How important are 4K random performance? Are sequential numbers pointless?

Feel free to correct me here.

I think most people agree that random 4K read/write numbers tell you the most about snappy performance. But the question is what's the point of sequential numbers and why do we still use them?

If we care about sequentials at all, it MUST be non-compressible, no? Because we're talking about large file sizes, then they must be media related in 99% of the cases: video and music.

So, if we're going to talk about sequential numbers in the future, shouldn't they always be for non-compressible speeds?

Here are my made up numbers, but it seems that out of all SSD buyers:
80% base purchase on random 4K numbers
20% base purchase on sequential numbers.

Of this 20%, It seems 80% would care about non-compressible (video and music) numbers.​

TL;DR - Nobody gives a crap about compressible sequential read/write so stop using those numbers. Therefore, anybody spending major time doing video and music shouldn't buy a SandForce. Thoughts?

 

Before SF, we don't need to care if it is compressible or not as all drive behave the same. It is SF that is the exception and only exception.

 

Since we are talking about read/write, here is a question i would like to know:

Is it because the files are so small that we don`t get any higher 4K read/write than 80-90 MB/s? If so, then how are they able to measure 90 MB/s with so small files like 4K? Are they hitting the drive with 1000s of 4K files simultainously? I see that with Vertex 3 you can get 300 MB/s with 512K files, why can`t they measure this high speed with lets say 1000 x 128 x 4K files? If so, then why do they stop with hitting the drive with 4K files when they reach 90 MB/s?

 

The reason we shouldn't filter out just SF is because at some point, SF could get bought or it might license its tech to other companies. What we care about is the random 4K and sequential non-compressibles. And besides, the fact that SandForce is so amazingly good at small, compressible data is still a huge advantage. I think if we start posting benchmarks without those stupid compressible + sequential numbers, people will get a clue.

I'd also like to know.

 

I think it's good to understand where the obsession on 4K performance comes from. Once upon a time there was a SSD called OCZ Core. It used a Jmicron controller. Everyone was running their normal benchmarks and these drives looked great. Until many users started experiencing stuttering, heavy stuttering. It was then found out that there was a way to predict whether a drive would stutter or not. It was the 4K performance. So 4K performance became very important. It was a way to tell the good drives from the bad drives. But now we are years later. No modern drive stutters any more. But people are still focused on 4K performance.

I don't 4K performance isn't as important as people make it to be. For example: Intel 510 has lower 4K performance than Sandforce 1200, yet the Intel is faster irl. Same story for Samsung 470 and Kingston V+: Not so good 4K performance, good real world performance.

Anand said something to the same extend in his Intel 510 review. 4K performance wasn't a good predictor of the real world performance.

Because everything matters. Latency, I/O performance, sequential performance, random 4K, 5K, 8K etc. performance.

Keep in mind that any user, also the ones that work with video and music, use their SSD in the first place for windows and applications. They are 50% compressible on average. This is the reason Sandforce 1x00 and 2x00 drives tend to do well in real world benchmarks.

Results with compressible data are important too. Sequential numbers are important too. Everything is important.

I suggest posting 1 ATTO result and 1 CDM result with random data. Real world performance is usually somewhere in the middle.

 

I didn't suggest that we should filter out SF, just that the once pretty good ATTO becomes meaningless on SF due to its unique way in handling all 0s or 1s.

Replacing with totally random data doesn't do any good either as that is not the real life data pattern as well.

I would say create a C: with a windows installation and a D: which can hold 30G then copy the whole C: to D: may be able to mimic the real life data. If sequential speed is one wants to measure on true data.

 

I believe I once saw a CrystalDiskmark run with partially compressible data. I'm not sure how to achieve it but that could be a good way too.

 

Were you talking about this quote in his Intel 510 review:

"In a horrible bout of irony Intel fixed its sequential performance and moved backwards in the random department. Random read performance, as it turns out, has a pretty major impact in the real world."
​

Agreed. Even compressible sequential numbers matter. However, we're in a world with too much information. I want the numbers that matter most; less of the numbers that matter less.

Agreed. That's why I said pretty much the same thing a couple of posts above:

"... we shouldn't filter out just SF ... And besides, the fact that SandForce is so amazingly good at small, compressible data is still a huge advantage."
​

If you love everyone, you love no one. If everything is important, nothing is important. Maybe you meant to say that everything matters. And I've already addressed that point.

