SM951 (first?) benchmark

according to pudgetsystem commnets:

That's what I thought to be the case but that wasn't how it turned out. Windows installed and booted just fine in the M.2 slot. It would install, but it wouldn't boot in the PCIe adapter though and I'm not sure why.


Intel dc p3500 can not be preorder anywhere. There was 1 website accepting preorder but now they stop and say they will remove the drive from their database. So maybe the p3500 won't be out for another 6 months. That is too long. Just use the sm951 for now.

The P3700 is already out, just not to consumers. Several next-gen native PCIe controllers are getting ready to launch (most notably OCZ's new JetExpress) so if Intel hopes to stake their place in the enthusiast SSD space again like they did with the X25 they'll need to release the P3500 in the near future.

Even if Intel doesn't release the P3500 for 6 months, I can't boot to the SM951 and use it at full speed so it does me no good which is why I sold it (not to mention the NVMe issue).

It has spoiled me though, Sata drives seem so mundane now :p
 
I think at this moment in Jan. of 2015 we just have to wait till March of 2015 to see what is really available.

I am sure with demand, supply will kick in.
 
The P3700 is already out, just not to consumers.

P3700 is easily available to consumers. He's right, though, about the P3500. P3600 and P3700 are out, but P3500 isn't yet. Personally I am pretty upset with Intel for it because they are clearly trying to force you to spend 50% or whatever more for the P3600 before they allow us to buy the P3500. I'm not biting. They can keep the 3600 and 3700s. Not sure why everyone but me seems to overlook that Intel never ships anything on time. So many people out there think that Intel actually meet their roadmaps, lol.
 
P3700 is easily available to consumers. He's right, though, about the P3500. P3600 and P3700 are out, but P3500 isn't yet. Personally I am pretty upset with Intel for it because they are clearly trying to force you to spend 50% or whatever more for the P3600 before they allow us to buy the P3500. I'm not biting. They can keep the 3600 and 3700s. Not sure why everyone but me seems to overlook that Intel never ships anything on time. So many people out there think that Intel actually meet their roadmaps, lol.

Oops, that's what I get for not checking, P3700 and P3600 are fairly available now.

I seriously doubt they're purposely holding back the P3500 to entice people to buy the more expensive versions. Very few enthusiasts are desperate enough for the extra performance to pay that kind of crazy premium and Intel knows that.
 
I'd say wait one more generation because the XP941 is still quite fast. I doubt you'd notice a difference. If you want a real upgrade, you can try an Intel P3700 instead, but those are like $1250 for 400GB. Crazy. Enterprise gear. Even it probably won't be all that noticeable in most workloads over the XP941.
Maybe grub would've worked but I decided to just wait for the Intel P3500 since I couldn't get NVMe working anyway.
Just providing my input on this.
The original Intel X25-M controller and it's minor updates in the G2 & G3 received critical acclaim and are really fast for consumer workloads.
The last drive released in that series was called Intel SSD 710 and is aimed for the enterprise.
All of these drives provide good performance in the Anandtech Oracle Swingbench benchmark.
It's a benchmark which highlights performance of a corner case: IO performance at transfer sizes below 4KB.

The first Intel controller to move away from that original design, and use a flat indirection mapping table for the NAND was the Intel S3500 & S3700.
Interestingly these drives have appalling performance in the test described above.
The Intel SSD 730 also does, it uses the same controller but for "prosumer" use.
The Intel SSD DC P3700 also has really bad performance in this test, and the controller in that one is an evolution in the one found in the S3500/S3700 with added NVMe support.

Now there is one final piece to the puzzle, it is the Intel SSD 910 controller, it is a drive released before the ones described above, but it also has really bad performance in this test.
What all these drives share in common is that their performance below 4KB IO transfers is really bad.
It can be easily tested using Atto, because it benchmarks IO transfer sizes from 0.5KB all the way up to 8MB.
I have a 800GB Intel SSD 910 which I have flashed with a different firmware making it support built in RAID and making it bootable.
I'm running a fresh Windows 7 install on it and it honestly feels really slow at certain things, at others it just flies.
My theory is that it's due to the poor performance below 4KB.
But try as I may I have not found a way to get a breakdown of IO transfer sizes like Anandtech shows in their reviews, i.e. that 40% where 4KB in size, 20% where 16KB etc etc.

