1) Over, say, the next three years, which aspects of SSD performance do you see advancing the most and the least.
2) The realistic way for an application to use NAND today is through an abstraction that makes it look like the flat, contiguous partition of it's rotational disk ancestor. Do you think that as SSDs get faster and applications use them in more sophisticated ways that the efficiency of this abstraction will hold up? Or will there a real need to change the abstraction, for example to include things like atomic multi-block writes.
As much as I love SATA drives, in the business space, SAS makes a huge difference. The fact that it is full-duplex vs. the half-duplex of SATA alone boosts what it can do. Are there any plans for Intel to release a native SAS based solution using this controller or a derivative?
No, but the same techinology should appear in next-generation SAS drives from Hitachi GST, since Intel has a Joint Development Program on SAS SSDs with Hitachi: http://www.intel.com/pressroom/archive/releases/20...
If I open 32 large files and append 32 bytes to the end of each file, and then close the files, how much NAND actually gets written in that process? Is it just 32 bytes x 32 (<1 page)? Or do 32 pages get written? Or does many hundreds or even thousands of pages get written? Do these writes get queued up so that only one page gets written? That's what I think the controller is doing but would be nice to know for sure.
Did they do any specific optimizations for databases? Specifically, how do they handle transaction logs that do large amounts small sequential writes that need to be persistent? Is a battery backed write cache still advisable for such workloads or can the onboard super-capacitors handle it by doing write combining?
Is there a concern with using these drives in a RAID-array that is constantly being written to?
Am I correct in assuming since these drives will be in an array most of the time, does the wear-levelling not work like usual? And does this reduce the life expectancy?
Any plans to bring TRIM to RAID arrays/controllers in the future?
Is there any possible way to increase SSD speeds for low QD usage, typical in consumer SSD'd ? or are SSD's going the way of x86 : best speed ups for paralle loads only?
Not really, there is nothing you can do. When you combine the disk IO from several machines all together, it is going to inherently be pretty random, as in reads and writes all over the place. That's just the nature of the beast.
Looking at some native PCI-E SSD's (Micron P320h in particular) you can see SSD's scale really high under certain work loads. The interface is always going to be a bottleneck with any decent controller and enough channels.
The next logical place to put an SSD controller would be into a SoC which is common place in the embedded area already. I suspect that x86 SoC's for laptops and possibly some desktops will start to incorporate an SSD controller on-die to reduce the number of components on a board as well as low power consumption.\ by a hair.
The thing I'm watching is performance under specific queue depths. Consumer workloads don't reach high queue depths that are required to make top of the line SSD's shine. With the Micron P320h as an example, it didn't hit its stride until QD of 256 was tested. Ultimately what needs to happen in the consumer space for SSD performance to improve is a refinement of storage drivers and better file systems. I strongly suspect that there are some optimizations there that can lead to improved performance for all SSD's but it may require a full break away from hard drives. Economics don't favor such development yet due to the capacity/price balance in favor of hard drives still.
All the independent drive vendors have had bad numbers the last few quarters. Enterprise is supplied by niche (mostly private) companies. The X-25E didn't go over so hot in Enterprise. Sun/Oracle released a flash appliance years ago, and Violin has one now. It appears that IBM is headed that way, too. So: is any SSD form factor in the future? Or will Linus be proven right:
... but Flash-based storage has such a different performance profile from rotating media, that I suspect that it will end up having a large impact on filesystem design. Right now, most filesystems tend to be designed with the latencies of rotating media in mind. -- Linus Torvalds/2007
What are the supported and non-supported RAID configurations of this drive? 5, 50, 60 and number of max drives. I assume 0, 1, and 10 work fine. Does TRIM pass to these RAID modes yet or simulate? Is TRIM necessary still for this drive?
I'd like to see low capacity drives with this controller for use as boot disks on a VM server 16 to 30 or 30 to 60 GB, any chance? I won't ever go lower than 16 GB These low capacity drives would likely just see a RAID 1.
Are there any software and/or hardware technologies on the horizon that will help to mitigate the issue of SSD endurance as process size continues to decrease?
My question is basically the same point Duke already raised.
Intel has already modified drivers to allow RAID 0 to pass TRIM commands under the 7-series chipset. Given the primary design of the DC S3700 drive is to give consistent IO performance and the targeting of data center customers, it would make sense to include TRIM pass through with RAID support. Will Intel be working on this (either with its own chipset, or possibly even other vendors as well) to enable this?
