[dm-crypt] Latency penalty

Ivan Babrou ibobrik at gmail.com
Fri Sep 22 04:25:06 CEST 2017


Thank you for your substantive reply. My apologies for not including
benchmark output initially, I neglected this detail as not very
important given other numbers attached.

Here's what I have:

$ sudo cryptsetup benchmark
# Tests are approximate using memory only (no storage IO).
PBKDF2-sha1      1014096 iterations per second
PBKDF2-sha256     632434 iterations per second
PBKDF2-sha512     420776 iterations per second
PBKDF2-ripemd160  569878 iterations per second
PBKDF2-whirlpool  196804 iterations per second
#  Algorithm | Key |  Encryption |  Decryption
     aes-cbc   128b   485.8 MiB/s  1666.9 MiB/s
 serpent-cbc   128b           N/A           N/A
 twofish-cbc   128b           N/A           N/A
     aes-cbc   256b   359.1 MiB/s  1258.6 MiB/s
 serpent-cbc   256b           N/A           N/A
 twofish-cbc   256b           N/A           N/A
     aes-xts   256b  1402.3 MiB/s  1422.9 MiB/s
 serpent-xts   256b           N/A           N/A
 twofish-xts   256b           N/A           N/A
     aes-xts   512b  1073.8 MiB/s  1097.8 MiB/s
 serpent-xts   512b           N/A           N/A
 twofish-xts   512b           N/A           N/A

>From what I see here, I should be getting around 1073MB/s for reads.

In my first email I linked two graphs that show relation between block
size and linear read speed:

* No FS: https://i.imgur.com/yar1GSC.png
* XFS: https://i.imgur.com/OQk6kDo.png

In neither case I'm bottlenecked on CPU (hovering around 30% between
kworkers). How come I can't read at device speed with 1MB block size
but can with 1GB?

I also mentioned that I used in-memory loopback device as a base for
LUKS and was able to get up to 850MB/s reads, where I presumably hit
single CPU limit.

In addition we checked completely different hardware (latest top
macbook pro running stock xubuntu in a vm). There we've got 2700MB/s
from "cryptsetup benchmark", 812MB/s from a raw block device and
551MB/s from a LUKS device. These numbers do not add up for me.

More information about the dm-crypt mailing list