More on SSDs as SSTs

12:45 PM -- My friend and fellow Byte and Switch blogger George Crump has certainly been making the case for flash SSDs over the past few weeks, and with the recent announcements from IBM, Sun, HP, and Pillar, flash technology has officially taken the record from data de-duplication as the 21st century's fasted technology to reach universal storage vendor support.

The latest trend is to put the flash chips right on the server's PCI-E bus, eliminating a few microseconds of latency, the cost of all that pesky storage networking kit and the SAS or SATA 3 Mbit/s bottleneck. All of this leads to truly blinding performance at a price of around $50 per Gbyte, which is around what you paid for the Clariion CX500 full of 36-GB, 15,000-rpm drives.

Rather than be topped by Texas Memory, Fusion-io has announced the ioDrive Duo -- taking back the crown of fastest drive alive with a claimed 167,000 write IOPS and 1.5-Gbit/s transfer rate. They even used a 4-KB block size for the IOPS rating, as Oracle and Exchange 2007 do in the real world, rather than the 512 byte I/Os those trying to fudge the benchmark usually do.

They're also pushing the envelope on capacity with a 640-GB model out next month and a 1.28-TB whopper promised for later in the year.

Of course, flash SSDs are just the latest answer to better disk performance. Back in the 1980s, I built a prototype digital special effects system for Orion Pictures that used head-per-track disks to hold a whopping three frames of 4Kx4K scanned motion picture film. Texas Memory has been selling DRAM-based SSDs for years, and those with deep pockets and little patience have found them to be an effective way to make things work faster.While there will always be applications that need better performance, I wonder if the PCI-E Flash drive is the right solution for the kind of general purpose problem George described in his last blog post. By moving from SAN disk to DAS flash, his proverbial DBA could speed up the database substantially.

But two years as InformationWeek's backup boy make me think about what he's giving up. That Oracle database server, probably running on Solaris or Oracle's unsinkable Unbreakable Linux, was already connected to at least a midrange array like a Clariion or Xiotech Magnitude, where the storage admin was taking periodic snapshots of the database LUNs just in case the database somehow got corrupted. He was also replicating those LUNs to the corporate undisclosed location just in case one of the four horsemen of the apocalypse should happen into the data center. Finally he was running backups off the snapshots so as not to slow the application down during the backup process, and storing them for five years -- calling the pile of tapes at Iron Mountain an archive.

If the Oracle team slipped in a PDI flash card and moved the "hot" tables to it without consulting the storage team, as George suggested they might, then the database is essentially unprotected. (Please note here I'm not saying George suggested they do such a low-down, underhanded, and foolish thing. He just suggested it could happen. George is smarter than that, even if he did steal my "regular storage guy" line.)

On the other hand, I really like the idea of using a PCI flash card in a server-based storage system like Falconstor's NSS or Open-E's DSS.

Enterprise flash is about a year old. I figure it will take us another three years to really know how to use it effectively.Find out more about innovative storage. InformationWeek and Byte and Switch are hosting a virtual event on this topic on March 25. Sign up now (registration required).

Howard Marks is chief scientist at Networks Are Our Lives Inc., a Hoboken, N.J.-based consultancy where he's been beating storage network systems into submission and writing about it in computer magazines since 1987. He currently writes for InformationWeek, which is published by the same company as Byte and Switch.