Secondly, does Intel assume TRIM is in place when it calculates the endurance of the SSD? Since TRIM would affect write amplification I would assume it would have somewhat of an impact on the drive longevity figures? Or would that impact be inconsequential with good idle garbage collection?
What technology do you guys see most likely replacing NAND in the next several years? And approximately when do you think that will happen? What will this new technology bring that allows it to scale beyond NAND?
Will anything like TRIM ever be implemented for Virtual File Systems? Foe example, if I have a VM with storage on top of a SAN that is thin provisioned, if that VM writes a bunch of data, the SAN will allocate storage, but if the VM then deletes that data, there is no way for it to signal up the chain, to the vmdk, and even to the storage underneath the vmdk that those blocks are no longer needed. Something like TRIM seems like it would be the logical solution here, but I have never heard anyone talking about it.
One question: How is the logical interface between the SATA and the internal storage improved since part of the memory access time is taken up by that process, not by the access time to the actual cells.
Can we expect to see a caching method for storage servers like we see on desktops, hopefully beyond the 60GB space? Perhaps something that could be implemented in current storage solutions without buying all new hardware.
If this is not possible it would still be nice to create a server that only did caching for other storage servers.
I am curious if this will be available for later viewing? I live in Australia, and 03:30 during the work week is a bit steep for me to watch live. If possible, enabling a way to watch later would be great. Thank you.
I'm by far no geek, and just try to get the knowledge I need for my current projects. I host a small website with 6 million pageviews/month. I'm running it on an older 120GB SSD drive. But now I'm concerned, that it will fail due to too much read/write operations. Would the new SSD DC S3700 be a good solution for me, or will I have to get myself a SLC based SSD drive - maybe even in the form of PCIe? My biggest problem is the price of these drives. Any suggestions would be appreciated very much.
2x An SSD in RAID1 (mirroring, 2x faster read performance, >1GB/s sequential).
I recommend 840 Pro, or even an enterprise drive, like S3700 for more reliability. Samsung 830 series are cheaper and reliable, though slower than 840 Pro.
I really should pay more attention to what is going on in town. The office that I work at is only a couple of blocks away from the salt palace. Exhibition tickets are $100 though....
What is the largest block that is guaranteed to be written atomically, i.e. when power is cut, either the whole block is overwritten or nothing is changed, but in no cases, one part of the block contains new data and one part old data.
It seems this technology makes TRIM somewhat less relevant both in regard to performance and endurance.
Was this technology/design born out of a desire to function without TRIM (in RAID, irrelevant of controller) and can we expect it to migrate to a consumer segment where TRIM becomes nice-to-have, but not a showstopper, when it comes to endurance?
Are SSDs not expected to be used in RAID configurations for Servers? What RAID configurations will be available? Will TRIM be implemented in RAID Controllers?
How should one expect this to perform compared to a 710 SSD when used over a SATA-II interface? Do most SATA-III SSDs behave the same when you do this? I've got one 710 and was looking to get some more but now I'm going to hold my horses, especially given the price differences. But a big unknown to me is exactly what one should expect when I go and place it onto an old Dell PERC 6 controller.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
33 Comments
Back to Article
dave_rosenthal - Friday, November 9, 2012 - link
1) Over, say, the next three years, which aspects of SSD performance do you see advancing the most and the least.2) The realistic way for an application to use NAND today is through an abstraction that makes it look like the flat, contiguous partition of it's rotational disk ancestor. Do you think that as SSDs get faster and applications use them in more sophisticated ways that the efficiency of this abstraction will hold up? Or will there a real need to change the abstraction, for example to include things like atomic multi-block writes.
James5mith - Friday, November 9, 2012 - link
As much as I love SATA drives, in the business space, SAS makes a huge difference. The fact that it is full-duplex vs. the half-duplex of SATA alone boosts what it can do. Are there any plans for Intel to release a native SAS based solution using this controller or a derivative?NandFlashGuy - Friday, November 9, 2012 - link
No, but the same techinology should appear in next-generation SAS drives from Hitachi GST, since Intel has a Joint Development Program on SAS SSDs with Hitachi:http://www.intel.com/pressroom/archive/releases/20...
Hitachi's next SAS SSD will also be utilizing the latest 12-Gbit SAS interface, which was demonstrated earlier this year:
http://www.tomshardware.com/news/Hitachi-Storage-s...
