EMC introduced Enterprise Flash Drives (EFDs) into their mid-range Clariion storage arrays last year with the promise of them being the ultimate “Tier 0″ performance drives. They provided some initial guidance on configuration, but as we see with most technologies, these recommendations have shifted somewhat over the last 12 months. It is still true that EFDs can provide tremendous performance gains over Fibre Channel and SATA drives; however, it is important to understand the configuration parameters and workloads that benefit from flash drives.

The biggest improvement with EFDs is seen from small, random, read-intensive small block size workloads. These workloads thrive on flash drives because there is no mechanical seek time and they are serviced by multiple parallel threads from an individual device. It is not uncommon to see 5000-8000 IOPS per drive in these environments (when configured with 4k or 8k block sizes). When using a larger block size though, these IO rates go down dramatically, often matched by Fibre Channel drives, due to the restrictions of the drive interface.

This behavior requires us to re-think recommendations provided by Microsoft and Oracle to format database volumes with large sector sizes (i.e. 64K). These recommendations are made because the database providers want to cache large blocks in memory to get around poor disk performance, so they read in large sectors in the hope they will have page-hits in memory. If the nature of the application is truly random, however, small disk block sizes on flash drives can provide better performance. Additionally, because of their random read-heavy nature, database drives on EFDs won’t benefit much from having array controller read or write caching enabled.

The next consideration is whether or not to put database transaction log volumes on EFDs. The conventional wisdom when EFDs were released was to put TLog LUNs on RAID10 FC drives because logging activity is primarily sequential writes. The original guidance for EFDs was to turn off SP write caching in order to allow the write cache to be better leveraged for FLARE LUNs. What we have learned in the wild, though, is that for write intensive LUNs (like transaction log volumes) enabling write cache has a positive effect on write-intensive EFD LUNs. We see this because the Clariion cache strategy is to coalesce consecutive writes into complete stripes using the MR3 algorithm. These complete stripes are then flushed to the “disks” in the RAID group. These large full stripe writes work very well with the EFD buffered write mechanism.

So where wouldn’t you use EFD drives? Flash drives will perform well under most circumstances and will provide excellent performance as well; however, FC—and even sometimes SATA drives—can provide comparable performance in single threaded large block sequential operations like backup and media. Flash drives can hold par in these types of operations, but the cost/IOP and cost/GB does not make sense.

With the release and propagation of FLARE 30 FAST EFD, drives are going to provide a critical component to storage tiering both in Tier0 and in FLASHCache. This should simplify decisions regarding what should be on EFD “spindles” and what should be on FC and SATA.