Understanding IOPS vs. Latency (in Random read/write)

[balnazzar]

I was wondering if you could help me understanding the importance of IOPS vs. Latency in SSDs.

Allow me to give a bit of context. I'm trying to decide between Firecuda 530 and Samsung PM9A3 for an AI-gpu workstation. These are very different beasts, the latter being an enterprise drive, but it's interesting because it provides a feature I deem to be quite desirable: power loss protection. The former, on the other hand, is considered to be one of the best consumer ssds with some prosumer features (namely high TBW and hardware encryption). I have to say also that for my use case, random, and not sequential performance, is important.

I was looking at storagereviews' reviews of both these products, and noticed that, for 4K random read, the figures are very different:

PM9A3: 905K IOPS at a latency of 562µs.
Fc.530 : 577K IOPS at a latency of 219µs.

How should I read these figures? One drive has almost twice the IOPS, but also more than twice the latency.. Which one will perform better?

Another thing that got me curious is that for the PM9A3 these figures are in line with the specs declared by Samsung, but it's worth noting that that's the U.2 version of the PM9A3.. For the M.2 version the specs are much worse.. So in case I decide for the samsung, I'd go for an adapter like the one linked below. My question is: would such an adapter compromise the drive's performance?
Thanks in advance.

Adapter, U.2 to PCIe - 2.5' U.2 NVMe SSD - Drive Adapters and Drive Converters | Italy

 

[Rand__]

I assume you have read the first few hits?

SSD performance parameters - latency vs IOPS?

The most used performance metrics for SSDs are said to be: Throughput (MBps), IOPS and latency (ms). I'm a bit confused from the latter 2, since: IOPS - measures how many input output operations are

stackoverflow.com

IOPS, Throuhput, Latency Explained: Understanding SSD, HDD Storage Performance Benchmarks | Gear Primer

If you've ever read some specifications for an SSD or HDD these days, you might be left scratching your head. That's because those who are familiar with

www.gearprimer.com

SSD Throughput, Latency and IOPS Explained – Learning To Run With Flash - The SSD Review

As fast as SSDs have found mainstream consumer use, they are unfortunately grouped in the same pictu

www.thessdreview.com

In the end all those numbers are just numbers, and which one is best for your particular use case depends.
It depends on your particular use case, the mix of read and write (even if you only read you still have writes, access dates, metadata etc), block size etc.

Then you need to understand that there is more to it then just the "Big 3" as they called in in one post - you need to look at the process from end to end to really identify how things are read (and why one is better than the other).
The major problem here is that you usually don't know half of the sh*t you'd need to to really make an educated guess - unnless you are writing your own client tools or they tell you how they access data you wont know if the run their own blocksizes, just leave it to the OS, parallelize their queries or not and so on.

So nobody can tell you if double iops for double latency is better for your use case or not... if you got a lot of small individual reads/writes then a lower latency for each individual write might be beneficial since you might never need more than 500k IOPS per drive ... or the other way round.

You also need to look at queue depth (from a client point of view) - do you run enough client processes (threads) to utilize parallel access pipes and or high QD's? That might totally skew the data that you use as reference material...

tldr; we need more data to provide an educated guess.

p.s. Each adapter induces latency, but an adapter as you linked should be okish. I'd run U2 with adapter over M2 anytime (higher performance of U2, significantly larger heatsinks, potentially easier placement in the chassis if you use one with a cable)-

edit: pps - Make sure to get an adapter that matches your PCIe version

 

[balnazzar]

Thanks for your reply. Indeed I should have provided more info.

It's an AI workstation, I work mainly with tabular data and NLP/NLU, so the ssd will read/write a lot of very small files. Queue depth will not be more than 8-16, but more likely it will be 1-2-4 the vast majority of times.

What do you think?

As for PCIe, I'll place the adapter upon a 3.0 X4 slot. That shouldn't create problems, since I don't value sequential speed so much. But I can always buy an adapter cable powered by a sata plug, and link the U.2 ssd to an M.2 pcie 4.0 slot.

 

[Rand__]

We are talking about a single drive? What size and budget?I actually would recommend Optane for low qd's if sizing allows it...

Else the one with lower latency might be the better choice, especially if you can cool it well and its a native pcie4 drive in a pcie4 m2 slot.

 

[balnazzar]

We are talking about a single drive? What size and budget?I actually would recommend Optane for low qd's if sizing allows it...

Of course an Optane would be the ideal solution. But I need 2Tb bare minimum, and I have 500 EUR budget. At this price, not even heavily used units on ebay can be had, sadly.

Else the one with lower latency might be the better choice, especially if you can cool it well and its a native pcie4 drive in a pcie4 m2 slot.

As I imagined. I can live with the shorter life of the consumer ssd, but one thing that pisses me off is the absence of PLP... Let's hope an UPS will suffice..