Kevin G - Friday, November 9, 2012 - link
Have you spotted any of the new Itanium 9500 series chips? The press release is out on Intel's site now. (For the curious: http://newsroom.intel.com/community/intel_newsroom... )Any chance of bringing one back for some testing?
mckirkus - Friday, November 9, 2012 - link
"reviewed it here"Should link here, assuming Intel has green lighted it this time ;)
http://www.anandtech.com/show/6432/the-intel-ssd-d...
TemjinGold - Friday, November 9, 2012 - link
Do you have any plans to introduce a consumer-level drive with these new technologies?Shadowmaster625 - Friday, November 9, 2012 - link
If I open 32 large files and append 32 bytes to the end of each file, and then close the files, how much NAND actually gets written in that process? Is it just 32 bytes x 32 (<1 page)? Or do 32 pages get written? Or does many hundreds or even thousands of pages get written? Do these writes get queued up so that only one page gets written? That's what I think the controller is doing but would be nice to know for sure.extide - Friday, November 9, 2012 - link
That depends largely on the application you are using to edit. It may write out all the bytes, or only the changed ones.stepz - Friday, November 9, 2012 - link
Did they do any specific optimizations for databases? Specifically, how do they handle transaction logs that do large amounts small sequential writes that need to be persistent? Is a battery backed write cache still advisable for such workloads or can the onboard super-capacitors handle it by doing write combining?DukeN - Friday, November 9, 2012 - link
Is there a concern with using these drives in a RAID-array that is constantly being written to?Am I correct in assuming since these drives will be in an array most of the time, does the wear-levelling not work like usual? And does this reduce the life expectancy?
Any plans to bring TRIM to RAID arrays/controllers in the future?
mayankleoboy1 - Friday, November 9, 2012 - link
Is there any possible way to increase SSD speeds for low QD usage, typical in consumer SSD'd ? or are SSD's going the way of x86 : best speed ups for paralle loads only?Jaaap - Friday, November 9, 2012 - link
According to Anand,It it possible to enhance VT-d to mitigate this "problem"?
extide - Friday, November 9, 2012 - link
Not really, there is nothing you can do. When you combine the disk IO from several machines all together, it is going to inherently be pretty random, as in reads and writes all over the place. That's just the nature of the beast.mayankleoboy1 - Friday, November 9, 2012 - link
Apart from the SATA3 interface, what is the biggest roadblock to increasing SSD speeds ?Kevin G - Sunday, November 11, 2012 - link
Looking at some native PCI-E SSD's (Micron P320h in particular) you can see SSD's scale really high under certain work loads. The interface is always going to be a bottleneck with any decent controller and enough channels.The next logical place to put an SSD controller would be into a SoC which is common place in the embedded area already. I suspect that x86 SoC's for laptops and possibly some desktops will start to incorporate an SSD controller on-die to reduce the number of components on a board as well as low power consumption.\ by a hair.
The thing I'm watching is performance under specific queue depths. Consumer workloads don't reach high queue depths that are required to make top of the line SSD's shine. With the Micron P320h as an example, it didn't hit its stride until QD of 256 was tested. Ultimately what needs to happen in the consumer space for SSD performance to improve is a refinement of storage drivers and better file systems. I strongly suspect that there are some optimizations there that can lead to improved performance for all SSD's but it may require a full break away from hard drives. Economics don't favor such development yet due to the capacity/price balance in favor of hard drives still.
FunBunny2 - Friday, November 9, 2012 - link
All the independent drive vendors have had bad numbers the last few quarters. Enterprise is supplied by niche (mostly private) companies. The X-25E didn't go over so hot in Enterprise. Sun/Oracle released a flash appliance years ago, and Violin has one now. It appears that IBM is headed that way, too. So: is any SSD form factor in the future? Or will Linus be proven right:... but Flash-based storage has such a different performance profile from rotating media, that I suspect that it will end up having a large impact on filesystem design. Right now, most filesystems tend to be designed with the latencies of rotating media in mind.
-- Linus Torvalds/2007
eanazag - Friday, November 9, 2012 - link
What are the supported and non-supported RAID configurations of this drive? 5, 50, 60 and number of max drives. I assume 0, 1, and 10 work fine.Does TRIM pass to these RAID modes yet or simulate? Is TRIM necessary still for this drive?
I'd like to see low capacity drives with this controller for use as boot disks on a VM server 16 to 30 or 30 to 60 GB, any chance? I won't ever go lower than 16 GB These low capacity drives would likely just see a RAID 1.
