2017 iMac SSD Slower than 2015 for small files

[Ylan]

According to Mac Performance Guide the new 2017 iMac's SSDs are slower than the 2015 models when writing small files.

Does anyone know how this could be? It's only for smaller files, on the whole they are a lot faster.

As i read this, I thought about two possible reasons:

1. Apple has made different compromises and prioritized bigger files.
2. The new drives are optimized for APFS coming in High Sierra.

I am however not an engineer, but am truly curious.

Humbly Ylan

 

[nambuccaheadsau]

Can depend on the size of the SSD. Smaller SSDs, such as a 256GB model, will be slower than a larger capacity SSD.

 

[bplein]

According to Mac Performance Guide the new 2017 iMac's SSDs are slower than the 2015 models when writing small files.

Does anyone know how this could be? It's only for smaller files, on the whole they are a lot faster.

As i read this, I thought about two possible reasons:

1. Apple has made different compromises and prioritized bigger files.
2. The new drives are optimized for APFS coming in High Sierra.

I am however not an engineer, but am truly curious.

Humbly Ylan

I am going to generalize, I have no architectural knowledge of the specific NAND chips used in the PCIe/NVMe SSDs in use on the 2015 vs 2017 Macs.

Newer NAND is optimized for capacity and price. That means denser chips, and that means fewer chips to hit the same capacity point. Denser chips also results in larger Erase Blocks. All of these things getting larger can have a detrimental effect on smaller I/Os.

Also it appears that large block I/O is getting better. In other words, it's getting better for the files of today: rich media (video and audio), which will generally drive larger block I/O and the 2017 will end up OUTPERFORMING the 2015 for these files.

What kind of small file high I/O work does one do on an iMac? Software compiling is the only thing that comes to mind.
[doublepost=1500045560][/doublepost]By the way, so I'm basically agreeing with #1, that it's optimized for the type of media that the majority of users will be using when they need higher performance I/O.

 

[cynics]

Different NAND most likely. Several things to consider though.

While there are specific work loads that you would be able to see a difference in performance for the most part we are looking through a magnifying glass at that chart. 100% difference seems like a lot, but if I said 100% slower than .05 seconds (example) would you really care? That maybe a bit of an extreme example but copy a dozen 1mb files on even a HDD is fairly quick.

Read speed is still the most important. This is where we see performance gains like boot times, programs loading, etc etc. Write is very important however it can often be limited by other factors like USB while importing data, GPU/CPU while exporting data from a program, etc. Its not often people run into bottlenecks due to PCIe SSD write speeds, although they do exist.

APFS is coming. With cloning, general file copying, regardless of size is just cloned near instantly. Further limiting user exposure to slower write speeds in real world use. Although many people say modern OSs are optimized for SSDs that not really the case, they are just faster and new OSs are fatter. With APFS there is some actually SSD optimization, even for FD owners with smarter use of the SSD vs HDD.

I have a Sandisk SSD Plus and a Samsung EVO 840 and the 4k file random reads is supposedly 50% faster with the Samsung. For the life of me I can't see/feel a real world difference.

Point is, even if that chart is accurate I wouldn't concern yourself too much over it. I could be doing the math wrong in my head so please forgive me, but using the low spot (dip) in that chart, 250MB size files at 750MiB (just going to round and say 750MB) = 3000 files in one second written.

 

[Fester1952]

Then that article goes on to say,
"Sustained transfer speed, large transfers
Both machines utilized the Apple 1TB SSD.

The 2017 iMac 5K delivers an outstanding performance that is 40% faster for reads and 38% faster for writes"

 

[cynics]

Then that article goes on to say,
"Sustained transfer speed, large transfers
Both machines utilized the Apple 1TB SSD.

The 2017 iMac 5K delivers an outstanding performance that is 40% faster for reads and 38% faster for writes"

We are referencing typical write behavior which is not "large transfers" (with exceptions obviously).

Your quote continues btw.

"The 2017 iMac 5K delivers an outstanding performance that is 40% faster for reads and 38% faster for writes. Few applications are even capable of taking advantage of this superlative performance."

 

[bplein]

We are referencing typical write behavior which is not "large transfers" (with exceptions obviously)."

First off, you can't say that everyone is typical.

OK, I do a little video editing and RAW photo editing on my iMac. But I'm not working in it 24x7 that way, so the rest is browsing, some gaming, email (web). Lot's of windows open dealing with other systems (VMware, FreeBSD, Linux) in my home lab, so that's not much file usage.

So my "typical" I/O is going to be small file.

But my important I/O (video and RAW photo editing) is all large transfer. That's the only time I am concerned about performance, because I don't copy/write large amounts of small files where wall clock time (the time to complete a process) is important.

I think this statement about large file transfers being untypical is not backed by facts, nor is it the right measurement. The measurement of performance should be taken for the types of I/Os that are slowing you down, not for the idle process that happen in the background and aren't slowing you down.