 

I agree that Intel might perform really well in "real life" situations. And I think the overall point is that benchmarks don't tell the whole story.

I completely agree with that.

The problem is that there's too much information out there. I want to get right to the meat of the issue. Just tell me what I freaking need. Just OS, Facebook and YouTube? You need higher 4K randoms. Do video and photo editing with large file sizes? You need higher non-compressible sequentials. Do a bit of both? Then get numbers in the middle.

All I'm saying is that the numbers that matter the most, are being obfuscated amongst data that matters little. People less technically inclined don't know what to buy because SF puts out these stupid sequential numbers. Don't get me wrong, I have a Vertex and I love it. I still want a Vertex 3. But I know what I'm looking for and I know how to interpret the data. That doesn't change the fact that these ppl are using stupid measurements. And maybe I expect the manufacturers to use stupid numbers because they are in the business of selling. I would just like it if we, the enthusiasts, stop using them. Then, maybe the reviewers and review sites will stop using them.

It's like deciding Playboy Playmates based on SAT scores. Does it matter how smart they are? Of course. Does it REALLY, REALLY matter? No.

 

(reply to earlier post) Unfortunately there is no 'one number catches all'. In the absence of one, I think it's best to go by real world benchmarks.

PC Mark is coming out with a Windows 7 suite. If anything comes close to a 'one number catches all' it will be the HDD test in that suite.

This is the quote from Anand I meant:

 

Of course no number catches it all. But some numbers are better than others. Compressible sequential numbers are relevant, but least useful out of all of them. I think we're agreed now?

And I think you'd also concede that random 4K read is what we do the most and also why Anand would prefer to own a Vertex 3.

 

If I were to choose between Intel 510 or Crucial C300, I'd choose Intel 510 even though the Crucial has higher 4K random reads and writes. I assume Anand would choose the same.

I don't think Anand prefers the Vertex 3 because of the 4K random performance. He prefers it because it's the fastest in real life.

 

Many hours since i have posted this. Nobody who knows

 

Fastest in benchmarks maybe but Intel 510 and Vertex 3 are switching places in real life tests although Vertex 3 is much faster with Seq Read and faster with 4K R/W. And the differences between them are so marginal i would pick reliability/price/reputation of the brand before speed

 

they're writing gigabytes to the disk, in one big file. just the pattern in which they write changes. sequencially, or randomly spread 4kbyte junks (or 512kbyte junks) with random generated data. so it's really the ssd not being able to handle more than 22.5 thousand random 4kbyte junks randomly placed per second (asuming 90MB/s write speed))

 

aha ok. So they test it to the limit, and 90MB/s is the fastest you will ever have with 4K read/write on a vertex 3. That leads me to a different question:
If 90 MB/s is the limit, most operations would require a lot less 4K blocks right? If so, then that score you get from CrystalDiskMark is a little deceiving because it is all about the fastest "start" or the fastest momentum that counts since it is 100 x 4K blocks, not 100000 x 4K blocks (limit)? A burst operation, not a long marathon where a SSD can reach its highest 4K speed. If you understand what i mean...

 

As a noob in means of SSDs and one who is on the verge of deciding which one to get (Vertex 3 or Intel 510 or M4/C400?) I would like to know how much one drive being a few miliseconds faster than another really matters?

Here are three Youtube videos displaying how wonderfully fast all of these SSDs are:

Intel X25-M video
OCZ Vertex
Corsair F80

All of them seem to be performing great in my view...

I am considering getting an Intel 510 mainly because of its reliability. What I would like to know is how would its relative slowness in 4K read and write speeds affect its performance? Would it perform like the ones above? Boot up and opening apps are based on 4K read/write speeds, no?

I will install this drive (240 gb) in my Dell XPS L502X. I will mainly use the notebook for websurfing; typing in MS Word and other office apps; watching movies and listening to music; downloading films and music via peer-to-peer networks like emule and bit torrent, but more than that I will be using a download accelerator for downloading stuff via direct downloads (http). I will also be using music format conversion (eg. wav to flac or flac to mp3) and ID-tag editing software. After conversion I will be moving multiple music album folders (each about 300 mb) and movies (varying in size from 1.5 to 4.5 GB) to external hard drives via usb 3.0 connection.

