Samsung 990 Pro (2TB) - Very poor random read and write performance

[DARIOcaptain]

I just installed a fresh Samsung 990 PRO 2TB drive in my laptop (in the PCIe 3.0 x4 slot). The read and write speeds are around 3200mb/s which is okay as it's installed in a gen.3 slot, however the random read and write values are terribly bad. I read that the PCIe generation shouldn't really affect the random read and write speeds so I have no idea where is the problem. The declared IOPS in this disk is up to 1.5M, while in my benchmarks it is around 120k.
The drive is used as my system drive, but I don't think it would limit the random read and write performance that much.

 

[Solaris17]

The declared IOPS in this disk is up to 1.5M, while in my benchmarks it is around 120k.

Are the benchmarks equal?

Are you cooling it? This things get toasty and they will throttle to nothingness.

 

[P4-630]

(in the PCIe 3.0 x4 slot)

That's one of your problems.

 

[DARIOcaptain]

Are the benchmarks equal?

Are you cooling it? This things get toasty and they will throttle to nothingness.

No, I have it without any heatsink. No heatsink fits in this laptop.
The temps are between 50 and 70 celsius.

That's one of your problems.

Isn't the random read and write speed dependant on the internal components of the drive only?

 

[R0H1T]

Not if you want 1.5M random IOPS on benchmarks.

 

[P4-630]

Since the laptop will only take PCIe 3.0 nvme's, I'd bought an PCIe 3.0 one, might as well saved me some cash...

Examples:

 

[massaker]

The declared IOPS in this disk is up to 1.5M, while in my benchmarks it is around 120k

@PCIe 3.0 theoretical maximum is around 810-850K IOPs, no way you could reach more than 900K. Just post your benchmarks, CDM (NVMe-SSD Setting!) and AS-SSD would be fine.

 

[dgianstefani]

I just installed a fresh Samsung 990 PRO 2TB drive in my laptop (in the PCIe 3.0 x4 slot). The read and write speeds are around 3200mb/s which is okay as it's installed in a gen.3 slot, however the random read and write values are terribly bad. I read that the PCIe generation shouldn't really affect the random read and write speeds so I have no idea where is the problem. The declared IOPS in this disk is up to 1.5M, while in my benchmarks it is around 120k.
The drive is used as my system drive, but I don't think it would limit the random read and write performance that much.

Drive performance depends on CPU/RAM performance. It's likely that your CPU, paired with a Gen 3 connection, simply isn't fast enough, bottlenecking the drive.

 

[DARIOcaptain]

Drive performance depends on CPU/RAM performance. It's likely that your CPU, paired with a Gen 3 connection, simply isn't fast enough, bottlenecking the drive.

Could be.

My CPU is Intel I7-7700HQ and the RAM is 32GB (2x16) 2600Mhz.
I notice that when the random read/write is being benchmarked the CPU usage goes from around 50% up to 100%.

@PCIe 3.0 theoretical maximum is around 810-850K IOPs, no way you could reach more than 900K. Just post your benchmarks, CDM (NVMe-SSD Setting!) and AS-SSD would be fine.

 

[chrcoluk]

I checked 7700HQ, its probably CPU saturated.

My 980 Pro was bottlenecked by my 9900k, and your chip is much weaker then that.

 

[DARIOcaptain]

I checked 7700HQ, its probably CPU saturated.

My 980 Pro was bottlenecked by my 9900k, and your chip is much weaker then that.

So - normally during the benchmark the CPU usage should stay below 100%? And if it doesn't then it's probably too weak to handle such SSD speeds?

 

[chrcoluk]

So - normally during the benchmark the CPU usage should stay below 100%? And if it doesn't then it's probably too weak to handle such SSD speeds?

On a 8 core chip below 12.5%.

 

[Assimilator]

Are you cooling it? This things get toasty and they will throttle to nothingness.

Not likely to be a problem if the drive is unable to hit its full speeds.

I notice that when the random read/write is being benchmarked the CPU usage goes from around 50% up to 100%.

The i7-11850H in my work laptop never goes above 1% total CPU usage while running the CDM benchmark, so I think it's safe to say you've found the bottleneck.

I do have to echo @P4-630 - not sure why you would buy a top-of-the-line PCIe 4.0 SSD and put it in a system that can only handle gen3 drives. But over and above that, why do you care that you get fewer IOPS than some arbitrary benchmark performed in an ideal-world scenario? If the drive is performing well enough for what you need it to do, nothing else matters; I see far too many people on this forum obsessing over benchmark results, when they could just be enjoying their purchase.