For people who process large files such as RAW photos and Video, large file transfers are the norm (although the metadata databases used by tools such as Lightroom may also do lots of small file I/O).

For people building software, they are going to have to profile the file access during their builds to see what the mix of small and large block will be, and estimate to the best of their ability if they need larger or smaller I/O.

 

[tipoo]

Apple seems to prioritize large file transfer speeds over 4kb random read/write for its SSD, in all of its products. If you look on NotebookReview for any of their modern macs, max transfer speeds are stunning, while random 4KB is often beaten by much cheaper competitors.

I trust that Apple knows what it's doing, perhaps macOS simply avoids 4KB worst case performance by reading and writing in batches (coalescing), it already defragments any file under 20MB, etc etc.

The speed speaks for itself in any competitive testing outside of these low level benchmarks, so I'd not worry about it.

 

[cynics]

First off, you can't say that everyone is typical.

OK, I do a little video editing and RAW photo editing on my iMac. But I'm not working in it 24x7 that way, so the rest is browsing, some gaming, email (web). Lot's of windows open dealing with other systems (VMware, FreeBSD, Linux) in my home lab, so that's not much file usage.

So my "typical" I/O is going to be small file.

But my important I/O (video and RAW photo editing) is all large transfer. That's the only time I am concerned about performance, because I don't copy/write large amounts of small files where wall clock time (the time to complete a process) is important.

I think this statement about large file transfers being untypical is not backed by facts, nor is it the right measurement. The measurement of performance should be taken for the types of I/Os that are slowing you down, not for the idle process that happen in the background and aren't slowing you down.

For people who process large files such as RAW photos and Video, large file transfers are the norm (although the metadata databases used by tools such as Lightroom may also do lots of small file I/O).

For people building software, they are going to have to profile the file access during their builds to see what the mix of small and large block will be, and estimate to the best of their ability if they need larger or smaller I/O.

I'm not sure why you think I feel the opposite, you even quoted me saying "with exceptions obviously". I should have chosen my words there more carefully I guess.

But regardless, everyone is typical.

I deal with large files near constantly. Dropping media files into video editors, moving around those files for safe keeping, moving around the export files etc. Even outside of work my main use of my iMac is a media server, so I'm constantly dealing with 30gb MKVs, and/or the 5-10gb output of Handbrake.

However everything up to that point was based on small file transfers. From boot to using the video editing software. All the performance realized from an SSD in real world use is based around random small files. This is why the 4k test are important in storage benchmarks (4k is the file size).

I know when I move a large file there will be delay, I'm prepared for it and accustomed to it. However slowness in overall system performance I prefer to avoid wherever possible.

While someone like me could benefit from faster large file size transfers, small are just as if not more important.

But please reference my first post of why I'm not personally concerned around it. Soon large same drive file transfers won't matter to even an HDD with APFS cloning. However small file read like from opening a program won't benefit from that, that is where we will want the performance. Again though....its not like its "slow" by any means.

 

[looking4anotebo]

Well, we know they took $300 off the price of the computer..... you don't think they took that off their margins, do you?

We know they switched to cheaper speakers for sure. Possibly cut back in other areas too.

 

[cynics]

Well, we know they took $300 off the price of the computer..... you don't think they took that off their margins, do you?

We know they switched to cheaper speakers for sure. Possibly cut back in other areas too.

I've seen users report this but is there any real testing? If so do you have a link?

People said the same about the 2015 iMac. By the time I get one there will be a tin can with a string coming out of the Mac.

 

[looking4anotebo]

I've seen users report this but is there any real testing? If so do you have a link?

People said the same about the 2015 iMac. By the time I get one there will be a tin can with a string coming out of the Mac.

I have the 2015 and spent some time in the store playing with the 2017's and I think it's fairly obvious.

 

[cynics]

I have the 2015 and spent some time in the store playing with the 2017's and I think it's fairly obvious.

Do you feel that is a fair comparison though? I haven't heard one but I know I the acoustics in my local Apple store isn't a good point of comparison for me personally. Be a shame if true because I was rather impressed with the sound quality of my 2013 given its size.

 

[Ylan]

Well, we know they took $300 off the price of the computer..... you don't think they took that off their margins, do you?

We know they switched to cheaper speakers for sure. Possibly cut back in other areas too.

This is news for me, do you have a source?

Humbly, Ylan

 

[SaSaSushi]

This is news for me, do you have a source?

I'm guessing that the source might be his behind on this one.

 

[looking4anotebo]

This is news for me, do you have a source?

Humbly, Ylan

source would be common sense.
[doublepost=1500718928][/doublepost]

Do you feel that is a fair comparison though? I haven't heard one but I know I the acoustics in my local Apple store isn't a good point of comparison for me personally. Be a shame if true because I was rather impressed with the sound quality of my 2013 given its size.

good point but it was in a dead best buy.

