Storage speed demon

Is faster really better?  Certainly not in regard to some aspects of our private lives, but it almost always is when it comes to computing environments.

A historically useful differentiator among storage hardware vendors has been how quickly their technology gets data on and off a disk.  When these read and write operations occur on disk subsystems – disk arrays and JBODs (http://www.nwfusion.com/links/Encyclopedia/J/724.html) – the speed at which the data moves through the machine has relatively little to do with the horsepower of the on-board processors. But it has a great deal to do with the machine’s basic architecture.

With network-attached storage (NAS) devices, RAID subsystems and JBODs, there is a lot of speed at the backend – where the disks are – and at the front end where the device connects to the storage-area network (SAN), LAN or (decreasingly these days) to a computer.

Today, the majority of disks shipped with high-end RAID systems spin at either 10,000 or 15,000 RPMs. RAID controllers and bus protocols – both SCSI and Fibre Channel – manipulate data more efficiently each year.  To an increasing degree, RAID devices are located on SANs, which (assuming they are running on Fibre Channel) move data at either 1- or 2-G bit/sec.

Despite all that horsepower in and behind the RAID controller, plus the high speed of the SANs, data still moves through an array at an appreciably slower rate than you might think.

The same thing happens with JBODs.  These are much simpler devices, but nonetheless are a substantial investment for an IT group.  JBODs too are often found on a SAN, and frequently have high-speed disks inside.  And like RAID systems, JBODs cannot seem to move data onto the SAN fast enough to feed the data hungry applications that run on our new, faster CPUs.

The problem is that shared-bus architectures of these devices have been unable to scale sufficiently to deal with the rush of data that hits subsystem controllers.

As a result, the buses are swamped and are a chokepoint for throughput. Within the last week, EMC and Vixel have announced new technologies for getting data through the machine and onto the SAN faster.

EMC has redesigned the guts of its high-end Symmetrix arrays, swapping its decade-old bus architecture for what it calls a “Direct Matrix Architecture” that offers up to 128 direct data paths.  This, plus a very large cache will provide 64G bytes of bandwidth through the machine.  With up to 64 Fibre Channel or ESCON ports, the new DMX series should not only get data through the box, but also onto the SAN faster as well.

For those of you more interested in JBODs and NAS, Vixel’s embedded switch technology provides competitive speed at what is likely to be a very competitive price.  Its embedded switches have been in use for several months now by a number of major storage vendors, including HP, NEC and LSI. It is at that market that Vixel is also aiming its newly announced 20-port InSpeed 320 system on a chip (SOC). The SOC is also available as a blade for existing devices, and as an independent unit.

You are likely to be impressed by InSpeed 320 because it provides a series of direct connections between the disks and the devices requesting the data.  As a result, each request holds the data path for as long as is necessary (rather than contending for shared bandwidth over a bus) which speeds up the way data is moved through the system.

System builders will certainly like that, but they are even more likely to appreciate both the scalability of the switched architecture and its inherent ability to improve data path monitoring and management (I/Os can be tracked discretely across the individual data path rather than across the shared bus).  These are the sorts of things that lower what it costs a vendor to build a system – and hopefully, what it costs the end-user to buy one as well.

Obviously, performance isn’t everything: manageability, availability, functionality and cost are just as important.  But other things being equal, when it comes to accessing data, faster really is much better.