On 03/12/2024 16:10, Bartosz Golaszewski wrote:
On Tue, 3 Dec 2024 14:53:21 +0100, neil.armstrong@xxxxxxxxxx said:
On 03/12/2024 10:19, Bartosz Golaszewski wrote:
From: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
Having switched to workqueue from tasklet, we are no longer limited to
atomic APIs and can now convert the spinlock to a mutex. This, along
with the conversion from tasklet to workqueue grants us ~15% improvement
in cryptsetup benchmarks for AES encryption.
Can you share on which platforms you did the tests and the results you got ?
Sure, I tested on sm8650 with the following results (they vary from
one run to other but are more or less in this range):
With this series:
# Algorithm | Key | Encryption | Decryption
aes-cbc 128b 94.1 MiB/s 138.6 MiB/s
serpent-cbc 128b N/A N/A
twofish-cbc 128b N/A N/A
aes-cbc 256b 94.8 MiB/s 128.5 MiB/s
serpent-cbc 256b N/A N/A
twofish-cbc 256b N/A N/A
aes-xts 256b 132.9 MiB/s 131.8 MiB/s
serpent-xts 256b N/A N/A
twofish-xts 256b N/A N/A
aes-xts 512b 122.6 MiB/s 122.4 MiB/s
serpent-xts 512b N/A N/A
twofish-xts 512b N/A N/A
Without it:
# Algorithm | Key | Encryption | Decryption
aes-cbc 128b 96.4 MiB/s 141.0 MiB/s
serpent-cbc 128b N/A N/A
twofish-cbc 128b N/A N/A
aes-cbc 256b 67.0 MiB/s 97.8 MiB/s
serpent-cbc 256b N/A N/A
twofish-cbc 256b N/A N/A
aes-xts 256b 131.7 MiB/s 132.0 MiB/s
serpent-xts 256b N/A N/A
twofish-xts 256b N/A N/A
aes-xts 512b 93.9 MiB/s 96.8 MiB/s
serpent-xts 512b N/A N/A
twofish-xts 512b N/A N/A
AES-CBC and AES-XTS with shorter keys remain pretty much the same. I'm not
sure why that is. I also tested on sa8775p but there are no visible
improvements there. :(
Thanks for the results !
Neil
Bart
