Why does random write performance go down with increasing size, on the P3608? 50KIOPS at 4TB, 80KIOPS at 3.2TB, and 150KIOPS at 1.6TB? Is that a misprint or is there a reason for the decreasing performance?
It could be that they are using higher-density NAND in fewer packages in the higher capacity drives, while the lower capacity drives are still using lower-density (older) NAND. I don't know if this is why, but it's a possibility.
I really think PCI-E sata-e NVM drives are doomed, the power consumption is simply too high. Considering that an spinning HDD consumes at most 8-10w writting data.
At 100MB/s sequentially, so you get 30x the performance on the 4TB model for 4x the power. Not to mention if you have random writes, in IOPS/watt this is vastly superior. The only place HDDs win is in pure bulk capacity but for a TB-size database that's not what matters at all.
High peak power consumption for a very short amount of time, as Lolimaster already said. There's no free lunch here - those electrons have to be pushed into their places if you want to write to flash.
I am pretty sure spinning a hard drive uses more power than pushing electrons. The high power usage is most likely because of the extra processing taking place in the SSD.
Time is a factor in power consumption. If the SSD uses 40 watts and finishes the job in 0.5 second, while the spinner needs only 10 watts but takes 4 seconds to do the job, then surprise surprise, the spinner is using twice as much power.
Good point, but it's important to distinguish between power (watts) vs energy (joules). FTFY: "Time is a factor in energy consumption..." "... the spinner is using twice as much energy."
Do'h, the SSD consumes more power idling than the spinner while working :) If your system storage IO is loaded 100% of the time it is a good thing, but if the storage IO is idling, the SSD will inevitably use more power.
It's an enterprise drive, and a high-end one, so there will be practically no idle time. That's why manufacturers don't employ any idle power states like client drives do.
Highspeed SSDs are used for IOPs limited situations not bulk storage; one SSD can service more random IO requests than a storage chassis stuffed with dozens of disks storing dozens of copies of the same data.
I'll be benchmarking the 1.6TB when I get back home from Samsung's SSD Global Summit, using our usual enterprise test suite. But what we can expect is that you need a really heavy multithreaded workload in order to make this drive really shine. Aside from implementing NVMe properly, there's simply not much that can be done to improve random reads at low queue depths, and that's the bottleneck for most interactive tasks.
I'm pretty sure I don't have a big enough CPU to do C++ compilation with enough parallelism to produce the high queue depths needed for full performance from these drives.
For my next developer box, im seriously considering a semi-silent mini-itx box. After seeing those numbers, for a sec, i thought ATX was going to be back in style.
I know im CPU limited for compiles, but havnt looked too closely. Was wondering if there was something to be eeked out of I/O instead of no-go super expensive Xeons.
You could experiment with a RAM drive to see if anything could be gained from crazy-fast IO during compile. A speedup there obviously wouldn't guarantee a high end SSD would help; but if a ramdrive doesn't give you anything, you know there's no need for anything beyond a consumer drive.