 

[Rand__]

You running transactions of any kind? Or is everything repeatable?
Plp is really only needed if loosing a write will corrupt something...

 

[balnazzar]

You running transactions of any kind? Or is everything repeatable?
Plp is really only needed if loosing a write will corrupt something...

Mh, no, as a matter of fact.
I thought that a power loss could break any ssd if it was writing something at that instant...

 

[Rand__]

Nahhh. where did you get that idea from

 

[NablaSquaredG]

There *are* SSDs that break on unexpected power loss (there was a study back in ~2013), but as far as I remember it was only one model and is HOPEFULLY fixed in all current non-no-name models on the market!

 

[balnazzar]

Guys, I was eyeing the P5510. Top-notch reliability and PLP.
It's not an Optane drive, that's for sure, but look at the review.. Intel P5510 NVMe Enterprise SSD Review

It does over 700K IOPS with a latency of just 165 microseconds at low loads (presumably, low QD and threads).

That's mazing.. Compare this with the arguably fastest consumer SSD in 4K random read at low QD/threads, that is the Firecuda 530Phison E18 with 176Lnand): Seagate FireCuda 530 SSD Review

The P5510 seems to beat it hands down both in IOPS and latency.. But then I found this:

Just 43 Mb/s! How can this be explained?? The Firecuda 530 does almost *twice* (85 Mb/s) in Q1T1 4k random read...

 

[Rand__]

You might be comparing apples to oranges here - which is a problem that you often have when comparing numbers from different sources.

The P5510 test was done on a fairly slow CPU (4110, max speed 3GHz) and its not clear (to me) what ssdreview was using.
Secondly its always a question of cache and a way of excluding the impact of it...

Where is the CDM result of the FireCuda you are referencing?

 

[balnazzar]

Where is the CDM result of the FireCuda you are referencing?

https://www.servethehome.com/seagate-firecuda-530-1tb-nvme-ssd-review-sustained-write-king/2/

Particularly: https://www.servethehome.com/wp-con...gate-Firecuda-530-1TB-CrystalDiskMark-1GB.png

Also: Is the FireCuda 530 the fastest NVMe yet?

 

[Rand__]

Well,
first @WillTaillac was using a better CPU (5950x) so that might account for something (although probably not doubling performance).

Then the question is how reliable these values are in the first place - I can conjure a worse value too - only reaching 60MB/s here...

and some issues to boot

So take all benchmarks with a grain of salt - I've had times when running the same benchmark 3 times on the same drive I'd get three different results (Pool performance scaling at 1J QD1) [fixed by secure erase] and just to mention it.

In the end performance numbers are just numbers, and whether you can recreate them with your specific setup and use case remains to be seen.
I'd suggest to get the drive that fits the bill best - from cost, to form factor, cooling, size, performance, resell value to whatever else is important to you. And dont forget, next year (or more lilkely right after you bought one) there's gonna be a new best drive out there

 

[i386]

There *are* SSDs that break on unexpected power loss (there was a study back in ~2013), but as far as I remember it was only one model and is HOPEFULLY fixed in all current non-no-name models on the market!

There were different model/vendors where this could happen, but after 2016 it seems that most ssds had a workaround/fix (even snia mentions this in one presentation)

 

[balnazzar]

Then the question is how reliable these values are in the first place - I can conjure a worse value too - only reaching 60MB/s here...

That's spites me quite a bit.. It seems that there is no reliable source upon which one can make a "best buy"... :-/

 

[balnazzar]

Mhh, the review is a bit suspect TBH.. Firstly, note that the guy seems to just pay attention to seq performance. Then, in the review he says "uh, the pc all of a sudden became very slow". Still, he was getting over 6Tb/s versus the original ~7Tb/s. No way that he could have "felt" this without benchmarking... Secondly, he did the two benchmarks differently.. The first time he was running CDB in Default mode (random shown in Mb/s) the second time, probably in Real World (or Peak) mode (random shown in IOPS and microseconds).

It stinks.

 

[Rand__]

I know.
Was just a way to illustrate that not all benchmarks are equal and you can't really rely on just 'any' out there...

 

[balnazzar]

I ended up picking a Samsung 983 ZET 480gb on ebay.. Couldn't resist, it was quite cheap (203 bucks shipped). Used but with less than 2 Tb written.

In the end, I managed to get some kind of poor man's Optane drive Will use it as system drive and for swap and the smaller datasets. As main data drive I'll use the 530 2Tb.

 

[Rand__]

Lol - there went the "at least 2TB" requirement

But whatever works for you is good o/c - good luck with it.

 

[Styp]

Did you test it? Do you see improvements on the performance of your DL Pipeline?
I am currently in the market of buying a new M2 drive, and I am wondering if its really worth the extra dollar.