Build or buy: A tale of two all-flash strategies Ever since flash was first added to traditional storage arrays back in 2008, there has been vigorous debate on the best approach to storage architectures for all-flash systems. There are a number of opposing camps, each with their own perception of the right way forward. Some vendors chose to use existing hardware and amend it …
I think the distinction is between really high performance (low latency) and really, really high performance. Adapted systems can do really high, ground up systems can do really, really high. Something like Pure or IBM FlashSystem or EMC's new DSSD can do sub 500 millisec. Something like 3PAR can do under 900 millisec or in that range. One way of thinking about it is that ground up has twice the performance of adapted systems... but another way of thinking about it is that either is going to be a huge improvement over disk systems, so close enough. Depends on the price and the performance you need. For most people straight SSD systems will probably work just fine.
"Something like Pure or IBM FlashSystem or EMC's new DSSD can do sub 500 millisec. Something like 3PAR can do under 900 millisec or in that range."
BTW HPE 3PAR can and does produce consistent sub 500 microsecond latency, take a look at the SPC-1 numbers if you need further evidence.
DSSD is in a very different class to all of the above the way it interfaces with the outside world is built specifically for reduced latency and since it offers no data services has a much shorter code path. IBM can do similar if used in bare metal form, but very few customers are willing to sacrifice data services to achieve those numbers.
IBM, Pure etc all use standard protocols and the market requires data services for those arrays which in turn have added overheads and latency as they've evolved to meet that requirement.
The built for flash argument maybe stacked up a couple of years ago and still does vs the more traditional arrays, but with the need for data services on flash solutions it no longer really holds true for anything other than the market for specialist kit like DSSD.
You are very smart, and I know you meant to say 500 microseconds and 900 microseconds. All IBM FlashSystem 900 runs sub-millisecond. My team sees Oracle RAC sustained latency of 234 microseconds on IBM's 900. That is roughly 4x faster than what we saw with NetApp EF-series. 4x is really really fast and I believe the market is as you say it is. However, to some CFOs, 4x faster is worth it, especially when FlashSystem is the EXACT SAME PRICE as what NetApp quoted us for their EF-560. SSD drives, in a box, are NEVER faster than custom flash DIMMs spread out on an FPGA. I know you understand that, but the guy writing the article may not. He does not even understand IBM has the fastest array out there, unfortunately. There was a guy on my team who made a ruckus about IBM's 900. All the other guys fought it and were like "IBM, heck no, they aren't doing relevant". Well, my other guys got pretty quiet after seeing IBM's 900 was about 4x faster than the EF-560 in the POC.
They did... that was Texas Memory's stock and trade, unique hardware. IBM decided to punt it to Micron and use their FortisFlash. Probably makes sense. Although the TMS Flash module provided better performance, over time they were not going to be cost competitive. It is still better than standard SSDs, but not really custom engineering. IBM's hardware is probably the best for AFA, but they need to work on the software stack... add VM based storage (not simple), dedupe, decent performance monitoring, redirect snaps. Probably would have been best, if IBM wanted to spend some money, to buy something like Tintri. That is a smoking software stack.
https://www.micron.com/products/nand-flash/mlc-nand/ibm-collaboration-fortis-flash
IBM Disclaimer
Minor correction AC. FortisFlash is the result of a joint collaboration between IBM and Micron that was announced when FlashSystem switched to Micron MLC flash. In concise terms since FlashSystem architects down to the chip level, engineers have very granular access for how to interface with the chips. The result was a significant increase in endurance for a standard MLC flash chip (about 9x). Flashcore is the umbrella name for the base TMS hardware/designs. It underpins all of the FlashSystem products and includes things like dedicated FPGA controllers, non-block crossbars, and patented sub chip level RAID protection. So Flashcore is how the overall system is designed and FortisFlash is specific to how FlashCore interacts with the flash chips.
Storage is bought with two logic points, bigotry and complacency. Even though IBM has the best flash array in the market, and they are #1 or #2 (PB shipped), many bigots at the customer site are asleep and complacent. They think IBM storage business 2006 instead of IBM storage business 2016.
Pure is drunk on it's success in the mid-market with VDI/VSI. They will eventually come out with a SKU that does not sell because it does not hit the right part of the market, at the right economics. This is probably the one. They will have the same frustrations that Infiniflash and DSSD are having in terms of where customers see it fits, and at what price versus all of the other cheap and deep converged options out there involving storage defined IP. Also, Pure's margins will get more compressed over the next year or two. Kind of explains the whopping 67% short interest on the stock right now, no?
Chris, great article, but how do you not mention IBM FlashSystem? You know that TMS has been building custom designs with dram and flash for 3 decades right? IBM FlashSystem is the best performing and densest box my team has ever looked at it. We made the plunge after looking at 10 competitors. Tell me a box that costs $4/GB usable that gets .235ms sustained latency running Oracle RAC. Go ahead, show it to us. No vendor could show us this. People, don't judge IBM by the 3 letters. FlashSystem blew my team away and no one has even come close.
I am not so impressed by the article. It looks subjective, like Chris has buddies at Pure. Does Chris even know what an SSD is?? Violin, IBM, Pure...are not making "custom SSD". They are making custom flash modules with nand flash DIMMs. Understand yet? An SSD, is a solid state drive, made by the usual suspsects....SanDisk, WDC, Seagate, etc...
I am impressed how Pure can create a group of 20 folks, doing stealth design within the company. When I say stealth, I mean "copy IBM's FPGA blade design". I guess IBM will sue soon, like EMC had to sue???
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
