Q&A with Dal Allan: Ultra3 SCSI and Fibre Channel
Dal Allan is president of ENDL Consulting, in Saratoga, CA. He has advocated storage interface standards for more than 20 years and has been involved with the development of SCSI and Fibre Channel since their inceptions.
InfoStor: Do we need Ultra3 SCSI?
Allan: That`s the most politically charged questions there is these days! I personally think it`s coming too soon after Ultra2. It takes time for the market to absorb new technology and you need a good run to recover the investment. Too many choices in a short time tends to disrupt the market, and that could actually benefit Fibre Channel. Ultra3 is a new step with the double-edged clocking, and many companies may decide to do Fibre Channel rather than parallel SCSI after the short time it will have been for Ultra2.
InfoStor: Within the time frame that Ultra3 SCSI devices will be available--probably next year--where will Fibre Channel be from a performance standpoint?
Allan: As an external interface for point-to-point use, you`ll be able to get double-speed Fibre Channel next year, but the loop--at least internally--will remain the same. The maximum benefit for Fibre Channel over the next few years will come from driving down the cost, not by boosting the speed. There is little to be gained with higher speeds for a while, because few processors are capable of taking advantage of what we`ve already got. The systems that do need it will be able to run at double speed.
InfoStor: Do you think there will be a time period when SCSI will be the faster interface?
Allan: Not unless SCSI figures out a way to run full duplex. Remember, your 100MBps Fibre Channel cable is really more than 106MBps In and also Out, so you have a nominal rate of 212.5MBps. I`m using nominal rates because that`s the fairest comparison with SCSI. SCSI is half duplex, Fibre Channel is full duplex.
The overhead is very different between the two. On Fibre Channel it`s less than the 6.25% that was originally allocated. The next big performance gain for loop will be FC-AL-3 with multiple concurrent virtual circuits, or spatial reuse. In real terms, you can figure on a 60% boost from full duplex, and about 4X from spatial reuse, so the bandwidth will be more than 500MBps.
Disk subsystems won`t benefit from full duplex because the disks will be single direction, but disk drives can do better than 4X on spatial reuse if you put large numbers of them on the loop. It`s hard to imagine a disk subsystem getting up to 500MBps on a single loop, but it would be possible.
Although data transfer rates have been boosted regularly on SCSI, the overhead has remained constant. It`s an unfortunate circumstance that only a percentage of the improved speeds can be achieved. Packetizing parallel SCSI is one answer, because it provides a dramatic reduction in overhead and comes with a CRC bonus. Packetizing can be applied to any speed physical interface, and CRC is good for OEMs, despite the downside that packetizing requires new drivers.
InfoStor: What`s the overall interface situation?
Allan: Interfaces are no longer a bottleneck. We`re way out in front of the capability of systems to process data at the rates available now. Until operating system vendors redo their I/O structure, we`re not going to get the kind of I/Os per second at the application level that the hardware is capable of delivering. Boosting the raw rate is challenging, but it`s actually an easy out compared to reducing operating system overhead.