 

[SaSaSushi]

We know they switched to cheaper speakers for sure. Possibly cut back in other areas too.

source would be common sense.

Wow. Case closed.

I realize it's a total waste of time but you wouldn't happen to have any other sources for this bombshell, would you? Maybe some factually-based ones?

 

[joema2]

According to Mac Performance Guide the new 2017 iMac's SSDs are slower than the 2015 models when writing small files....Does anyone know how this could be? It's only for smaller files, on the whole they are a lot faster....

I have both 1TB 2015 and 2TB 2017 SSD iMac 27s and I don't see any real-world difference in SSD speed. Testing with QuickBench shows the 2015 is a little faster at some small transfer sizes but this varies based on whether it's sequential vs random, read vs write, and what exact size. See these tests:

https://joema.smugmug.com/Computers/2015-vs-2017-iMac-27-SSD-performance/n-vJNW6N/

Using the command-line dtrace utility, it's possible to make an I/O distribution histogram which monitors the size of all I/Os. Doing this during importing and transcoding media within FCPX, the most common I/O size is 64k, and the next most common size is 1 megabyte. However these operations are CPU-bound, not I/O-bound, so in the real world the 2017 iMac 27 is significantly faster.

When scrubbing forward and backward on the timeline in FCPX, the most common I/O size is 1 megabyte, and the next most common size is 512k bytes.

 

[looking4anotebo]

Wow. Case closed.

I realize it's a total waste of time but you wouldn't happen to have any other sources for this bombshell, would you? Maybe some factually-based ones?

Looks like they are possibly using runt 7700ks as well.

 

[SaSaSushi]

Looks like they are possibly using runt 7700ks as well.

Still waiting for the proof that Apple has switched to cheaper speakers.
[doublepost=1500765485][/doublepost]

Looks like they are possibly using runt 7700ks as well.

Yeah, this was disproven in another thread where someone posted this guy's video. His benchmarks are no faster than average.

Listen, if I had the 2015 model like you then I'd want to believe all this crap too.

 

[looking4anotebo]

Still waiting for the proof that Apple has switched to cheaper speakers.
[doublepost=1500765485][/doublepost]

Yeah, this was disproven in another thread where someone posted this guy's video. His benchmarks are no faster than average.

Listen, if I had the 2015 model like you then I'd want to believe all this crap too.

Sure thing, buddy. I'll just let you go on your way thinking Apple cut $300 off their margin per computer.

 

[SaSaSushi]

Sure thing, buddy. I'll just let you go on your way thinking Apple cut $300 off their margin per computer.

If I had the 2015 and then Apple released the 2017 at $300 less, then I would really want to believe all this crap.

It's called a fluctuating market. Component prices change. The prices are what they are. Apple's margins are simple speculation on your part.

 

[cynics]

Looks like they are possibly using runt 7700ks as well.

That video is deceiving, probably not intentionally (I like snazzy labs) but just from lack of better understanding of Geekbench. He used the generic geekbench score found on their webpage.

However if we look at the actual scores you'd see this....

Have to go a couple of pages in to find a score as low as his.

Since its a synthetic benchmark and you can run it 20 times and get 20 different scores, therefore there are scores much lower than his.

Point is you are going off bad information. Kaby Lake i7 CPUs, especially unlocked CPUs (K series) don't have runts. The runts have features turned off, cores disabled etc then rebadged and moved down the chain to become other models.

 

[2010lexus]

Sure thing, buddy. I'll just let you go on your way thinking Apple cut $300 off their margin per computer.

Just because the computer is cost less does not mean apples margins per Computer is any less. They could be getting the same parts or even better parts that cost less.
As far as comparing sound of the speakers. I don't care if you are at Best Buy or the Apple Store the acoustics are different then your home.
I have 2 iMacs in my home one 2015 the other is a 2017 and I can say they sound the same.

 

[Marx55]

I have both 1TB 2015 and 2TB 2017 SSD iMac 27s and I don't see any real-world difference in SSD speed. Testing with QuickBench shows the 2015 is a little faster at some small transfer sizes but this varies based on whether it's sequential vs random, read vs write, and what exact size. See these tests:

https://joema.smugmug.com/Computers/2015-vs-2017-iMac-27-SSD-performance/n-vJNW6N/

Many thanks for the benchmarks. That is 2,718 MB/s read & 2,126 MB/s write for iMac 2017. Apple should have chosen (or at least allow user upgrades) a much better and cheaper disk like:

Samsung 960 PRO Series - 2TB PCIe NVMe - M.2 Internal SSD (MZ-V6P2T0BW)
Sequential R/W (read/write) speeds up to 3,500/2,100 MB/s and random R/W speeds up to 440/360K IOPS, respectively.
http://www.samsung.com/semiconductor/minisite/ssd/product/consumer/960pro.html
https://www.amazon.com/Samsung-960-PRO-Internal-MZ-V6P512BW/dp/B01LY3Y9PH