SSD/Hard Disk Hybrids Bridge Storage Divide

Solid state is colonizing every corner of the storage ecosystem, from internal server caches to standalone SSD arrays. Heretofore, solid state has been treated as a distinct storage layer. However, a hybrid approach blurs the lines between electronics and magnetics by closely pairing solid-state and mechanical disk storage to achieve optimal price/performance balance.

The goal of dramatically expanding access to solid state's performance-enhancing benefits without breaking the bank, while simultaneously solving some of the biggest storage problems in virtual servers, is the dream of Enmotus, a stealthy startup that opened its kimono at last fall's Server Design Summit by demonstrating what it calls MicroTiering technology. As CEO and co-founder Andy Mills puts it, the company's "whole focus is an affordable tiered hybrid storage, at a lower price point than Fusion-io." Translated: solid-state cache-like performance at the price point of a commodity hard drive.

The trade-off between performance and capacity is a fundamental principle of digital storage. Regardless of the physical means of storing bits, system designers can have speed or size, but not both. It's the tenet undergirding the canonical multilevel cache architecture used in all modern processors, with a three (or more)-deep hierarchy for data access where fetches from each successive level take longer.

The same phenomena is now manifest in the realm of persistent storage as flash-based solid-state devices have finally scaled to the point where they're feasible replacements for magnetic disks in many situations. Here the yin and yang between performance and capacity is particularly striking. The sweet spot of for SSDs is currently 256 Gbytes versus 2 Tbytes for SATA hard drives (8 times more). In return, SSDs provide two to three orders of magnitude higher IOPs and about a hundred-fold reduction in write latency.

But engineers hate trade-offs and constantly seek design optimizations to provide the best of both worlds, at least under normal circumstances.

Enter hybrid storage, where Mills claims MicroTiering delivers nearly identical performance--within 10% on real-world applications that haven't been specifically tuned for local solid state caching--to a caching-plus-HDD or all-SSD implementation.

MicroTiering gets its name from the fact that it gloms storage volumes into a cohesive union of SSD and HDD that to the outside world look like a single device. But unlike server-side caching products made famous by Fusion-io that simply front-end HDDs with a veneer of flash, Mills emphasizes that in MicroTiers the SSD acts as primary storage. The MicroTiering software automatically migrates "hot," active data to the SSD tier and "cold," inactive data to the HDD. The scheme sounds a lot like Apple's Fusion Drive, but when offered this comparison, Mills hastened to point out that MicroTiers work at the block layer, while Fusion operates on files.

Conventional wisdom, perhaps, but as my colleague Howard Marks points out, not true. Fusion drives do indeed operate below the file system as detailed by these tests showing that when reading just the first megabyte from 300 100-Mbyte files, each came out of the SSD while subsequent reads for the remainder of a file came from disk. Still, Fusion is a bundled device that only works on a select set of proprietary systems. In contrast, Enmotus builds MicroTiering into a an ACHI-compliant PCIe storage adapter that works on any server and with virtually any SAS or SATA device, much like a conventional RAID HBA.

Next page:Inside MicroTieringBut building MicroTiering software into a hardware adapter doesn't just offer maximal flexibility in choice and mix of drives; it also enables the tiering action to happen outside and independent of the guest OS or hypervisor. As Mill points out, this means MicroTiered storage, much like a RAID stripe, is presented to the host as a standard disk volume that can be carved up and spread across all the guest VMs. But unlike RAIDs, MicroTiers don't spread blocks equally among available devices, but intelligently place active data on SSD.

And in contrast to server-based caching, MicroTiers not only aggregate solid-state and disk storage, treating SSDs as a persistent device, but also don't require any system-level software to manage data movement between the two.

Mills claims that operating below the hypervisor mitigates the infamous I/O blender effect, in which sequential I/O streams from guest OSs, when merged and serviced by the hypervisor, turn into a mishmash of random I/O to the physical storage devices. Mills says the MicroTier will always serve the data with the highest activity out of flash

Hybrid storage also simplifies design and deployment of VM environments using live migration because, unlike server-side caches, the guests VMs need not be specially configured to take advantage of flash storage, while the self-tuning MicroTier means those workloads with the greatest I/O demands will automatically get a greater share of the flash tier.

Enmotus' hybrid adapter uses RAID-like redundancy for both the SSD and HDD layers. MicroTier storage pools are fully virtualized so that new devices are automatically incorporated into the pool, a feature Mill notes makes it easy to increase the SSD/HDD ratio should for new applications with a larger memory footprint. Its software also detects failing or end-of-life SSDs, automatically moving existing data to other devices and reincorporating replacement drives into the SSD pool.

Solid-state/disk hybrids promise the best of both worlds, speed and size and after years of false starts, as storage consultant Greg Schulz of StorageIO puts it, "the teething problems are being solved." Products like the Seagate Momentus XT and Apple Fusion Drive demonstrate the feasibility of a commercially successful hybrid product in the consumer market, but newcomers like Enmotus hope to do the same for enterprise storage.

As Mills put it in a presentation to last year's Flash Memory Summit, "Ease of use and transparent integration of SSDs are more important than raw IOPs for the next two to three." Translated: The 80/20 rule will apply to storage. A product that delivers most of the performance improvements of an all SSD or solid-state caching system at a fraction of the cost and implementation complexity should be a winner. The thesis will get a test once Enmotus starts actually selling products later this year.