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
27 Comments
Back to Article
makerofthegames - Wednesday, September 23, 2015 - link
Why does random write performance go down with increasing size, on the P3608? 50KIOPS at 4TB, 80KIOPS at 3.2TB, and 150KIOPS at 1.6TB? Is that a misprint or is there a reason for the decreasing performance?nathanddrews - Wednesday, September 23, 2015 - link
It could be that they are using higher-density NAND in fewer packages in the higher capacity drives, while the lower capacity drives are still using lower-density (older) NAND. I don't know if this is why, but it's a possibility.baconforall - Wednesday, September 23, 2015 - link
"...the smallest 1.6TB configuration having more overprovisioning to boost random write speeds."makerofthegames - Thursday, September 24, 2015 - link
They got a 3x performance boost out of overprovisioning?! What did they do, use 4TB raw capacity on all three of them?baconforall - Thursday, September 24, 2015 - link
It sure looks like they did!http://www.anandtech.com/show/6489/playing-with-op
BrokenCrayons - Wednesday, September 23, 2015 - link
I'll take two 4TB models thanks. :)Shadow7037932 - Wednesday, September 23, 2015 - link
That would be one kidney.ddriver - Thursday, September 24, 2015 - link
I highly doubt it will be that cheap.BrokenCrayons - Thursday, September 24, 2015 - link
Totally worth it. Kidneys are overrated anyway.Lolimaster - Wednesday, September 23, 2015 - link
I really think PCI-E sata-e NVM drives are doomed, the power consumption is simply too high. Considering that an spinning HDD consumes at most 8-10w writting data.Kjella - Wednesday, September 23, 2015 - link
At 100MB/s sequentially, so you get 30x the performance on the 4TB model for 4x the power. Not to mention if you have random writes, in IOPS/watt this is vastly superior. The only place HDDs win is in pure bulk capacity but for a TB-size database that's not what matters at all.MrSpadge - Wednesday, September 23, 2015 - link
High peak power consumption for a very short amount of time, as Lolimaster already said. There's no free lunch here - those electrons have to be pushed into their places if you want to write to flash.ddriver - Thursday, September 24, 2015 - link
I am pretty sure spinning a hard drive uses more power than pushing electrons. The high power usage is most likely because of the extra processing taking place in the SSD.wbwb - Wednesday, September 23, 2015 - link
Time is a factor in power consumption. If the SSD uses 40 watts and finishes the job in 0.5 second, while the spinner needs only 10 watts but takes 4 seconds to do the job, then surprise surprise, the spinner is using twice as much power.baconforall - Wednesday, September 23, 2015 - link
Good point, but it's important to distinguish between power (watts) vs energy (joules). FTFY:"Time is a factor in energy consumption..."
"... the spinner is using twice as much energy."
willis936 - Wednesday, September 23, 2015 - link
energy*ddriver - Thursday, September 24, 2015 - link
Do'h, the SSD consumes more power idling than the spinner while working :) If your system storage IO is loaded 100% of the time it is a good thing, but if the storage IO is idling, the SSD will inevitably use more power.Kristian Vättö - Thursday, September 24, 2015 - link
It's an enterprise drive, and a high-end one, so there will be practically no idle time. That's why manufacturers don't employ any idle power states like client drives do.DanNeely - Thursday, September 24, 2015 - link
Highspeed SSDs are used for IOPs limited situations not bulk storage; one SSD can service more random IO requests than a storage chassis stuffed with dozens of disks storing dozens of copies of the same data.SeanJ76 - Friday, September 25, 2015 - link
11 watts is too much? Lmao!!piiman - Sunday, September 27, 2015 - link
Only if power is your main concern. Its a trade off more speed or less power???blahsaysblah - Wednesday, September 23, 2015 - link
Any chance of a benchmark run? A compile of UE4 or AOSP as reference? (under Win 10 please)Or is everything CPU limited except actual DBs versus good SATA3 SSD?
Billy Tallis - Wednesday, September 23, 2015 - link
I'll be benchmarking the 1.6TB when I get back home from Samsung's SSD Global Summit, using our usual enterprise test suite. But what we can expect is that you need a really heavy multithreaded workload in order to make this drive really shine. Aside from implementing NVMe properly, there's simply not much that can be done to improve random reads at low queue depths, and that's the bottleneck for most interactive tasks.I'm pretty sure I don't have a big enough CPU to do C++ compilation with enough parallelism to produce the high queue depths needed for full performance from these drives.
DIYEyal - Wednesday, September 23, 2015 - link
Any word on pricing?Venya - Monday, September 28, 2015 - link
Intel SSD DC P3608 priced at $2.19 per GB.So we have $3,009 (1.6TB), $7,009 (3.2 TB) and $8,759 (4 TB)
blahsaysblah - Wednesday, September 23, 2015 - link
Thanks for info.For my next developer box, im seriously considering a semi-silent mini-itx box. After seeing those numbers, for a sec, i thought ATX was going to be back in style.
I know im CPU limited for compiles, but havnt looked too closely. Was wondering if there was something to be eeked out of I/O instead of no-go super expensive Xeons.
DanNeely - Thursday, September 24, 2015 - link
You could experiment with a RAM drive to see if anything could be gained from crazy-fast IO during compile. A speedup there obviously wouldn't guarantee a high end SSD would help; but if a ramdrive doesn't give you anything, you know there's no need for anything beyond a consumer drive.