Code:
Atto 0.5KB Transfer - 2GB Sequential Write & Read - QD4 - KB/sec - higer is better

Product:		Write:	Read:	Controller:		Source:
Intel S3500 300GB	314	9679	Intel PC29AS21CA0	[URL="http://hardforum.com/showthread.php?t=1789271"]watooit[/URL]		(First Intel controller with flat indirection table)
Intel DC P3700 1.6TB	321	47369	Intel CH29AE41AB0	[URL="http://anandtech.com/bench/product/1239"]Anandtech[/URL]	(Controller is an evolution of the design used in the S3700/S3500 with added NVMe support)
Intel DC P3700 800GB	346	41856	Intel CH29AE41AB0	[URL="http://www.thessdreview.com/our-reviews/intel-ssd-dc-p3700-nvme-ssd-enthusiasts-report/2/"]The SSD Review[/URL]	(-||-)
Intel 730 480GB		329	9528	Intel PC29AS21CA0	[URL="http://anandtech.com/bench/product/1141"]Anandtech[/URL]	(Same controller as S3500/S3700 but runs at 600Mhz instead of 400Mhz)
Intel 910 800GB (200GB)	413	19356	Intel EW29AA31AA1	[URL="http://www.anandtech.com/show/6124/the-intel-ssd-910-review/5"]Anandtech[/URL]	(Test on single 200GB partition, same SSD controller as Hitachi Ultrastar SSD400M, SSD400S & SSD400S.B)
Intel 910 800GB (RAID0)	426	17536	Intel EW29AA31AA1	Per Hansson	(800GB RAID0 on all four 200GB SSD's using the built-in LSI controller)
Intel 710 200GB		14848	22344	Intel PC29AS21BA0	[URL="http://anandtech.com/bench/product/429"]Anandtech[/URL]	(Same controller as 320 but with a newer firmware revision for enterprise, 320 itself being the same as the X25-M G2)
Intel 320 160GB		15488	23503	Intel PC29AS21BA0	[URL="http://anandtech.com/bench/product/370"]Anandtech[/URL]
Samsung 470 256GB	17619	15283	Samsung S3C29MAX01-Y340	[URL="http://anandtech.com/bench/product/354"]Anandtech[/URL]
Intel 310 80GB (mSATA)	20224	34176	Intel PC29AS21BA0	[URL="http://anandtech.com/bench/product/567"]Anandtech[/URL]
Intel X25-M G2 160GB	17830	48140	Intel PC29AS21AA0	[URL="http://www.ssdreview.com/review/compare/intel-x25-m-g2-postville-160gb-2cv102hd/attodiskbenchmark.html"]SSD Review[/URL]
Intel Pro 2500 240GB	21578	28416	SandForce SF-2281	[URL="http://anandtech.com/bench/product/1299"]Anandtech[/URL]
Samsung XP941 512GB	26112	16981	Samsung S4LNO53X01	[URL="http://anandtech.com/bench/product/1201"]Anandtech[/URL]	(PCIe Gen2 x4 mode)
Samsung 830 512GB	27706	18722	Samsung MCX		[URL="http://anandtech.com/bench/product/428"]Anandtech[/URL]
Samsung 850 Pro 256GB	39708	45056	Samsung MEX		[URL="http://anandtech.com/bench/product/1253"]Anandtech[/URL]
Samsung SM951 512GB	44177	45696				[URL="http://hardforum.com/showpost.php?p=1041335229&postcount=11"]Aibohphobia[/URL]	(Probably in PCIe Gen2 x4 mode)
Samsung 840 Pro 512GB	53120	59882	Samsung MDX		[URL="http://anandtech.com/bench/product/743"]Anandtech[/URL]

Swingbench Intel 910 http://www.anandtech.com/show/6124/the-intel-ssd-910-review/5
Swingbench Micron P320h http://www.anandtech.com/show/6371/micron-p320h-pcie-ssd-700gb-review-first-nvme-ssd/3
Swingbench Intel S3700 http://www.anandtech.com/show/6433/intel-ssd-dc-s3700-200gb-review/5
 
Last edited:
Very interesting, when you say the Windows 7 install runs slow sometimes, what kind of tasks are you talking about?
 
Very interesting, when you say the Windows 7 install runs slow sometimes, what kind of tasks are you talking about?
It's really slow at booting, and it's also slow at launching a browser with many tabs open.
Both actions that I would assume cause allot of IO activity below 4K.
Looking for example at the Windows folder there are over 50000 files below 4KB in size, and the Program Files folder has another 20000.

