Hacker News Reader

Comment by Osiris | Hacker News Reader

Another reason is that read/write speeds on APFS encrypted volumes are 30-40% slower than non-encrypted volumes[1].

[1] https://news.ycombinator.com/item?id=15333036

As the speed test was using a beta of High Sierra / APFS, I performed some tests of my own. I've written /dev/zero to the disk, and read Xcode_9_GM_seed.xip from the disk using different block sizes using `dd`. I ran the read tests a few times, as there was a lot variation in the results. I've excluded the first result of the read results for each FS. The test was run on a Samsung Evo Pro 850 500GB SSD inside a Sharkoon USB-C case connected over USB-C to a MacBook Pro 2016 w/ Touchbar.

    Speed in MB/s
                HFS+    HFS+ Encrypted  APFS        APFS Encrypted
    1M write    498     473             490  (-2%)  444 (-6%)
    1M read     6896    6815            6860 (-1%)  6827 (0%)
    4k write    486     477             485  (-0%)  445 (-7%)
    4k read     2734    2709            2599 (-5%)  2545 (-6%)

While there is still some performance degradation between HFS+ and APFS, it's within a single digit now.

dd is not an acceptable benchmark. You need to eliminate the effects of buffering/caching to get accurate results.

/dev/zero is also not an acceptable source of data; since \0\0\0\0\etc is unnaturally easy to compress. A large real file (e.g. installation ISO) may have been a better idea.

Following the comments on the test method, I've run the 'Blackmagic Disk Speed Test'. The results are the following.

    Speed in MB/s
            HFS+    HFS+ Encrypted  APFS    APFS Encrypted
    read    513     513             508     506
    write   477     460             478     432

The earlier results of 7 GB/s are quite impossible with the Samsung Evo Pro 850 -- those max out around 500 MB/s. So my earlier test results are invalid. The new test results mostly show that both file systems can saturate the disk's raw throughput.

-- update

On my internal Apple SSD (PCI-Express) I've created a partition for testing. I've kept "Blackmagic Disk Speed Test" running for a few minutes and took an average of the tests. The results are in favor of APFS, beating HFS+ by 3-17% in throughput.

    Speed in MB/s
            HFS+    HFS+ Encrypted  APFS        APFS Encrypted
    read    2400    2240            2700 (+13%) 2400 (+7%)
    write   1900    1930            2000 (+5%)  1980 (+3%)

  Speed in MB/s
             HFS+     HFS+ Encrypted  APFS     APFS Encrypted
  1M WRITE   1375     1373            1372     933
  1M READ    2446     2340            2162     1304
  4K WRITE   852      797             502      378
  4K READ    2106     1486            2156     1001

https://news.ycombinator.com/item?id=15333754

That's really strange. Why would encryption affect APFS so badly but not HFS+?

Lots of reasons.

If crypto is part of the syscall path, then you definitely pay with latency. If you can push IO to a worker thread then you can release the user call faster. This also requires tuning the read-ahead cache more aggressively.

It could also be a change in crypto algorithm, unaligned page IO, or any other overlooked IO latency problems.

At a high level, I believe APFS encrypts files individually (and uniquely) instead of the whole drive. Have to do some more reading :)

Not instead of, but optionally in addition to (https://developer.apple.com/library/content/documentation/Fi...)

On Mac OS, it seems you only get full disk encryption, though (https://arstechnica.com/gadgets/2017/09/macos-10-13-high-sie...)

Appreciate the clarification. Both are better than just one.

This was high sierra beta though.

Just tested this. Was consistently getting 2000mb/s+ writes before, now getting ~1600mb/s. Reads are inconsistent, but most results are also 20-40% slower than before.

Just curious, at these speeds, what use case would really affect you, or people you can imagine?

30-40% in a stress test is significant. but if I'm not editing 4k videos perhaps, where would I notice this?

Even amongst us programmers who use horrible inefficient IDEs and make horrible inefficient electron apps, disk speed isn't really a noticeable bottleneck right now?

Even a convoluted local environment to simulate a server environment with a ton of microservices wouldn't really be affected much by this

Disk speed is always a bottleneck, much more than CPU generally. Almost every thing you do leads to, or is dependent on, IO speeds. If you disk speed drops by 40%, for example, every application will be at least 40% slower to load. That's a major hit. That horribly inefficient electron app is now slow to load, not just slow to run.

Or, to put this another way: Do you have an SSD? Why?

The filesystem is cached in RAM so it’s too simplistic to talk about applications like that.

yeah, not always true, and this is more about noticeable performance. The answer is to why I have an SSD is speed and throughput and the point of calling this file system change out is because there are diminishing returns to the productivity in real world use

a benchmark is just a repeated process which is not reflective or real world use except in very small narrowly tailored applications

Reply on news.ycombinator.com