Bandwidth scales easily, IOPS do not. If you stripe a number of SSDs via RAID 0 their bandwidth scales almost linearly - provided the interconnect (the required number of PCIe 3.0/4.0 lanes) is not saturated of course. However the IOPS are barely affected in RAID 0 mode. Depending on the RAID implementation the IOPS might slightly improve, they might remain the same or even get a little worse.
It also depends on what IOPS we are talking about : 4K-QD1 IOPS will either not break a sweat or get a bit worse due to striping overhead. Longer QDs and 4K+ IOPS should see a gain from RAID 0 though. The same apparently applies to Cray's implementation, though I don't think they would ever use RAID 0, which is unsuitable for professional storage solutions due to high risk of data loss.
Nope you're wrong. My QLC NVME SSD is faster, the benchmark i dont remotely understand said so. This "Cray" thing didnt even mention M.2 or RGB so you know they arent the real deal. /s
Seriously though, they must mean 5 Million or something, as much as we understand how IO works under different types of load, EVERYONE quotes best case and this thing can easily do more than 50,000 if all the stars are aligned. 50,000 "Kilo" IOPS would probably be closer...
With 2 drives you gain a bit as you don't have to scan gigabytes of index file AND you look on other thing. You look at bursts, that would make your ssd's overheat within 2-20 minutes. Thing is that with such massive raid, locating things is the issue. I think they guarantee 50k on Q1 and usual speed is way above that but sometimes spikes down to 50k/s. In this case it does not matter that it can burst to 5M iops on favorable conditions, they need to know if it will hold up the sustained load.
So it was a misprint, bit I like your idea of quoting Q1 and worst case specs. Imagine a world where SSD and other storage manufacturers allowed you to have any understanding of min or sustained performance from the spec sheet alone, you could almost understand where the product sat in the market, shocking!
You're right, it is 50,000 "Kilo" IOPS. From press release: Unprecedented storage performance: ClusterStor E1000 systems can deliver up to 1.6 terabytes per second and up to 50 million I/O operations per second per rack – more than double compared to other parallel storage systems in the market today.
Maybe just post a link to the press release next time guys, if you can't even copy and paste, let alone add any commentary :( It scales to 50 *million* IOPS per rack, not 50,000. Thought that sounded odd.
30GB/s from 60TB is pretty impressive. Fair bit of flash involved there, presumably!
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
11 Comments
Back to Article
rpg1966 - Wednesday, October 30, 2019 - link
50,000 IOPS per rack?Cellar Door - Wednesday, October 30, 2019 - link
Must depend on the drivers used but that does seem low.Santoval - Wednesday, October 30, 2019 - link
Bandwidth scales easily, IOPS do not. If you stripe a number of SSDs via RAID 0 their bandwidth scales almost linearly - provided the interconnect (the required number of PCIe 3.0/4.0 lanes) is not saturated of course. However the IOPS are barely affected in RAID 0 mode. Depending on the RAID implementation the IOPS might slightly improve, they might remain the same or even get a little worse.It also depends on what IOPS we are talking about : 4K-QD1 IOPS will either not break a sweat or get a bit worse due to striping overhead. Longer QDs and 4K+ IOPS should see a gain from RAID 0 though. The same apparently applies to Cray's implementation, though I don't think they would ever use RAID 0, which is unsuitable for professional storage solutions due to high risk of data loss.
danielfranklin - Thursday, October 31, 2019 - link
Nope you're wrong. My QLC NVME SSD is faster, the benchmark i dont remotely understand said so.This "Cray" thing didnt even mention M.2 or RGB so you know they arent the real deal.
/s
Seriously though, they must mean 5 Million or something, as much as we understand how IO works under different types of load, EVERYONE quotes best case and this thing can easily do more than 50,000 if all the stars are aligned.
50,000 "Kilo" IOPS would probably be closer...
deil - Thursday, October 31, 2019 - link
With 2 drives you gain a bit as you don't have to scan gigabytes of index file AND you look on other thing. You look at bursts, that would make your ssd's overheat within 2-20 minutes. Thing is that with such massive raid, locating things is the issue. I think they guarantee 50k on Q1 and usual speed is way above that but sometimes spikes down to 50k/s. In this case it does not matter that it can burst to 5M iops on favorable conditions, they need to know if it will hold up the sustained load.danielfranklin - Thursday, October 31, 2019 - link
So it was a misprint, bit I like your idea of quoting Q1 and worst case specs.Imagine a world where SSD and other storage manufacturers allowed you to have any understanding of min or sustained performance from the spec sheet alone, you could almost understand where the product sat in the market, shocking!
xelc - Thursday, October 31, 2019 - link
You're right, it is 50,000 "Kilo" IOPS.From press release: Unprecedented storage performance: ClusterStor E1000 systems can deliver up to 1.6 terabytes per second and up to 50 million I/O operations per second per rack – more than double compared to other parallel storage systems in the market today.
abufrejoval - Wednesday, October 30, 2019 - link
As much as I like Anandtech for the personal compute stuff, when it comes to HPC, there is nothing like the source: https://www.nextplatform.com/2019/10/30/cray-revam...gfkBill - Wednesday, October 30, 2019 - link
Maybe just post a link to the press release next time guys, if you can't even copy and paste, let alone add any commentary :(It scales to 50 *million* IOPS per rack, not 50,000. Thought that sounded odd.
30GB/s from 60TB is pretty impressive. Fair bit of flash involved there, presumably!
close - Thursday, October 31, 2019 - link
[Checks author]... Anton, yeah, checks out. Copy/pasta.demu67 - Thursday, October 31, 2019 - link
Can't find a shop to buy one. Please help...