I still have not found anything for Windows that will give me a IO breakdown like Anandtech has in their reviews.
But I did find one for Linux which is really nice.
I tested it on a RHEL v6.6 webserver that sees really heavy traffic. Disks are two 7200RPM 1TB drives formatted with EXT3 filesystem.
The software is called iosnoop

The first column is total count, and the second column is IO transfer size, I ran the trace for 60 seconds:

Code:
     83 512
     63 1024
     58 1536
     84 2048
     79 2560
     82 3072
    109 3584
    271 4096
I have removed the transfers above 4KB but as you see there are allot of transfers, all the way down to 0.5KB
I think this in itself prooves that Intel are wrong in the following quote, taken from Anandtech in their review of the S3700 in relation to the poor performance in Swingbench.
Because serving files using Nginx on RHEL v6.6 is certainly not a corner case!

Intel views this type of a scenario as unlikely (apparently the latest versions of Oracle allow you to force the redo log to use 4KB sectors instead of 512B sectors, which would make these transfers 8KB - 12KB in size), and thus simply didn't optimize for it with the S3700's firmware. Intel did confirm that should customer feedback indicate this is a fairly likely scenario that it would be willing to consider some alternate solutions to improving performance here, but at this point it seems like it doesn't matter for the vast majority of use cases.
Source: http://www.anandtech.com/show/6433/intel-ssd-dc-s3700-200gb-review/5
 
It is curious... as it might be more future looking for OS's to be optimized around a 4K aligned access structure, the current OS access logs are very mixed on transfers and file access size paths.

I don't personally think ATTO is worth all that much though as its simply sequential in nature, and nearly 100% compressible (which is why sandforce drives tend to use it to gauge their performance on the box).

The point you made Per Hansson is true though, the current intel drives are not very good at smaller than 4k file access patterns, even so they tend to do fairly well for most people.
 
It's really slow at booting, and it's also slow at launching a browser with many tabs open.
Both actions that I would assume cause allot of IO activity below 4K.
Looking for example at the Windows folder there are over 50000 files below 4KB in size, and the Program Files folder has another 20000.

I still have not found anything for Windows that will give me a IO breakdown like Anandtech has in their reviews.
But I did find one for Linux which is really nice.
I tested it on a RHEL v6.6 webserver that sees really heavy traffic. Disks are two 7200RPM 1TB drives formatted with EXT3 filesystem.
The software is called iosnoop

The first column is total count, and the second column is IO transfer size, I ran the trace for 60 seconds:

Code:
     83 512
     63 1024
     58 1536
     84 2048
     79 2560
     82 3072
    109 3584
    271 4096
I have removed the transfers above 4KB but as you see there are allot of transfers, all the way down to 0.5KB
I think this in itself prooves that Intel are wrong in the following quote, taken from Anandtech in their review of the S3700 in relation to the poor performance in Swingbench.
Because serving files using Nginx on RHEL v6.6 is certainly not a corner case!


Source: http://www.anandtech.com/show/6433/intel-ssd-dc-s3700-200gb-review/5

Oops double post...

You might take a look at HDTune as it has a Disk Monitor section on it that will show you realtime feedback of disk access sizes.
 
I guess we can't use sm951 full potential until motherboard manufacture rolls out support for nvme boot in uefi bios
 
I guess we can't use sm951 full potential until motherboard manufacture rolls out support for nvme boot in uefi bios

You should be able to initially boot to a non-NVMe device and then load an OS from an NVMe device if that OS has NVMe drivers. GRUB2 has NVMe support, I believe.
 
You should be able to initially boot to a non-NVMe device and then load an OS from an NVMe device if that OS has NVMe drivers. GRUB2 has NVMe support, I believe.
That is very interesting. Does that suffice for a Windows boot? I thought that the Windows kernel loader has to be able to read from the device.
 
That is very interesting. Does that suffice for a Windows boot? I thought that the Windows kernel loader has to be able to read from the device.

I'm not sure. Windows can be booted from GRUB in general and GRUB does have NVMe support, so I would think there is a good chance that you could do this with Windows (if your Windows install does have NVMe support). I'm not sure if Windows will require UEFI boot with NVMe or if it can do a legacy boot. I just don't have that much experience with GRUB and UEFI in general. I use rEFInd for UEFI booting at the moment instead of GRUB, so I have little UEFI experience.

Sorry I couldn't give you a specific answer but hopefully I've given you a hint that leads you to a solution.
 
Just providing my input on this.
The original Intel X25-M controller and it's minor updates in the G2 & G3 received critical acclaim and are really fast for consumer workloads.
The last drive released in that series was called Intel SSD 710 and is aimed for the enterprise.
All of these drives provide good performance in the Anandtech Oracle Swingbench benchmark.
It's a benchmark which highlights performance of a corner case: IO performance at transfer sizes below 4KB.

