I can understand the difficulty. Here’s another data point…Intel Research did an analysis of Flash vs. HDD for synchronous IO in transactional database applications (SIGMOG’09). The results show that the Intel X25 SSD was only twice as fast as a 10K HDD on random writes…so the 351 IOPS number that Microsoft measured in not that far off. The SSD was about 1/5th the speed of the HDD in small, sequential IOPS.

See Fig. 3(d) and 3(b)

http://www.pittsburgh.intel-research.net/~chensm/talks/Flashlogging-sigmod09.pdf

As you go through the paper, note that the “Ideal” performance scenario is represented by a single HDD with write cache enabled, and that the X25 SSD never manages to exceed about 70% of the HDD w/WCE performance (see Fig. 17, compare “Ideal” vs. SSD).

It’s no secret that NAND Flash Erase/Program (Write) cycles are 100x slower than reads, but most people mistakenly assume that the DRAM write caches on modern SSDs solve the problem. Unfortunately, the most performance-sensitive enterprise applications do a lot of synchronous I/O. When Flash SSD write caches fill up, the whole system slows down to the speed at which writes are completed to Flash. Another feature of synchronous I/O workloads is that they don’t generate deep queues at the device. Without deep IO queues, the multichannel paralellism of modern SSDs can’t be leveraged.

> MSUK research team (my link above) used after-the-fact analysis of application traces
> to decide in advance exactly what data could benefit from placement on SSD vs. HDD
Sorry but it’s difficult to take seriously the analysis of an SSD with 351 write IOPS (your link, Table 4).

I think the “hot data” end is the problem, not the “cold data” end. $100/TB is plenty cheap, what we need is something like $0.25/IOP (SPC-1C/E) and $5/GB from an “enterprise class” SSD cache tier before it’ll fly.

With E-class SSD at $2/SPC-1 IOP and ~$50-$150/GByte today, we’re a long way off…

I was specifically referring to the “technical costs” in terms of adding the SSD hardware and software logic for making automated decisions about what goes in which tier. This was meant in the context of looking at the potential ROI, or more specifically — at what point does a two-tier storage architecture (i.e. SSDCheap HDD) actually have the potential to reduce storage costs compared to a single HDD-only tier.

>>”Once you own operating system/file system/volume/drive layer, you’ve got all the usage data one needs to make intelligent placement decisions…”

Sure you do, but the MSUK research team (my link above) used after-the-fact analysis of application traces to decide in advance exactly what data could benefit from placement on SSD vs. HDD. In other words, they presume the intelligence to make 100% perfectly accurate placement decisisions is just “there” irrespective of how or where it’s done.

>>…and all technical capabilities to do it transparently(with zero extra management costs).

They also presume that the incremental cost of this intelligence is zero.

So…even with a hypothetical perfect intelligence that hypothetically comes for free, there was still no ROI to be found in switching from single HDD tier to two-tiered SSD-HDD architecture .

Without an underlying business case (ROI) for two-tiered architecture, how can there be a business case for the intelligence to manage it?

In my first comment I have intentionally said “he who owns the whole stack”. Once you own operating system/file system/volume/drive layer, you’ve got all the usage data one needs to make intelligent placement decisions and all technical capabilities to do it transparently(with zero extra management costs).

ZFS L2ARC made first step by introducing extra level of read cache, and I believe intelligent (meta)data placement is next on their agenda. And they don’t make money in the “access gap”, they make it on an intelligent system delivering more bang for less bucks .

Agreed. I laughed when AutoRAID came out – it was obvious that there was no economic benefit. Haven’t run the numbers but it feels like $100/TB disk might put us at the tipping point.

If not $100, then maybe $50/TB. It is coming – soon – and at some point it works.

Robin

MS UK research group assumed “perfect” intelligence and yet found the cache tier is a money-loser…

http://research.microsoft.com/en-us/um/people/antr/ms/ssd.pdf

25 or so years of research at places like HP Labs has never resulted in a commercially successful product. When the geniuses at Carnegie Mellon’s renowned Parallel Data Lab spent 4yrs building and testing automated “tiered learning architectures”, they started from the same premise — “of course, this is obvious”. They were surprised that they couldn’t make them work, even after trying a slew of sophisticated AI and genetic algorithms.

http://www.pdl.cmu.edu/PDL-FTP/SelfStar/CMU-PDL-04-109_abs.shtml

Finally, upon examining the “architectural economics” of a discrete caching tier, it turns out that you get tremendously more bang for the buck either by (a) increasing system DRAM or (b) using more spindles, or (c) both, so its almost impossible for companies to make money in the access gap between DRAM and disk.

http://www.idema.org/_smartsite/modules/local/data_file/show_file.php?cmd=download&data_file_id=2007

Joe.

Also take a look at grid iron http://www.gridironsystems.com/ …

As for tiered storage never really taking off, I think it’s because there was not much difference between the tiers, talking only ~3-4x performance boost from SATA to 15k FC. And with SSD, modern enterprise arrays simply cannot handle them with remotely the level of density that they can 15k drives.

My company’s last storage refresh we went to a 100% SATA solution, with big caches in the front end, certainly not as fast as these specialized products but it’s good to know we’re headed in the right direction

Joe.

Not sure at all if at the end run you will save money (not to
speak of hassle) in the long run with such a solution.

Robin