Slow NVMe performance

[Ad Hominem]

I recently got a Samsung 990 Pro 2TB M.2 drive to be my fast game library folder. Issue I'm having is that the performance seems to be well below expectations.

Running CrystalDiskMark for benchmarking, sequential read and write are about what you'd expect but the random is below half what I see the drive has done in benchmark tests.

[3 test count] [1GiB] [E:72%] [MB/s]

SEQ1M Q8T1 - 6539.62 / 6277.59
SEQ128K Q32T1 - 6528.03 / 6278.32
RND4K Q32T16 - 2463.65 / 1782.94
RND4k Q1T1 - 79.37 / 137.74

Compare that to the numbers on this review: https://www.thessdreview.com/our-reviews/nvme/samsung-990pro-gen-4-ssd-series-review-samsung-reigns-true-yet-again/3/

The sequential speeds are a little under but nothing that stands out to me, but the random speeds seem significantly below. I notice this in games, where friends also have a samsung NVMe - they are loading in to levels significantly faster than me.

I already have a samsing 970 for my OS that has been working great, but I went and downloaded the samsung drivers as well as samsung magician hoping that would help somehow, but nothing changed.

Overall my system is as follows:
ASUS ROG STRIX X570-E GAMING
AMD Ryzen 9 3950X
32 GB RAM @ 3600
GIGABYTE RTX 3070 VISION OC (GV-N3070VISION OC-8GD)

Will provide more information if it helps identify the issue. Starting to wonder if I just got a lemon. Is the problem that the drive only has ~28% free space (514GB free out of 1.81TB)?

[Electric Cupcake]

Slowing down as they get closer to full is a thing, though I wouldn't expect it until there's closer to 10% left.

You can try enabling Overprovisioning mode.

What's your TRIM schedule set at?

[Ad Hominem]

Samsing Magician does seem to recognize my 990, and checking the firmware in Magician it says it is up to date with the latest rev. Magician as well as HWInfo say the 990 is in good health. I have Windows set to trim the drive weekly. This thing is relatively new, it was delivered July 20th 2024.

[ˢᵈˣ FatCat]

maybe you using all the Sata port, about lane sharing? i forgot where i read those, anyway my 990 pro write 6700 supposedly 6900, i make my peace with that

i mean your friend could be using Low setting or 1080p while you higher so he loading faster, it's not end of the world

[Phénomènes Mystiques]

My WD Black NMVe drive has a "Game Mode" which helps it a bit.

[Illusion of Progress]

Keep in mind that when comparing to reviews, unless you have "controlled your variables" to ensure everything is the same, you have to expect that results may vary from those of the review.

Your Q1T1 number seems ~20% lower than the review, and that may be explained by a combination of the drive not being empty and other variables.

Your Q32T16 number seems to be just under 50% of the result the review gets. I don't know if that much of a difference is expected with the drive being largely filled.

Something else to point out is that drives on chipset connected ports will generally be a bit slower, at least sometimes, and I know this is often the case with the X570. I see it myself on mine (different motherboard though) where my second identical drive turns in slightly lower numbers. I think this might only dampen the sequential peaks a bit though and not so much the rest.

I'd generally recommend your fastest drive to be devoted to being your system drive because that's where you'd want it to be, unless you were devoting it a bandwidth heavy task like content creation. Faster drives are generally wasted on use cases like games. That's not really a use case that takes advantage of the benefits of faster drives.

All that being said, I don't know if the sub-50% number you got compared to one random review is expected or not, but unless you're willing to benchmark the drive when empty, it's hard to say.

[Ad Hominem]

Originally posted by Illusion of Progress:

Keep in mind that when comparing to reviews, unless you have "controlled your variables" to ensure everything is the same, you have to expect that results may vary from those of the review.

Your Q1T1 number seems ~20% lower than the review, and that may be explained by a combination of the drive not being empty and other variables.

Your Q32T16 number seems to be just under 50% of the result the review gets. I don't know if that much of a difference is expected with the drive being largely filled.

Something else to point out is that drives on chipset connected ports will generally be a bit slower, at least sometimes, and I know this is often the case with the X570. I see it myself on mine (different motherboard though) where my second identical drive turns in slightly lower numbers. I think this might only dampen the sequential peaks a bit though and not so much the rest.

I'd generally recommend your fastest drive to be devoted to being your system drive because that's where you'd want it to be, unless you were devoting it a bandwidth heavy task like content creation. Faster drives are generally wasted on use cases like games. That's not really a use case that takes advantage of the benefits of faster drives.

All that being said, I don't know if the sub-50% number you got compared to one random review is expected or not, but unless you're willing to benchmark the drive when empty, it's hard to say.

I can understand a 10-20% drop in performance and just chalk it up to differences between my rig and the review test rig, or the fact that I do have some stuff on the drive slowing it down some. But the random performance being 50% worse seems like it's beyond just that. Combined with the fact that my brother got the same exact drive and when we play hunt showdown he is loading in to the map and loading back into the menu twice as fast. We live in the same area and have comparable internet speeds as well.