The first Intel controller to move away from that original design, and use a flat indirection mapping table for the NAND was the Intel S3500 & S3700.
Interestingly these drives have appalling performance in the test described above.
The Intel SSD 730 also does, it uses the same controller but for "prosumer" use.
The Intel SSD DC P3700 also has really bad performance in this test, and the controller in that one is an evolution in the one found in the S3500/S3700 with added NVMe support.

Now there is one final piece to the puzzle, it is the Intel SSD 910 controller, it is a drive released before the ones described above, but it also has really bad performance in this test.
What all these drives share in common is that their performance below 4KB IO transfers is really bad.
It can be easily tested using Atto, because it benchmarks IO transfer sizes from 0.5KB all the way up to 8MB.
I have a 800GB Intel SSD 910 which I have flashed with a different firmware making it support built in RAID and making it bootable.
I'm running a fresh Windows 7 install on it and it honestly feels really slow at certain things, at others it just flies.
My theory is that it's due to the poor performance below 4KB.
But try as I may I have not found a way to get a breakdown of IO transfer sizes like Anandtech shows in their reviews, i.e. that 40% where 4KB in size, 20% where 16KB etc etc.

Code:
Atto 0.5KB Transfer - 2GB Sequential Write & Read - QD4 - KB/sec - higer is better

Product:		Write:	Read:	Controller:		Source:
Intel S3500 300GB	314	9679	Intel PC29AS21CA0	[URL="http://hardforum.com/showthread.php?t=1789271"]watooit[/URL]		(First Intel controller with flat indirection table)
Intel DC P3700 1.6TB	321	47369	Intel CH29AE41AB0	[URL="http://anandtech.com/bench/product/1239"]Anandtech[/URL]	(Controller is an evolution of the design used in the S3700/S3500 with added NVMe support)
Intel DC P3700 800GB	346	41856	Intel CH29AE41AB0	[URL="http://www.thessdreview.com/our-reviews/intel-ssd-dc-p3700-nvme-ssd-enthusiasts-report/2/"]The SSD Review[/URL]	(-||-)
Intel 730 480GB		329	9528	Intel PC29AS21CA0	[URL="http://anandtech.com/bench/product/1141"]Anandtech[/URL]	(Same controller as S3500/S3700 but runs at 600Mhz instead of 400Mhz)
Intel 910 800GB (200GB)	413	19356	Intel EW29AA31AA1	[URL="http://www.anandtech.com/show/6124/the-intel-ssd-910-review/5"]Anandtech[/URL]	(Test on single 200GB partition, same SSD controller as Hitachi Ultrastar SSD400M, SSD400S & SSD400S.B)
Intel 910 800GB (RAID0)	426	17536	Intel EW29AA31AA1	Per Hansson	(800GB RAID0 on all four 200GB SSD's using the built-in LSI controller)
Intel 710 200GB		14848	22344	Intel PC29AS21BA0	[URL="http://anandtech.com/bench/product/429"]Anandtech[/URL]	(Same controller as 320 but with a newer firmware revision for enterprise, 320 itself being the same as the X25-M G2)
Intel 320 160GB		15488	23503	Intel PC29AS21BA0	[URL="http://anandtech.com/bench/product/370"]Anandtech[/URL]
Samsung 470 256GB	17619	15283	Samsung S3C29MAX01-Y340	[URL="http://anandtech.com/bench/product/354"]Anandtech[/URL]
Intel 310 80GB (mSATA)	20224	34176	Intel PC29AS21BA0	[URL="http://anandtech.com/bench/product/567"]Anandtech[/URL]
Intel X25-M G2 160GB	17830	48140	Intel PC29AS21AA0	[URL="http://www.ssdreview.com/review/compare/intel-x25-m-g2-postville-160gb-2cv102hd/attodiskbenchmark.html"]SSD Review[/URL]
Intel Pro 2500 240GB	21578	28416	SandForce SF-2281	[URL="http://anandtech.com/bench/product/1299"]Anandtech[/URL]
Samsung XP941 512GB	26112	16981	Samsung S4LNO53X01	[URL="http://anandtech.com/bench/product/1201"]Anandtech[/URL]	(PCIe Gen2 x4 mode)
Samsung 830 512GB	27706	18722	Samsung MCX		[URL="http://anandtech.com/bench/product/428"]Anandtech[/URL]
Samsung 850 Pro 256GB	39708	45056	Samsung MEX		[URL="http://anandtech.com/bench/product/1253"]Anandtech[/URL]
Samsung SM951 512GB	44177	45696				[URL="http://hardforum.com/showpost.php?p=1041335229&postcount=11"]Aibohphobia[/URL]	(Probably in PCIe Gen2 x4 mode)
Samsung 840 Pro 512GB	53120	59882	Samsung MDX		[URL="http://anandtech.com/bench/product/743"]Anandtech[/URL]