Do you think there would be noticeable speed and performance difference between Intel 510 and Vertex 3 on a notebook used the way described above?

I feel like people are too much absorbed by numbers representing speed differences of miliseconds, am I wrong?

PS: I would consider purchasing an M4/C400, too, as they seem to have performed very well on tests, but I don't want to wait for them to become available for purchase...

 

Can't say you're right or wrong on this. I recall he said C300 is the best in a TRIM enabled OS for people with lighter workloads. I know he's a Mac user, at least for notebooks, so it's not going in the MacBook. He probably needs to do a lot of testing, so it might not go in his desktop. But, knowing what he knows and if he were recommending a drive to his impatient grandmother that doesn't do a lot of computing, but hates slow speeds, he might recommend the C300 over the Intel.

That's a good point. But I'm not saying that 4K random and fast in real life are mutually exclusive. All I'm really saying (and I admit the "pointless" in the topic title is a bit heavy) is that some measurements are a lot more important than others. Compressible sequentials are the first things that people report on. Everyone does it, even Anand. Just saying that we should place random 4K and non-compressible sequentials at the top. That isn't so bad, right?

Regarding the perfect test bed, I'm pretty sure I know what it is, but nobody will do it. But the best way would be to have a user created database for SSDs the way they have for 3DMark. People would test their applications, file transfers, etc. Each little test would get a score.

Then, when you and I go to shop for an SSD, we can input what % we spend with Photoshop, Flash sites, Facebook, etc. The website gives us a weighted score depending on what we do on the computer and the best SSD for that environment. It's perfect because the test will be tailored to each person. The only problem is that nobody will build this thing.

 

Start by reading this review of 510, M4, C300 and Vertex 3
Crucial m4 256GB SSD (C400) Review - Introduction

 

^ that's the best advice.

Turns out there's only 4 seconds difference between Intel 510 and Vertex 3 over all benchmarks (if Tiller did the maths correctly).

So it doesn't really matter much which one you pick.

It would be nice if manufacturers from now would give:
compressible sequentials
incompressible sequentials
4K random read and write performance

I don't really care about the order.

 

It looks like Intel 510 is 4 times slower than M4 and Vertex 3 in 4k read and write times which means less responsive OS, right? But in a few messages back in this thread it was mentioned that 90 mb/s is the upper limit for 4k read and write speeds if I have not understood it incorrectly. So Intel 510 shouldn't be noticably less responsive than others since its 4k read and write speeds are 80 and 50 mb/s respectively. Am I incorect?

How much less responsiveness are we talking about here? Is it measured in miliseconds, or would it be noticably less responsive?

 

I could live with that.

One step better would be to say what files are included.
You could transfer an Outlook.pst file for compressible seq.
You could transfer .avi for non-compressible.

Then people can actually decide what speed/measurement is relevant for them.
I would love to see enthusiasts start with this though.
Manufacturers will post numbers that help their sales.
It's better if we force them to change by changing the market itself.

 

No. During normal usage Intel 510 is just as responsive as those other drives.

Even during very heavy multi tasking the Intel is nearly as responsive as the Vertex 3.

 

The reason why Vertex 3 gets 285MB/s read and 245MB/s write there instead of 35MB/s read and 95MB/s write like i mentioned is because they tested the SSD with a Queue Depth of 32 instead of Queue Depth of 1. The more Queue Depth, the more workload it should represent i believe. Example: Queue Depth of 1 represent 10 000 4K blocks and Queue Depth of 32 represent 320 000 4K blocks. It is to visualize how drives perform under heavy heavy load i think.

So i think this really answered my own question lol. The reason why CrystalDiskMark don`t show any faster speed than 35/95 MB/s with 4K is because it is a set of 4K bytes. Not an infinite number of 4K blocks until the drive cannot get any higher speed.

 

And could the reason why Intel 510 is only slightly slower than Vertex 3 with real life tests be because of this although Vertex 3 is much faster in Seq Read/Write and 4K?

 

What does this difference translate to in real world use? Also, why are the much lower intel writes not more of a disadvantage to Intel?

 

Nr.1 and Nr.2 are Vertex 3 and Intel. 12.7 in 200 points difference is about 6%.

 

doesn't matter. it will do the 100 x 4K blocks in 1/1000th the time of the 100000 x 4K blocks. there's no real burst situation on an ssd.