Hulk - Friday, November 9, 2012 - link
Are there any software and/or hardware technologies on the horizon that will help to mitigate the issue of SSD endurance as process size continues to decrease?Kougar - Friday, November 9, 2012 - link
My question is basically the same point Duke already raised.Intel has already modified drivers to allow RAID 0 to pass TRIM commands under the 7-series chipset. Given the primary design of the DC S3700 drive is to give consistent IO performance and the targeting of data center customers, it would make sense to include TRIM pass through with RAID support. Will Intel be working on this (either with its own chipset, or possibly even other vendors as well) to enable this?
Secondly, does Intel assume TRIM is in place when it calculates the endurance of the SSD? Since TRIM would affect write amplification I would assume it would have somewhat of an impact on the drive longevity figures? Or would that impact be inconsequential with good idle garbage collection?
extide - Friday, November 9, 2012 - link
What technology do you guys see most likely replacing NAND in the next several years? And approximately when do you think that will happen? What will this new technology bring that allows it to scale beyond NAND?extide - Friday, November 9, 2012 - link
Will anything like TRIM ever be implemented for Virtual File Systems? Foe example, if I have a VM with storage on top of a SAN that is thin provisioned, if that VM writes a bunch of data, the SAN will allocate storage, but if the VM then deletes that data, there is no way for it to signal up the chain, to the vmdk, and even to the storage underneath the vmdk that those blocks are no longer needed. Something like TRIM seems like it would be the logical solution here, but I have never heard anyone talking about it.Andre Oliveira - Friday, November 9, 2012 - link
It would be nice if you could wear Spok's pointed ears.DDR4 - Friday, November 9, 2012 - link
One question: How is the logical interface between the SATA and the internal storage improved since part of the memory access time is taken up by that process, not by the access time to the actual cells.Dug - Friday, November 9, 2012 - link
Can we expect to see a caching method for storage servers like we see on desktops, hopefully beyond the 60GB space?Perhaps something that could be implemented in current storage solutions without buying all new hardware.
If this is not possible it would still be nice to create a server that only did caching for other storage servers.
Spoony - Friday, November 9, 2012 - link
Anand,I am curious if this will be available for later viewing? I live in Australia, and 03:30 during the work week is a bit steep for me to watch live. If possible, enabling a way to watch later would be great. Thank you.
Betina - Saturday, November 10, 2012 - link
Hi.I'm by far no geek, and just try to get the knowledge I need for my current projects.
I host a small website with 6 million pageviews/month. I'm running it on an older 120GB SSD drive. But now I'm concerned, that it will fail due to too much read/write operations.
Would the new SSD DC S3700 be a good solution for me, or will I have to get myself a SLC based SSD drive - maybe even in the form of PCIe? My biggest problem is the price of these drives.
Any suggestions would be appreciated very much.
Best regards
Betina
thebluephoenix - Sunday, November 11, 2012 - link
Hi, geek here, not expert.Alternative and faster solution:
2x An SSD in RAID1 (mirroring, 2x faster read performance, >1GB/s sequential).
I recommend 840 Pro, or even an enterprise drive, like S3700 for more reliability. Samsung 830 series are cheaper and reliable, though slower than 840 Pro.
tkll - Saturday, November 10, 2012 - link
I really should pay more attention to what is going on in town. The office that I work at is only a couple of blocks away from the salt palace. Exhibition tickets are $100 though....stepz - Sunday, November 11, 2012 - link
What is the largest block that is guaranteed to be written atomically, i.e. when power is cut, either the whole block is overwritten or nothing is changed, but in no cases, one part of the block contains new data and one part old data.casper.bang - Monday, November 12, 2012 - link
It seems this technology makes TRIM somewhat less relevant both in regard to performance and endurance.Was this technology/design born out of a desire to function without TRIM (in RAID, irrelevant of controller) and can we expect it to migrate to a consumer segment where TRIM becomes nice-to-have, but not a showstopper, when it comes to endurance?
ReliableTech - Monday, November 12, 2012 - link
Are SSDs not expected to be used in RAID configurations for Servers?What RAID configurations will be available?
Will TRIM be implemented in RAID Controllers?
floam - Tuesday, November 13, 2012 - link
How should one expect this to perform compared to a 710 SSD when used over a SATA-II interface? Do most SATA-III SSDs behave the same when you do this? I've got one 710 and was looking to get some more but now I'm going to hold my horses, especially given the price differences. But a big unknown to me is exactly what one should expect when I go and place it onto an old Dell PERC 6 controller.FoV001 - Tuesday, November 13, 2012 - link
will you be uploading the SSD S3700 demo to the website later on today??