Swingbench Intel 910 http://www.anandtech.com/show/6124/the-intel-ssd-910-review/5
Swingbench Micron P320h http://www.anandtech.com/show/6371/micron-p320h-pcie-ssd-700gb-review-first-nvme-ssd/3
Swingbench Intel S3700 http://www.anandtech.com/show/6433/intel-ssd-dc-s3700-200gb-review/5
This is most interesting. I had came across the same data on the 730 drive a few months back when comparing to the 850 Pro via Anandtech bench. I found the results curious but didn't see much discussion of the implications anywhere. I must have skimmed over or forgotten the information from the s3700 review. It's enough to make me reconsider my boot drive choice from a 730 to a 850 pro. Any more information you have on this would be appreciated, I think it deserves its on thread.
 
Yawn, different SSDs for different purposes. Comparing an enterprise level drive made to sustain the same performance while mildly utilized to FULLY utilized is different than a Pro/Consumer drive made for booting.

If you look at other tests in enterprise workloads the ENTERPRISE Intels you mentioned totally smash the Samsung, and that puts the data upside down practically.


Buy a SSD for what YOU will use it for, and YOUR workload.
 
Yawn, different SSDs for different purposes. Comparing an enterprise level drive made to sustain the same performance while mildly utilized to FULLY utilized is different than a Pro/Consumer drive made for booting.

If you look at other tests in enterprise workloads the ENTERPRISE Intels you mentioned totally smash the Samsung, and that puts the data upside down practically.


Buy a SSD for what YOU will use it for, and YOUR workload.
730 series is marketed as a consumer/enthusiast drive same as 850 pro
 
Yawn, different SSDs for different purposes. Comparing an enterprise level drive made to sustain the same performance while mildly utilized to FULLY utilized is different than a Pro/Consumer drive made for booting.

If you look at other tests in enterprise workloads the ENTERPRISE Intels you mentioned totally smash the Samsung, and that puts the data upside down practically.


Buy a SSD for what YOU will use it for, and YOUR workload.

Bullshit.
Any server load will have a mix of both high queue depths and medium to low queue depths.
I just copied a large folder with many files to the Intel SSD 910.
The speeds ranged from insane down to 91KB/s
Any SSD developed for enterprise use is supposed to have a minimum sustained speed rating; you think Intel would have sold many if they said the drive can sometimes write at 91KB/sec if you copy the "wrong" type of files to it?
As an example the I/O transfer size that MySQL uses is 16KB, something this drive sucks at.
Older versions of MSSQL even had transfer sizes below 4KB, how would you feel about purchasing this drive for such a server only to have it perform slower than a mechanical HDD?

Transcend where given allot of flack for their shitty SSD's using JMicron JMF602B controllers 5 years ago that would take 1000ms to complete some operations.
I don't see how this is any different, some operation occur at 1500MB/sec while other are counted in the Kilobyte per seconds range!
 
Last edited:
I was specifically talking about the Intel S3700 and how it's an outlier in the group, and how the graph makes it look horrible, yet that's not really the case. This is especially true if you've tuned mysql and or/are also using percona.

http://www.storagereview.com/intel_ssd_dc_s3500_enterprise_review
Or maybe a review of the Samsung 840 Pro with the Intels thrown in too.
http://www.storagereview.com/samsung_ssd_840_pro_enterprise_ssd_review
Or how about the Samsung drives dropping near 50mb/s in mySQL testing?
http://www.percona.com/blog/2013/08/28/testing-intel-samsung-sandisk-sata-ssd/

The huge decrease in latency = why Intels seem 'faster' but have lower IOPs overall in desktop usage, Intel iirc also shows their IOPs at different queue depth than other manufacturers too so comparing is hard from Consumer to Enterprise drive even.

Like I said, it depends on your usage patterns but Intel has proven to be steady reliable over the years.

Just because the enterprise grade drive is slower than the 850 Pro running Windows doesn't mean the 850 Pro is still faster running mySQL or other enterprise workloads.

I have over 20 new SSD in a box waiting for me to run numerous performance tests on them utilizing mysql for database testing and percona.... as that's my current fun project to "test" and then I'll use those drive(s) for my analysis/research project. My plan was to always go with the Intel NVME PCie drive but for OS and other servers I wanted to test/compare which works overall best for mysql and all around CentOS "do it all" web-server since I can't afford PCIe SSD for every server I have, although that would be nice :)

From MySQL:

The default database page size in InnoDB is 16KB, or you can lower the page size to 8KB or 4KB by specifying the innodb_page_size option when creating the MySQL instance.

