RSSSolid-State Drive Industry Leaders Seem to Be Hedging Bets
By Scott M. Fulton, III | Published May 17, 2007, 5:31 PM
LOS ANGELES - A panel of four leading figures in the marketing and engineering of solid-state drives, convening toward the end of WinHEC 2007, appeared to agree that the development of solid-state disk drives and hard drives that use flash memory as intermediate caches, even as flash technologies evolve and NAND prices continue to avoid bottoming out, will never result in the replacement of traditional HDDs for any market category.
A representative of Texas Memory Systems, which produces flash-equipped components for storage, admitted that customers approach his company when they are looking to improve their systems' performance, not storage. That sentiment was validated by three other members of the panel and a moderator representing Microsoft.
At one point, Seagate's market development manager Josh Tinker invoked an analogy dealing with the evolution of the personal transportation industry, where the mode of transport migrated over the past few hundred years from the backs of animals to the internal combustion engine. The migration ahead of us there, he said, was to the electric engine. How does that come about? "The first step," Tinker said, "is hybrid."
Which would seem to say the whole industry needs to move to flash/HDD hybridization first, before moving to totally solid-state memory solutions. But none of the panelists would commit to the idea implied by the analogy.
Tinker challenged panelists and attendees alike to inject some reality into the discussion, pointing out that while the new entry point price for HDDs is about $40 for a 60 GB 2.5" notebook drive, the entry point for SSDs is about $500 for 32 GB capacity. As capacities increase, the price scales increase at different rates, with flash-supplemented device prices increasing more rapidly.
As time goes on and prices continue to fall, Tinker and the other panelists seemed to agree with BetaNews' suggestion that while both price scales decline, a formidable gap may continue to exist between SSDs/hybrid drives and conventional HDDs. But as SanDisk's representative believes, the size of the gap itself will diminish over time. Then, he believes, total cost of ownership will be evaluated, after which time, the benefits of SSD over longer periods may become self-evident.
Which brought up a question from the audience: When will there be a flip-flop, where hybrids take over as standard PC components, and the price scales swap places? Five years? Ten years? Never? Even Texas Memory Systems agrees...that's a Very Good Question. "There's this time it takes for adoption in the enterprise," he said, "and that's really held it back. But I do see this opportunity for the future as things become more developed. It just takes more time."
As the panel was convening, Microsoft confirmed a development here at WinHEC that the company is committing to producing a new bus controller interface for non-volatile memory (NVM) by the end of this year.
SSD are the drives of the future as I see it. That is if microsoft can stop fragmenting in their OS of the future. The price will drop
as they become more popular. 2 gigs are very cheap now. I believe HDD have hit there ceiling for reads and writes.SSD is the only way to evolve. Unless Nano Tech can change HDD as we know them today.
Score: 0
|Capacity is the killer though... 1TB Hard drives are out now....can you imagine the cost of equivalent capacity in Solid-State format? The only thing that is appealing to me is a small 32GB SS disk that can be used for your OS, but thats it.... for mass storage HDD's are going to stay in the lead... the cost per GB differential between the two are too great.
Score: 0
|Another thing to consider is that, afaik, while the access times of flash devices is great, the sustained throughput doesn't compare to HDDs. I don't see this mentioned anywhere in this article, but isn't it a major factor? Or am I mistaken?
Score: 0
|You're mistaken. SSDs are now rivalling the sustained read/write speeds of spinning magnetic media, at least at the consumer level. They've also surpassed the durability/reliability of consumer HDDs as well, in some areas.
Score: 0
|Sustained read/write is not the issue.
If all data was stored in a contiguous non-fragmented read only format this would be the case. But its not! Thus data can be read and placed in look ahead buffers to your hearts content, but as soon as a request is made for data in a noncontiguous region, then these buffers page fault and the data much be located and read into memory, thus negating any sustained capability. in other words, the momentum is lost and you start from scratch.
In this sense SS memory may have an advantage in that it does not rely upon the physical access mechanisms to locate the non-contiguous data. But the cost is simply not competitive to disk drive storage.
