slot to ensure maximum performance.
Tests were run on three block sizes: 1 KB, 8 KB, and 64 KB. These sizes were selected because they are the most common block sizes used by database applications. On each test, 100,000 seek and read operations were performed. The results have the average of 10 runs at each block size.
The operating system used for the tests was RedHat Linux version 7.1 using the unmodified RedHat supplied 2.4.2 Linux kernel. The 2.4.2 kernel supported all devices and protocols except the iSCSI protocol. The Cisco iSCSI driver, version 1.8.7, was used for the iSCSI tests. This driver was loaded as a module, so the Linux kernel did not have to be modified. The driver and the test program were the only software loaded on the system that was not part of the RedHat 7.1 distribution. No RedHat patches were applied to the system.
A single Seagate 15,000 RPM Fibre Channel disk drive was used for all the tests. A Vixel 2100 Fibre Channel hub was used between the Cisco SN 5420 and the Fibre Channel enclosure to convert the multimode fiber interface on the Cisco SN 5420 to the copper interface on the fibre-channel disk enclosure. Figure 1 shows the iSCSI setup with the Cisco SN 5420, and Figure 2 shows the Fibre Channel test setup.
Chart 1 shows the average latency for each block size tested. The latency differences between the direct Fibre Channel-attached disk tests and the iSCSI tests using the Cisco SN 5420 are minimal. In all test cases, either technology could be used for most applications. In future iSCSI initiator and target products, Cisco expects that the latency will be reduced even further and any latency differences between Fibre Channel and iSCSI-attached storage will be insignificant.
Chart 1 shows iSCSI latency is a small portion of the overall latency.
Databases are one of the most latency-sensitive applications, and the primary measurement for disk performance of a database is the number of I/O per second that the disk system can sustain. Chart 2 shows the average number of I/O per second for both the Fibre Channel tests and the Cisco SN 5420 or iSCSI tests. Like the latency measurements, this chart shows that there is only a minimal reduction in I/O per second for each of the block sizes between the Cisco SN 5420 and the Fibre Channel-connected storage. The chart shows that the Cisco SN 5420 is well suited for database operations.
Chart 2 shows iSCSI is valid for consolidating storage and business continuance.
Cisco measured the latency of both iSCSI attached and Fibre Channel attached storage and found that the SN 5420 is suitable for traditional storage applications. The Cisco SN 5420 provides an extension of the existing Fibre Channel SAN using the iSCSI protocol. This extension allows for storage consolidation, remote data access, and the ability to do remote backup. For storage consolidation, the SN 5420 cost-effectively extends the benefits of a Fibre Channel SAN to servers throughout your enterprise using TCP/IP. All of these features of the SN 5420 allow for a lower cost of storage management without a significant impact to system performance or latency.