Note:

Increasing the page size is not a supported operation: there is no guarantee that InnoDB will function normally with a page size greater than 16KB. Problems compiling or running InnoDB may occur. In particular, ROW_FORMAT=COMPRESSED in the Barracuda file format assumes that the page size is at most 16KB and uses 14-bit pointers.

A MySQL instance using a particular InnoDB page size cannot use data files or log files from an instance that uses a different page size. This limitation could affect restore or downgrade operations using data from MySQL 5.6, which does support page sizes other than 16KB.

Although being limited on where you can restore would suck if you didn't own/lease your own root access level hardware (VM/baremetal).
 
Last edited:
I was specifically talking about the Intel S3700 and how it's an outlier in the group, and how the graph makes it look horrible, yet that's not really the case. This is especially true if you've tuned mysql and or/are also using percona.

http://www.storagereview.com/intel_ssd_dc_s3500_enterprise_review
Or maybe a review of the Samsung 840 Pro with the Intels thrown in too.
http://www.storagereview.com/samsung_ssd_840_pro_enterprise_ssd_review
Or how about the Samsung drives dropping near 50mb/s in mySQL testing?
http://www.percona.com/blog/2013/08/28/testing-intel-samsung-sandisk-sata-ssd/

The huge decrease in latency = why Intels seem 'faster' but have lower IOPs overall in desktop usage, Intel iirc also shows their IOPs at different queue depth than other manufacturers too so comparing is hard from Consumer to Enterprise drive even.

Like I said, it depends on your usage patterns but Intel has proven to be steady reliable over the years.

Just because the enterprise grade drive is slower than the 850 Pro running Windows doesn't mean the 850 Pro is still faster running mySQL or other enterprise workloads.

I have over 20 new SSD in a box waiting for me to run numerous performance tests on them utilizing mysql for database testing and percona.... as that's my current fun project to "test" and then I'll use those drive(s) for my analysis/research project. My plan was to always go with the Intel NVME PCie drive but for OS and other servers I wanted to test/compare which works overall best for mysql and all around CentOS "do it all" web-server since I can't afford PCIe SSD for every server I have, although that would be nice :)

From MySQL:



Although being limited on where you can restore would suck if you didn't own/lease your own root access level hardware (VM/baremetal).
Again, I'm looking at the Intel 730 series which is a consumer drive just like the 850 pro series for Windows/regular desktop usage. I'm not interested in mySQL.

That said, the 730 series is based on the S3500/S3700 series and I don't see any significant performance difference between them in desktop or enterprise situations.
 
ToddW2, please this isn't about performance at higher queue depths or I/O transfer sizes.
I said in my initial post "It's a benchmark which highlights performance of a corner case: IO performance at transfer sizes below 4KB."
All I'm trying to do is confirm if Intel are wrong in their assumption to not optimize for transfer sizes below 4K.


Oops double post...

You might take a look at HDTune as it has a Disk Monitor section on it that will show you realtime feedback of disk access sizes.
Thank you so much for the recommendation, it's amazing I've been using HD Tune for a long time but never used this feature, and it's exactly what I was looking for!

I have waited a couple of days to respond just to get a better feel for the drive, while writes & reads below 4K are not many, they have so bad performance that I think they may account in large for why the performance feels so "meh" from this drive...

I have been trying to get Windows to not write files that small, testing for example KB967351 which allows you to change the "Bytes Per FileRecord Segment" from Windows default 1024 bytes > 4096 bytes when formatting the drive by specifying the /L switch.

But it seems to have made no discernible difference.
I've also increased the NTFS cluster size from the default 4KB to 64KB without feeling much difference.

The Intel SSD 910's LSI SAS2008 controller itself does not allow changing the stripe size, which is fixed at 64KB, therefore all writes below 64KB will obviously only go to one SSD, limiting performance.

But still I feel that if I could just get Windows to not write files below 4KB to it that would make a large difference, or atleast prove me wrong ;)
As you can see from the fsutil fsinfo ntfsinfo output below all values are 4KB except for "bytes per sector" which is something the drives & LSI controller report, i.e. what is the minimum writeable size.
This got me thinking that being able to set any SSD to what in the mechanical HDD world is called "advanced format" (native 4K sector support) would take care of this, but so far I have not been able to find any way to force Windows into this mode.

Code:
Bytes Per Sector  :               512
Bytes Per Physical Sector :       4096
Bytes Per Cluster :               4096
Bytes Per FileRecord Segment    : 4096
 
Again, I'm looking at the Intel 730 series which is a consumer drive just like the 850 pro series for Windows/regular desktop usage. I'm not interested in mySQL.