Score: 0
|my take on ss drives is that they are incredibly useful for notebooks, extending battery life, not being damaged as easily from drops and a decent read/write speed. sustained read/write is not such a big deal for me as the notebook i am looking to upgrade is using a 5400 rpm drive, and thus i will still get better speeds out of the ss drive then the current one.
Score: 0
|Personally I've kept all my apps (everything except games) on a USB flash drive that's never taken out. It's a higher end one, around 30 MB/s read. Works great and keeps the load off the HDD.
There's just some concern about the lifespan of the datablocks. Usuaully they get about 1 million rewrites...*shrug* I'll prolly end up upgrading it well before then. Only like 30 bucks anyway.
Score: 0
|I'd be more worried about a simple surge on the USB port/hub wiping critical data, as is so common with USB.
I'd have a little more faith if they made Firewire drives for the same purpose, but can't really see the advantage of doing things the way you are anyway? Why not just backup?
Score: 0
|Please don't make the mistake of equating non-volatile SS memory techniques with your USB drive, regardless of how convenient the drive its!
This is NOT one of its major advantages! Quite the contrary, it is one of its disadvantages.
Use it for convenience, not for reliable availability!
Score: 0
|What I want to know is, why haven't they made any improvements to the read/write speeds of current disk drives? We moved from 5400RPM to 7200RPM drives in the mainstream, and there are high end drives that spin at 10,000 or even 15,000 RPM. But those are expensive and tend to heat up. Even at 15,000 RPM, I don't think you double the transfer speed and such a high speed motor is surely less reliable. Over the years they even increased on-board cache size and improved pre-fetching, and prediction logic of the on-board controller. Still, even with these improvements the overall transfer speeds have not climbed very fast or very high.
In my opinion, the next logical step would be to decrease the amount of data we read/write, through the use of compression or read/write multiple streams in parallel. They could add a dedicated hardware chip to compress/decompress data as it is written and retrieved from disk. It would be completely transparent to the user and even the operating system itself. This would not only increase transfer speeds, but increase the amount of storage too. Some might argue that this approach would compromise data integrity.
The other idea is to read/write onto multiple platters in parallel. Sort of a RAID-0 configuration at the hardware level. Most modern high capacity drives use 3 or 4 platters. Imagine if data was written to all 4 platters in parallel? The speed increase would be almost exactly 400%! Again, some might argue this would compromise data integrity. I'm not suggesting these ideas be used in sensitive areas where data is of utmost importance. But for the average user, the trade off would be well worth it. When the average person's hard disk fails, it fails. It doesn't matter if it's platter 1, platter 2, or all platters developed a problem, the drive is dead.
Score: 0
|Excellently said, kashin.
Not much of a statistics man myself, would anyone able able to tell me the average read/write time for an HDD at 7200RPM and 15000RPM vs SSD?
Score: 0
|What is strange is that someone would expect a doubling of linear rotational velocity to result in a corresponding linear doubling of angular velocity and tangential physical access times...
And pre-fetch and look ahead buffers ONLY work if what you read is sequential! Otherwise you essentially page fault and must reload the data based upon what is required next after a physical access!
This is something that anyone who has experience with striping knows! Such technology works great for large contiguous read only databases that enable large stripe sizes , but is a complete waste with fragmented editable filesystems (but don't tell that to the PC folks who have only discovered this technology in the last few years! And large stripe sizes also result in an incredible aste of space in an editible files system. Even if you change 1 character, the minimum write size if the size of your strip and thus you may easily waste 128MB, or much larger depending upon the strip size, for a single character edit that is not placed in a non-contiguous location thus negating the look ahead buffer efficiencies and requiring a page-fault and a physical seek - the slowest most inefficient process in a drive.)
And you are going to increase throughput by employing the overhead of compression and decompression. Right! Completely transparent huh? Well, I guess if you are ignorant of what is actually involved, its transparent.
Unfortunately any glitch cannot be transparent to the OS or to the drives - especially if the controllers employ a software and a redundant hardware based bad block relocation schema that is so common in the even that you encounter the common bad block on a drive. Compromised data integrity? Its called corrupted data.
