Re: cryptsetup benchmark vs openssl speed, huge difference

Hi John,

On 24/10/2023 07:10, John Smith wrote:
> Is there a reason why the AES-XTS speed is vastly inferior to openssl?

cryptsetup benchmark does not measure performance of openssl (or other 
configured userspace crypto backend). For detailed answer see below.

> 
> cryptsetup benchmark
> aes-xts        512b      2503.5 MiB/s      2514.3 MiB/s
> 
> openssl speed -evp aes-256-xts
> type             16 bytes    64 bytes     256 bytes    1024 bytes
> 8192 bytes   16384 bytes
> AES-256-XTS      633950.11k  2343011.04k  4795982.60k  6688253.60k
> 7698595.84k  7806526.67k
> 
> This is also borne out by the NVMe performance, or lack thereof.
> 

couple remarks to better establish cryptsetup benchmark baselines:

1) The benchmark does not perform any I/O on storage device. There's no 
correlation with block device installed in the system.

2) All block cipher results (aes, serpent, ...) are collected measuring 
kernel crypto layer performance (accessed via netlink API from). It 
depends what implementation of the cipher kernel crypto layer loads 
(based on system arch, CPU installed, kernel modules compiled-in etc.).
There's ab overhead for using kernel crypto layer via netlink API.

It does not reflect real performance of dm-crypt mapped over available 
block device in the system, by far. For example: dm-crypt encrypts data 
in 512B to 4KiB sized blocks based on dm-crypt table parameters 
(sector_size). The cryptsetup benchmark on the other hand measures 
performance by encrypting much larger data chunks (1MiB IIRC).

O.