That said, the 730 series is based on the S3500/S3700 series and I don't see any significant performance difference between them in desktop or enterprise situations.

OK, then we're talking about different things ;)

However...

Are your performance tests simply transfer rates or do you do latency testing as well?
If I can find X 2 times before you find it once then your transfer-rate of 3x doesn't mean too much unless my work-load includes moving a lot of medium+ sized files around and not doing a lot of 'searching' or 'random requests' that windows does.

IIRC Intel was faster (latency) than a lot of other drives, and specifically almost 2x faster in terms of latency as the 840Pro which should say something when comparing to other more cheap drives. That's something most users FEEL compared to sustained transfers. This is also why the Intel G2 80gb in a desktop TODAY will feel fast but when you start doing transfers it's not as fast as the 'fasest' SSD around.

Anyway, we can agree to disagree but all I'm trying to point out is for enterprise workloads consistent, reliable performance is greater than having the most badass transfer speed, and when your latency is among the lowest you're still going to 'feel' fast in desktop usage too.

For what it's worth I have a # of the 730s I plan to test, but for MySQL and linux file system testing only if I have time I'll do Win comparison. (Single Drive & Various RAID configs.)
I almost pulled the trigger on some S3700s but I don't need the endurance, yet if I stumble upon a couple for a deal I'll get them just to test and for the endurance throw them on a write heavy DB :D
 
ToddW2, please this isn't about performance at higher queue depths or I/O transfer sizes.
I said in my initial post "It's a benchmark which highlights performance of a corner case: IO performance at transfer sizes below 4KB."
All I'm trying to do is confirm if Intel are wrong in their assumption to not optimize for transfer sizes below 4K.



Thank you so much for the recommendation, it's amazing I've been using HD Tune for a long time but never used this feature, and it's exactly what I was looking for!

I have waited a couple of days to respond just to get a better feel for the drive, while writes & reads below 4K are not many, they have so bad performance that I think they may account in large for why the performance feels so "meh" from this drive...

I have been trying to get Windows to not write files that small, testing for example KB967351 which allows you to change the "Bytes Per FileRecord Segment" from Windows default 1024 bytes > 4096 bytes when formatting the drive by specifying the /L switch.

But it seems to have made no discernible difference.
I've also increased the NTFS cluster size from the default 4KB to 64KB without feeling much difference.

The Intel SSD 910's LSI SAS2008 controller itself does not allow changing the stripe size, which is fixed at 64KB, therefore all writes below 64KB will obviously only go to one SSD, limiting performance.

But still I feel that if I could just get Windows to not write files below 4KB to it that would make a large difference, or atleast prove me wrong ;)
As you can see from the fsutil fsinfo ntfsinfo output below all values are 4KB except for "bytes per sector" which is something the drives & LSI controller report, i.e. what is the minimum writeable size.
This got me thinking that being able to set any SSD to what in the mechanical HDD world is called "advanced format" (native 4K sector support) would take care of this, but so far I have not been able to find any way to force Windows into this mode.

Code:
Bytes Per Sector  :               512
Bytes Per Physical Sector :       4096
Bytes Per Cluster :               4096
Bytes Per FileRecord Segment    : 4096

You've mentioned it's an edge case.

You've mentioned the limitations of the controller.

It sounds more like your usage doesn't align too well with the drive you purchased :confused: which is what I've been saying all along... get a drive for YOUR purpose.
If your file system limits you to ___ or mysql limits you to ___ or you do lots of big transfers or small transfers then get ____. There isn't 1 drive that does well with everything, that's obvious.

With that said, I'm still very interested in your testing and your results. LIke you I like to figure things out to figure them out :) and obv. increase performance ;) Keep at it, and I'll keep checking back but stop with the point making in this thread ;) Curious if you find something _major_ to help.

Carry on!
 
Anyway, we can agree to disagree but all I'm trying to point out is for enterprise workloads consistent, reliable performance is greater than having the most badass transfer speed, and when your latency is among the lowest you're still going to 'feel' fast in desktop usage too.
Don't you see the error in your own text?
I wrote in my initial post that transfer speed is 321KB/sec for the 1.6TB P3700 when doing small I/O's
How much less consistent do you need the performance to be for you to accept it's bad?
 
OK, then we're talking about different things ;)

However...

Are your performance tests simply transfer rates or do you do latency testing as well?
If I can find X 2 times before you find it once then your transfer-rate of 3x doesn't mean too much unless my work-load includes moving a lot of medium+ sized files around and not doing a lot of 'searching' or 'random requests' that windows does.