And you already have the ability to compress and decompress data, and CPU speeds are far faster than the physical data transfer speeds! You have your bottlenecks reversed.
Multiple platters in parallel. What you are describing is mirroring. And you mirror for availability, NOT performance! You net increase is NOT 400% but 0%!!!!! And if you encounter a bad block relocation on one of the platters? or if all of the filesystems are not defined as symmetric across all platters in a drive? (Many OSes already have logical volume managers that allow low level inter and intra drive level storage policy and control of filesystem placement and optimization! Although this is also another indication of how far behind the curve PCs are.)
Parallel writes only make sense across multiple drives ad multiple controllers without a common point of failure. Again, this is nothing more than mirroring, and you gain NO performance gain.
For average users such techniques offer little. And the focus on hard drives alone as separate devices will offer only small improvements int he future, despite non-volatile memory schemes (which can offer some initial speed up as data reads and buffering is minimized.)
And they offer nothing for more advanced users where much more sophisticated data management procedures have been in place for more than 10 years.
Optimized access involves much more than drive design. It is an interactive system requiring coordination from the adapters/controllers, to the drives, to lower lever inter-physical volume allocation policies (ie for max performance the logical volumes should be located near the center of the disk; to intra-physical volume allocation policies where for maximum performance a maximum number of physical drives should be used for the logical volume logical partitions. This allows the LVM to schedule requests for long sequential reads or writes across physical disks in parallel.
Mirroring should generally disabled for maximum performance. If it is required however, the scheduling policy should be set to parallel and the allocation to strict which will cause the the copies to be places on separate physical drives and to perform the writes in parallel, thereby maximizing performance. In addition, reads would then be scheduled to the copy of the data required that is closest to the disk head, improving write performance. Write verification and mirror write consistency should be disabled preventing additional disk revolutions on each write to read back the data for validity, and it will also stop the monitoring and waiting for all writes to succeed before returning successful completion to the write.
Maximizing performance from the file system point of view requires other consideration at creation time, ranging from fragment size (using the largest fragment size will maximize throughput) at the expense of space utilization management. Compression increases the overhead of reads and writes and therefore from a performance point of view should not be used, and each filesystem must have its own log file - sharing a common log file for a journaled filesystem simply produces bottlenecks. Ands also, using a JFS can increase response time as the actual disk physical write is not performed until the JFS buffer is full. (but some DBs depend upon the writes being written promptly upon request in order to avoid data inconsistency if the write fails to occur before a system or device crash - as the DB logs will be out of sync with the JFS logs, resulting in an inconsistent state.
Additionally, the OS cannot be removed from this process, as parameters such as sequential read ahead parameters to determine how many extra pages of data to read ahead of the current one to read in. This provides for requests for subsequent pages to already be in memory and the time is saved. But a non sequential filesystem or fragmented drive placement will result in a page fault and the physical seek to locate and read the data into memory - defeating the look ahead buffering! Thus striping becomes a hindrance rather than an asset. Also, disk I/O pacing is needed to prevent applications that generate very large amounts of I/O from saturating the I/O queues with requests. High and low water marks must be managed and defined in order to put the process to sleep until a certain number of requests have reached the low water marks to render the process more efficient. Here SCSI drives offer a real performance benefit. When there are multiple requests in the a SCSI device driver queue, attempts to coalesce these requests into a smaller number of larger requests. Dynamic allocation of these values allow tuning of the physical devices with the operating environment characteristics for maximum throughput efficiencies. Additionally, the ability to define device queue limits can significantly aid devices that do not provide for sophisticated queue handling algorithms natively.
And finally, applications have the most control over actual performance in that they can provide for asynchronous disk I/O which means that control returns to the application immediately after a read /write request has been queued. Additionally, sync and fsync calls can schedule and write all memory data pages to disk and return immediately.
All of which goes to show that the actual disk drive is only a very small part of the performance issue.
Especially in Windows which lacks a large measure of this granular control which is much more common and fundamental on such systems as AIX, HP-UX and other enterprise systems.