 

deleted due to duplication

 

I guess the question is what Queue depth corresponds to regular usage or put a different way, at what point does Random read/write performance become 'good enough'?

 

At the end of the Intel 510 review, Anand said we've hit the limits of utility in random write speeds. So, random read is still king and everyone has a long way to go on that.

 

as most computing algorithms that are used today (the complex ones, not the simple ones. those an ssd can handle easily) are gathering-based, instead of scattering based, they're more random-read dependent than random-write dependent indeed. if that changes for something important (a new video codec or what ever), that might change the limit for random write speed needs again.

 

That doesn`t make sense. If it does it 1000th of the time it must have the same speed. CrystalDiskMark says you get much greater speed with more queue

 

no. more queues does not mean more data. it means more data in PARALLEL. the intel ssd has 10 queues => it can perform 10x as fast when reading / writing independent data, compared to just one stream of data.

 

OK but then why did you say that it finished 100 x 4K blocks 1000 times faster than 100 000 x 4K blocks (just an example) if it is paralell and it perform the same speed on all (since with QD=10 = 10x faster speed)?

lol i forgot what my original question was about. Damn this confusion

 

I didn't understand you question from the very beginning.

 

lol.

The best part was when Cloudfire quoted himself. I didn't want to interrupt because it seemed like he was about to find out some good stuff from himself.

 

lol that`s allright. Will have to think this through a little more and get back to you. I did have a question in there somewhere but it got lost

 

Am I misinterpreting these graphs or do the graph you posted and the one I posted contradict each other?

I'm about to pull the trigger on the Intel 510 SSD and I don't want to regret it later, so I'll ask as noobly as the noob I am: Would I notice any difference in OS and apps performance (or anything else) if I were to use Vertex 3 as my sole drive for a while and then switch to Intel 510? I mean are the differences between the two negligible?

 

Differences are negligible.

The graph I posted is a real world benchmark, it measures real performance. 4K random performance is a synthetic benchmark that's not necessarily relate to real world performance.

 

Ok, thanks a lot!

One last noobness:

This is the one to get, right? There's no other version, nothing else I need to know (about the model code I mean)?

Intel 510 Series (Elm Crest) SSDSC2MH250A2K5 2.5" 250GB SATA III MLC Internal Solid State Drive (SSD)

 

That's the one.

Ps. Vertex 3 is also in stock but cheaper.

 

Well Vertex 3 looks tempting, but I live in Turkey, and I will have the item delivered to a friend of mine in USA and then he will bring it to me. So I have to get a more reliable SSD, because returning the item would be such a hassle!

Crucial M4 is another one one my list but it's not available yet and I don't know if I waited for it to become available, it will be justifiably cheaper? Is there an estimated price for the 250 GB M4?

 

Well install times are basically sequential writes...and this is why Intel 510 or the M4 are well rated. However a typical usage seems to be about 50% operations being small random writes and about 7% random reads.

"Using Microsofts Diskmon, he simply monitored his typical computer usage in doing things such as using the internet, running applications, playing music etc. In short, he did his best to recreate the computer use of a typical user and then used the program to break down the percentage that specific disk access speeds were being utilized.

In the end, it confirms something we always thought but just didn’t really understand. Large sequential read and write access is utilized by the average user less than 1% of the time yet the most used method of access is smaller random write access as shown by the 8k write at over 50%."

source: The SSD Manufacturers Bluff | The SSD Review

I would call the latter a much better real world test...

 

The install times is only one page of the review. There are several pages with real world benchmarks.

The idea that random performance is the best indicator of real world performance is outdated. We recently had a whole thread about it. I'll try to find it.

And by the way, install times are more dependent on random write performance.

90% of those jobs you're not waiting for them to finish. In my opinion real world benchmarks are much more interesting when you time the jobs that the user has to wait for to finish.

Edit: I tried to find the thread where we had this discussion before but I can't. Can anyone provide the link? It's called something like "4K random performance".

 

Honestly who cares if it takes 5 seconds more to install Office?? This is something you do only a few times. I find it difficult to understand that you prefer optimizing the things you'll do 1% of the time rather than optimizing those you do 50% of the time... Ever heard about cumulative performances? A tiny gain of 0.1 sec on something you'll do 10000 times a day (like reading or writing 4K random blocks) will give you a total gain of close to 15 minutes while your single Office install (per day) and a few large file transfers would barely bring more than a few seconds.... It's up to you to decide what is significant in real usage. The ssdreviews article quoted above has been updated in March 2011. I wouldn't call this as being outdated On specific applications sequential operations are the primary interest but this is clearly not my focus nor the one of most people I guess.