IIRC Intel was faster (latency) than a lot of other drives, and specifically almost 2x faster in terms of latency as the 840Pro which should say something when comparing to other more cheap drives. That's something most users FEEL compared to sustained transfers. This is also why the Intel G2 80gb in a desktop TODAY will feel fast but when you start doing transfers it's not as fast as the 'fasest' SSD around.

Anyway, we can agree to disagree but all I'm trying to point out is for enterprise workloads consistent, reliable performance is greater than having the most badass transfer speed, and when your latency is among the lowest you're still going to 'feel' fast in desktop usage too.

For what it's worth I have a # of the 730s I plan to test, but for MySQL and linux file system testing only if I have time I'll do Win comparison. (Single Drive & Various RAID configs.)
I almost pulled the trigger on some S3700s but I don't need the endurance, yet if I stumble upon a couple for a deal I'll get them just to test and for the endurance throw them on a write heavy DB :D
I haven't seen any benchmarks that show the Intel 730 having a latency advantage over the 850 pro but if you have a source, please post it
 
Don't you see the error in your own text?
I wrote in my initial post that transfer speed is 321KB/sec for the 1.6TB P3700 when doing small I/O's
How much less consistent do you need the performance to be for you to accept it's bad?

What? I never based any of my forum posts on your data why do you think I would now, or agree with your data?

The same benchmark you linked to ranked your drive high in performance under REAL work loads:
http://www.anandtech.com/show/6371/micron-p320h-pcie-ssd-700gb-review-first-nvme-ssd/4
http://www.anandtech.com/show/6433/intel-ssd-dc-s3700-200gb-review/6

The point I argued was buy for your usage ;) and disagreed with your data. Your 1 case doesn't change my view of what is "consistent performance", sorry,


I haven't seen any benchmarks that show the Intel 730 having a latency advantage over the 850 pro but if you have a source, please post it
Grrr... I closed those tabs, when I get them found again I'll show you. But, I never mentioned the 850 Pro (vs730 latency), I said 840 Pro; there may not be data on the 850 Pro or the gap may be closer. I'll post or PM you when I find it.
 
Don't you see the error in your own text?
I wrote in my initial post that transfer speed is 321KB/sec for the 1.6TB P3700 when doing small I/O's
How much less consistent do you need the performance to be for you to accept it's bad?

To those who have disagreed - personally, I think this guy has a pretty good point based on the data that we currently have available. Even excluding my disappointment with Intel from their Feminist Frequency hate group partnership, there is no way I would buy one of their PCIE SSDs right now without seeing further testing done on desktop and gaming workloads. Personally, at this time, I no longer care to buy one of these and would probably sooner buy the SM951, even if it is AHCI-only.

Hopefully it's a problem that can be fixed with a firmware or driver update (especially considering Intel does have their own NVMe drivers) or even that it was some sort of mistake in how the benchmarks were run (though I'm not sure how).
 
Looks like he's using Windows 7 which doesn't have a native NVMe driver
With Windows 7 you need to install KB2990941.

Update to add native driver support in NVM Express in Windows 7 and Windows Server 2008 R2

Maybe to Windows 10 if I see conclusive proof the keylogger (which is unacceptable even in a free beta test) is 100% removed and not left in even if disabled by default in RTM).
I see you've fallen for the keylogger FUD started by Neowin (who have since deleted their article because is was blatantly false).

zdnet.com/windows-10-youve-got-questions-ive-got-answers-7000034462/
lifehacker.com/windows-10s-keylogger-fiasco-has-been-blown-out-of-pr-1642931793
 
I've got a 512GB SM951 inbound but not sure how long I'm going to keep it without NVMe support. Maybe the first drive that is just as fast and has NVMe support I will swap it out with.
 

Nope, not FUD. I happen to have read the EULA terms long before you replied to my post. They are unacceptable terms and it doesn't matter if some random people tell me it's nothing to worry about because it's simply unacceptable, period. Even in a free, test OS, nothing should be going back to MS without my express consent. It boggles my mind that ANYONE would agree to those terms, again, even for beta testing. Every beta test I participated in in the past required submitting bug reports when I had something to report, not hijacking my computer and sending whatever data it wants over.
 
Nope, not FUD. I happen to have read the EULA terms long before you replied to my post. They are unacceptable terms and it doesn't matter if some random people tell me it's nothing to worry about because it's simply unacceptable, period. Even in a free, test OS, nothing should be going back to MS without my express consent. It boggles my mind that ANYONE would agree to those terms, again, even for beta testing. Every beta test I participated in in the past required submitting bug reports when I had something to report, not hijacking my computer and sending whatever data it wants over.

Then don't accept the terms?
 
Back
Top