A Contrarian Viewpoint on SSDs

Peter, other articles in my series from the Flash Memory Summit did talk about wear leveling.

interesting article but it fails to mention that many of the manufacturers have already worked past the reliability shortcomings of SSD by forcing the writes, and erase to different spots on the SSD significantly reducing wear. First generation type SSD in the SAN industry did nothing to address this. Equallogic, EMC, Fusion I/O, as well as many others have their algorithms write and erase in ways that greatly improve the SSD lifespan and reliability. Keep in mind that Xiotech had some problems with SSDs in their non ISE platform and discontinued the use. In addition they have struggled to get SSD into their highly proprietary ISE platform. I would defiantly look at how widespread and accepted the use of SSD is across the industry right now, it's fairly strong and growing fast. I would be more worried about a storage company not offering and SSD option.

Regarding primary vs secondary storage, many old core memory computers had non-volatile primary storage, but still used hard disks. It is not a matter of enough memory, but of cheap enough. Regarding I/O speed, I too see a certain consistancy over time. My desktop PC boots from HD in about the same time that my Osborne booted from floppy. Ultimately it is user expectation, not capability of the technology that sets the level of compromise, and therefore system performance.

There are some non-obvious problems and advantages of SSD, worth considering alongside the usual price/performance issues. On the plus side, SSD have a knowable failure mechanism---the firmware can predict the remaining useful lifetime based on the available-vs.-current erase count; this is a huge advantage for managing a reliable storage system. On the problem side, there's a capacity gap. If all semiconductor plants in the world were converted to make flash memory, they would provide 12% of current worldwide demand for storage. There's a very interesting talk by IBM storage expert Steve Hetzler about those issues; he argues that SSDs have a hard time finding their niche on the price/capacity vs price/performance graph, just like several other promising storage technologies had problems in the past. Check it out: on www.caiss.org/docs/DinnerSeminar/TheStorageChasm20090205.pdf Sorry for an incomplete URL but EDN rejects valid links in comments.

Some really smart write cache algorithm could handle many of the wear issues. Something which reads a small record and writes back new data to the record and then to the next sequential record, etc could wear out a FLASH sector quickly. However, stalling writes and grouping them will solve this kink of problem. If something is written over and over,it never actually goes to FLASH. It sits in the write cache. Of course, the data must be read from the write cache because the actual FLASH data is stale. Finally, a big enough energy storage device is needed for transferring from the write cache (DRAM) to the FLASH when power is removed from the FLASH disk.

This approach is something that would be included in SSD control circuitry / software. Just as "wear leveling" and other considerations coded into the controller of a SSD. The final application coders address the hardware at this level? not happening....

DM: I've heard of no one taking the approach you suggest.

By coding, I mean how the data is coded in the flash memory. The properties of flash are fast read, medium write, slow erase, so one can do something like code the data so that one can do multiple writes to one memory location before one has to erase it and start over. The data coding to achieve this has some memory overhead, but reduces the number of erase cycles.

DM, I don't see much need to argue with the figures. The truth is that SSDs consistently deliver 100-200x the IOPS performance of hard drives. In an industry that's short-stroking HDDs to double average access times (2x) while throwing away half the capacity, it seems self-evident to me that drives that perform 100-200x more IOPS would have obvious (IOPSvious?) inherent value. A previous blog entry discussed some metrics showing that one SSD outperforms eight HDDs yet costs about 4-5x what an HDD costs per byte. There's a clear application for performance-oriented requirements. As for clever coding, I think that died with disco in the 1980s. I don't believe you get clever coding any more, partly because you can't tune an app for the myriad systems configurations it might run on. Can it be done? Sure. If the hardware configuration's fixed. Coders in the 1950s hard coded assembly-language apps so that the timing matched the rotational speed of the drum memory. I don't see how you can get that sort of tight hardware-software bonding today when your disks might be as closely attached as a PCI Express connection or as far out as somewhere on a SAN.

I'm surprised you don't call him on these bogus figures. Like storage performance isn't much better today than a 5.25" floppy! In large scale storage systems (forget small form factor here), then the question is about flash in the storage hierarchy, in the same way one considered DRAM vs. SRAM, magnetic bubbles or CCD memories (R.I.P.), etc. What does SSD add compared to a file server with big cache RAM? How does the performance and cost of large arrays of small (e.g. 1") disks compare to SSD? How do SSD's unique performance characteristics fit into the memory hierarchy? Can clever coding minimize the funny performance characteristics (at the expense of some capacity)? Given his discussion, I am surprised he didn't mention the IBM System/36, System/38 or AS/400 with single level memory. My wife used to sell them and her customers loved them.