 

Oh man you keep going on about install times. The review is 13 pages total, load times, video editing, file copies, the works.

If random performance is really as important as you and Les claim, please explain my why the Intel 510 does so well in every real world scenario you throw at it, including the Anandtech Storage benchmarks. Everyone knows the Intel 510 has mediocre random performance. Yet it's real world performance is very good. So that's proof. The idea that random performance is the best indicator of real world performance isn't true.

Real world performance is influenced by many things. Random performance is important but so are sequential performance and access times.

Can anyone please provide the link where we had this discussion before? That way we don't have to go through the same discussion again and again.

Edit: found it, here you go:
http://forum.notebookreview.com/sol...-sequential-read-write-numbers-pointless.html

Please use that thread if you want to dicuss this further.

 

Because your real world scenarios are not mine You seem interested in read and write operations on medium to large files and there yes you're right the Intel 510 wins (at least over the C300). However as mentionned in the ssdreviews article, if you want an overall snappy system you also need to load those tiny DLLs fast when launching programs, reading small temporary files etc is what makes your system more responsive and thus usable. I'm not talking about boot times here because here again I don't care if windows boots in 14 or 15 seconds. I mean what's more real-world than a guy monitoring his own typical usage (shared by 90% of people probably...) and see what matters more?

Don't worry I did not only read the install time section of these reviews. The other parts are interesting as well and some specifically mention one thing: random performances with 3rd gen drives are down...and if this wasn't a concern it wouldn't even be mentionned.

Don't take me wrong: I'm not saying the 510 is a poor drive and I'd choose it for compatibility with sandy bridge (that's not to say reliability since we do not know that on the 510) since it pisses me off to buy a 200+$ drive and have to apply registry hacks and deactivate power management features to have the drive working (dixit Crucial...). The thing is that it comes at a significant price premium over last generations drives without beating them up significantly except in sequential operations...which are important to you but not so much to me. That's just because we do not focus on the same aspect of what "useful" performance means.

EDIT: sorry didn't see your link earlier. I'll stop poluting this thread then

 

During that kind of usage, no one in the world will be able to distinguish between Intel 510, Vertex 3 and Crucial M4. Even though their random performance is different.

I wasn't talking about boot times. I was talking about Anand's Storage Bench.

Yet real world performance is up. Go figure.

Anand said it very well: It looks like we may have hit the upper limit of what we need from 4KB random write performance (at least given current workloads).

I moved the discussion to the appropriate thread. Please continue to educate us why random performance is so important

I'm curious Kilou, how many SATA III drives have you personally used during your own work patterns?

 

I do not pretend to educate anyone here, but I sure want to remind that we do not all share your typical usage of an SSD

As you correctly said not only random or sequential performance are important. The overall balance is important and many different things do count. I'm just trying to argue that 3rd gen drives seem to favour sequential over random. Since you seem in love with Anand, I repeat what they say on Intel 510 review (and what has already been said in this thread, page 1):

"In a horrible bout of irony Intel fixed its sequential performance and moved backwards in the random department. Random read performance, as it turns out, has a pretty major impact in the real world."
source: AnandTech - The Intel SSD 510 Review
This is what Anand says, not me...

Now since you still love real world performances besides Anand, check these PCMarkVantage scores from the link you provided us: AnandTech - The 2011 Mid-Range SSD Roundup: 120GB Agility 3, Intel 510 and More Compared.

Yes you're right Intel 510 comes on the top but it's the slowest of the 3rd gen drives in the comparison! The C300 (which cost a lot less) is right behind and is especially ahead in many of the subtests of PCMarkVantage.

Last but not least, a pitty the M4 was not in the comparison because in the 256Gb settings, it beats the 510 (see http://www.anandtech.com/show/4253/the-crucial-m4-micron-c400-ssd-review/8 )! Even the C300 is ahead in 256Gb... And what is mostly different between the 510 and the M4 or C300? Guess what? Random performances!

Want any other "education"? Double check what you read (no offense)