 

[massaker]

So - normally during the benchmark the CPU usage should stay below 100%?

@chrcoluk already meant it - look at the number of threads and queues! At QD=1 (Q1T1) it's 12,5%, but while it measures "in the 3rd line" (Q32T16 in CDM or Q1T64 in AS-SSD) it will use all cores/threas of your CPU. Especially when writing in the third line, it is normal for your CPU usage to reach 100%. So, if you have too few IOPs there (and therefore low score in AS-SSD) - your CPU is just too weak and bottlenecked your SSD @high queue depth. And yes, high QD and maximum IOPs is the least important discipline for normal use scenarios (deal with it if you actually have a multi-user server that thousands of users access simultaneously - only then is it really the most important discipline!).

The declared IOPS in this disk is up to 1.5M, while in my benchmarks it is around 120k.

No, 120K IOPs @4K are equivalent to ~490 MB/s (as you know 1M IOPs @4K = 4096 MB/s, so 4096 x0,12 = 491,52 MB/s exactly).
Although, in your Benches you reach ~1980 MB/s and ~480K IOPs (1979x244,14 = 483K IOPs).
Maximum write IOPs you reach in AS-SSD (AS-SSD shows MiB, so 1M IOPs = 3906,25 MB/s) -> 1766:3,91 ~ 452K IOPs.
The numbers are not bad for such old CPU AND the OS-SSD with Windows running on it.
My 990Pro Benches with 3 y.o. Ryzen CPU don't show 1,5M IOPs as well:

as you can see, max. read 5775 x244,14 ~1410K IOPs, here ist the same Benchmark with IOPs indication as a proof:

so, in high-QD writes i couldn't even reach 1,2M IOPs with my 8-core CPU @100% (assume 16-core CPU needed for 1,5M in reads/writes).

Now, in my old PCIe3.0-system i only benchmarked my older 980Pro:

As you can see, in reads/writes i had only 810K/660K IOPs.
Maximum IOPs i have reached with 980Pro @PCIe3.0:

So, 3486,24x1M:4096 = 851K IOPs (achieved with 16-core 3950X Ryzen CPU).

But, as already said, for general use, better look @ SEQ1M and RND4K results @ Q1T1, not max. IOPs.

 

[DARIOcaptain]

@chrcoluk already meant it - look at the number of threads and queues! At QD=1 (Q1T1) it's 12,5%, but while it measures "in the 3rd line" (Q32T16 in CDM or Q1T64 in AS-SSD) it will use all cores/threas of your CPU. Especially when writing in the third line, it is normal for your CPU usage to reach 100%. So, if you have too few IOPs there (and therefore low score in AS-SSD) - your CPU is just too weak and bottlenecked your SSD @high queue depth. And yes, high QD and maximum IOPs is the least important discipline for normal use scenarios (deal with it if you actually have a multi-user server that thousands of users access simultaneously - only then is it really the most important discipline!).

No, 120K IOPs @4K are equivalent to ~490 MB/s (as you know 1M IOPs @4K = 4096 MB/s, so 4096 x0,12 = 491,52 MB/s exactly).
Although, in your Benches you reach ~1980 MB/s and ~480K IOPs (1979x244,14 = 483K IOPs).
Maximum write IOPs you reach in AS-SSD (AS-SSD shows MiB, so 1M IOPs = 3906,25 MB/s) -> 1766:3,91 ~ 452K IOPs.
The numbers are not bad for such old CPU AND the OS-SSD with Windows running on it.
My 990Pro Benches with 3 y.o. Ryzen CPU don't show 1,5M IOPs as well:
View attachment 353992as you can see, max. read 5775 x244,14 ~1410K IOPs, here ist the same Benchmark with IOPs indication as a proof:
View attachment 353994so, in high-QD writes i couldn't even reach 1,2M IOPs with my 8-core CPU @100% (assume 16-core CPU needed for 1,5M in reads/writes).

Now, in my old PCIe3.0-system i only benchmarked my older 980Pro:
View attachment 353996As you can see, in reads/writes i had only 810K/660K IOPs.
Maximum IOPs i have reached with 980Pro @PCIe3.0:
View attachment 353997 So, 3486,24x1M:4096 = 851K IOPs (achieved with 16-core 3950X Ryzen CPU).

But, as already said, for general use, better look @ SEQ1M and RND4K results @ Q1T1, not max. IOPs.

Alright, thank you.
That makes the situation a whole lot clearer to me.

I bought a 4.gen PCIe drive, because I plan to use it in a PC I'm going to build soon.