Early examples of flash conversion among incumbents were not impressive. The EMC VNX 5500-F and HP Lefthand P4900 were so hamstrung by their underlying architectures that few storage analysts viewed them as serious alternatives to the shiny new systems from startups.
These early attempts were either IOP-limited or overly expensive (or both), in part because the traditional RAID logic their ancestor systems used for data protection on disk drives created a lot of write amplification. This amplification forced the designers to exclusively use expensive SLC SSDs.
However, recent product announcements by Dell and HP, and a lengthy conversation at Dell's Enterprise Forum, have me thinking that some disk-based systems make the transition well enough to make them contenders with purpose-built arrays.
One reason is the new Dell/Compellent and HP/3Par systems are based on architectures that are only about 10 years old. This is important, because it means they never used traditional RAID to protect data spindle by spindle but instead used "chunklet RAID," which is much better suited to both the requirements of flash and data movement.
However, my first reaction to Dell Compellent's flash-optimized system announcement was, "Really? Isn't the SLC era over?" This was an opinion echoed by my fellow analysts. That's because, in general, the flash controller folks have made MLC a much better storage medium that it was a few years ago, even though the smaller-geometry MLC chips, other than being cheaper per gigabyte, are actually worse.
Yet on closer inspection, the SLC/MLC hybrid architecture that Dell is pitching makes a lot of sense. In fact I've proposed a similar architecture for an enterprise SSD of the future. Incoming data is written to the SLC SSDs, which are faster on writes and have much greater write endurance. When the data in the SLC layer exceeds some high-water mark, the data progression process is triggered, moving the cooler chunks down to MLC.
Because Compellent's data progression was the first implementation of sub-LUN tiering, Dell Compellent is fond of the process. I generally prefer caching, but the progression process seems a better fit here than in most other cases.
Any really hot data, like the inventory record for popular items during the Christmas rush, will stay in SLC and absorb all the write traffic. This lets Dell use MLC SSDs that are "read optimized," which mostly means that they have lower levels of overprovisioning than typical enterprise MLC drives, thus keeping costs down. In fact, Dell claims the flash-optimized system has a cost-per-gigabyte price comparable to 15K RPM disk systems. By contrast, startups use data reduction to master this trick.
The key advantage the incumbents have is the wealth of storage management available for a platform that's been in a thousand customer sites for years. Not only do the latest flash-optimized Dell Compellents and 3Pars have all the snapshots, replication and host integration features their disk-based systems do, they also have the kind of third-party storage resource management software support that Pure Storage, Whiptail and Violin Memory can only dream of.
If we compare systems announced in the past month or so, Pure Storage's latest update has boosted its performance to 400,000 IOPS.
By comparison, the latest system from Dell Compellent tops out at 300,000 IOPS, while HP's four-controller 3Par 7450 hits 500,000, with latency under 1ms. Pure Storage manages all that performance with deduplication while the mainstream players are still storing straight data, but I'm adding the latest Dell Compellent and 3Par systems to my all-solid-state options list. Perhaps you should, too.
Disclaimer: Dell and HP have both been clients of DeepStorage, LLC over the years, as have some of the startups.