I do have the drive in question installed on the 2nd M.2 slot, but I'm only using a single GPU so I shouldn't be limited on PCIe lanes or anything. Again I could write off 10 or 20% reduction from an ideal benchmark but 50% seems too steep

[Illusion of Progress]

Yeah, like I said, I'm not sure if the 50% difference is expected or not. The drive being empty in the review versus yours not, and other factors, may or may not explain it, but generally, when two identical things give different results, the answer lies in some other variable. So exploring the known ones is your best place to start.

[pwnograffik]

For comparison I have a 990 Pro 2TB (C: boot) and 4TB (G). Here are results with multiple apps open ~22GB and Marvel Rivals running on the 4TB G: drive. 3 test count, 1GB. X670E Steel Legend, 7950X3D.

2TB: 29%
7455.98 / 6031.44
7442.88 / 6908.72
5740.28 / 5044.19
85.19 / 238.88

4TB: 47%
6573.20 / 6423.85
6568.00 / 6417.96
6393.01 / 5256.68
85.62 / 250.59

Magician:
2TB:
7444 / 6924
1400390 IOPS / 1189697 IOPS

4TB:
6567 / 6437
1560302 IOPS / 1323974 IOS

The review test was performed on an empty drive so unless you also test empty then it's not an equal comparison. Also, not every slot is equal. Slot 1 on your mobo is controlled by CPU and slot 2 is controlled by chipset. I don't think that your results should be that bad on slot 2 but slot 1 will give you the best possible performance. Switch and re-test.

BIOS and chipset drivers up to date?

[Chad]

I have the same issue. 990 Pro 2TB with heatsink. About 10x slower in just the Random writes. Samsung magician shows random writes at 120,000 IOPS (it should be about 1.5 million). RND4K Q32T16 was 700M/b with Samsung High performance mode enabled, 1,100MB with it disabled.

EaseUS found the 4k alignment was off (even though the partitions were divisible by 4096). That didnt solve anything.

Ive tried just about everything, updated all firmware, disabled AMD Expo, disabled all power saveing features, bitlocker is disabled, etc.

Ive read about every thread on the internet about this, and there is alot. This seems common, and the only resolutions some have found were the 4k alignment. I think this might be a junk POS NVME. My other PC has the WD equivalent and its it's rated speed (so it's not a W11 issue).

[Ad Hominem]

Originally posted by tbialoof:

consumer ssds are designed for cheapness and will slow down as they get full because you only get full performance in slc mode, meaning you need the drive to be mostly empty to get full perf

also know that windows 10 doesnt support pcie 4 ssds at full iops, you need to use windows 11 to have the perf fixes

also make sure you have 10% overprovisioning enabled to avoid accumulating performance issues

That's all good info. I do have overprovisioning enabled. I need to double check my numbers against my 970 that is also my OS drive and ~70% full. That could be a good apples to apples test on the same system with the drive having data on it

[Ad Hominem]

Originally posted by tbialoof:

Originally posted by Ad Hominem:

That's all good info. I do have overprovisioning enabled. I need to double check my numbers against my 970 that is also my OS drive and ~70% full. That could be a good apples to apples test on the same system with the drive having data on it

yeah also as chad mentioned you do wanna check to make sure the drive is "4k aligned"

the way i understand the issue, if its out of alignment you essentially get each filesystem block overlapping between physical blocks of cells, for sequential write that doesnt really matter because you're reading or writing multiple blocks of data in a row, but for random synthetic tests, when you're trying to read or write a single block, because of the overlap it'll be reading two blocks of cells to read one filesystem block, so you're essentially having to do double the read and write per 4k block, which will half your performance in the test, this could genuinely explain what you're seeing

basically the drive could be formatted poorly, you want the virtual blocks of the filesystem and the physical blocks of the ssd to be 1 to 1

How do I check that, is that in like magician? Also is that something I would have to reformat to solve?

[Ad Hominem]

Originally posted by tbialoof:

if you open msinfo32 aka system information

expand components, storage

click disks

you'll see a list of partitions, and for each one there is a partition starting offset below it

to check each of those numbers, type them into calculator and divide them by 4096

(edit: or sorry, you might have to divide buy 512, because 8 bits per byte, but 4096 returns whole numbers for me too)

the result should be a whole number for each

if the number is not whole, ie, one of them has a decimal place, then there is your issue



and to fix the issue, i would suggest software like minitool, its nice as it will let you resize, move, align partitions without having to wipe them

ironically i think magician used to have a alignment fix as part of their "advanced optimisation" but i believe they got rid of it

Thank you for the instructions. For the drive in question I have a drive offset of 16,777,216 bytes. Dividing that by 4096 and 512 both come out as whole numbers.