Score: 0
|"Multiple platters in parallel. What you are describing is mirroring. And you mirror for availability, NOT performance! You net increase is NOT 400% but 0%!!!!! And if you encounter a bad block relocation on one of the platters? or if all of the filesystems are not defined as symmetric across all platters in a drive? (Many OSes already have logical volume managers that allow low level inter and intra drive level storage policy and control of filesystem placement and optimization! Although this is also another indication of how far behind the curve PCs are.)"
I think he meant platters operating concurrently, not in parallel. This would be more like striping, which would give him the kind of speed increase he's talking about.
It makes some sense, but the costs associated with building such a device to be reliable would be a bit on the high side of prohibitive.
Score: 0
|Parallel is concurrently, if you will (and no, I am Not confusing them in my tongue in cheek reductionism)- both configurations already exist and are common in various raids and in various OS supported volume groups afforded by the intra and inter-policy of various logical volumes and filesystems spread across multiple drives which are the norm larger databases and distributed storage systems.
This is rather unusual only in the PC world.
Bottom line - you mirror only for availability (as there is an unavoidable performance hit due to overhead) and you stripe for performance. But striping has fundamental limitations. And these limitations come back to cream you when you have a fragmented randowm read/write volume where the large stripe size results in a gross waste of space, greater fragmentation resulting in physical mechanical seek times and significant memory page faults - all which completely negate the advantages of a striped system with a contiguous read only database...not to mention the maintenance 'nightmare' where if ANY part of the filesystem is damages on any drive or any partition, you must restore it all.
This is an old debate which has been demonstrated in the labs over 10 years ago - where clients were demanding both striping and mirroring be available simultaneously across multiple drives. They offset each other and you are better off simply mirroring for availability without the limitations of the striped volume damage requiring complete restoration.
The irony is that the capability was finally provided, not because of functional advantage, but only to placate user desires. The real pain came when so many would call in for support only to discover that the gains they thought they had provided for themselves in the event of failure were Pyrrhic, as many thought that they could simply avoid making regular backups, much to their chagrin.
Lots can be done with careful planning an placement through the inter and intra-policies of the volumes (placement within and across drives). But in order to really take advantage of this, OSes such as Windows need to substantially beef up their functionality. This is a major area where the large enterprise UNIXes continue to shine and where admin responsibility is a cakewalk. Once you have dealt with the 'built-in' tools provided within these environments, its hard to go back without significant regrets as the functionality and ease of implementation is simply not there.
Oh, and I might add, another thing that makes such environments so elegant is the tight integration of the OS with the 'smart' technology such as the SCSI variants. This ability to 'communicate' in tight fashion provides significant advantages. So there Are significant advantages to tight hardware and software integration to be had. But it comes at a price, as all of the various pieces and parts are not necessarily commodity pieces, as they must adhere to a common set of drivers and controller schema.
Score: 0
|Sorry, you missed my meaning.
In regards to striping vs. mirroring;
Parallel means both drives write the same bit at the same time (as though they were one unit). Concurrently means both drives are writing at the same time but independently (separate bits).
Perhaps my definition of the terms needs minor tweaking.
You are right though. This really wouldn't work in an "on device" multi-platter situation unless the drive itself controlled the placement/fragmentation of files, which would introduce a whole new nest of dragons.
Perhaps we should leave the striping to the gamers.
Score: 0
|An option not considered by you, or the industry?
Why not two sets of independent read/write heads per disk cylinder? Along with separate connections to controllers. Essentially two current disk drives, sharing the platters.
Disadvantages: Air turbulence, command queing.
Advantages: double thoughput to platter, read-after-write-checking, reading streamed-data (Tivo-pause), redundency - except for head crash.
I have not heard of anyone trying this - but it seemed like the next logical step, for the last 30 years.
Score: 0
|"the entry point for SSDs is about $500 for 32 GB capacity"
Can someone tell me WHERE I can actually get a 32GB SSD for only $500. I'd pay it, but I have never seen anything close to that price - certainly not for something I can just plug into my system as an IDE/SATA/SCSI device.
Score: 0
|