Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 21 Dec 2018 16:16:39 +0700
From:      Alexey Dokuchaev <danfe@nsu.ru>
Cc:        FreeBSD Current <current@freebsd.org>
Subject:   Re: AESNI, /dev/crypto, and new OpenSSL
Message-ID:  <20181221091639.GA53513@regency.nsu.ru>
In-Reply-To: <20181220181007.GA2374@regency.nsu.ru>
References:  <20181220173535.GA2505@regency.nsu.ru> <CAOjFWZ6WB4Wgy%2BAgib%2B5OqsPxKSSfK1=_CY8VycWuHpAKe=bcw@mail.gmail.com> <20181220181007.GA2374@regency.nsu.ru>

next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, Dec 21, 2018 at 01:10:07AM +0700, Alexey Dokuchaev wrote:
> On Thu, Dec 20, 2018 at 09:33:41AM -0800, Freddie Cash wrote:
> > On Thu, Dec 20, 2018 at 9:21 AM Alexey Dokuchaev <danfe@nsu.ru> wrote:
> > > Had something got broken here, or I'm misunderstanding how this machinery
> > > now works?
> > 
> > Start reading here:
> > 
> > https://lists.freebsd.org/pipermail/freebsd-stable/2018-December/090195.html
> > 
> > That thread covers this issue.  :)  Along with the "fix" for it.
> 
> Thanks for the pointer.  I've checked both -current and -hackers MLs prior
> to posting, but didn't expect this would show up on -stable first. :)

In case people find this thread and want quick answers without having to
deviate to -stable, here's a quick summary and my speed test results, with
some quotes from delphij@, jhb@, et al.:

1) aesni(4) and crypto[dev](4) modules are not required now for OpenSSL,
   and userland acceleration in general, to work;

2) On capable systems, AES-NI would be used automatically.  In fact, it's
   much faster to use the AES-NI instructions in userland than to use a
   system call that copies the data into a kernel buffer, uses the same
   AES-NI instructions, then copies the data back out again along with
   the overhead of a pair of user <--> kernel transitions.

  (Note from me: if you've observed very strange results when using -evp
   with aesni(4) + BSD cryptodev engine on OpenSSL 1.0.2, it was probably
   because of that user <--> kernel multicopying.)

Some quick naive benchmarks on AMD A8-5550M APU (results were trimmed for
brevity):

baseline: openssl speed -elapsed aes-128-cbc:

16 bytes    64 bytes    256 bytes   1024 bytes  8192 bytes  16384 bytes
35922.35k   39346.28k   40492.29k   94625.81k   95194.36k    95619.24k

hardware extensions: openssl speed -elapsed -evp aes-128-cbc

16 bytes    64 bytes    256 bytes   1024 bytes  8192 bytes  16384 bytes
133823.08k  186960.39k  226363.05k  238189.15k  241782.56k   241646.38k

AES-NI disabled: env OPENSSL_ia32cap="~0x200000000000000" openssl speed
                 -elapsed -evp aes-128-cbc:

16 bytes    64 bytes    256 bytes   1024 bytes  8192 bytes  16384 bytes
54820.92k   64884.98k   69229.02k   70424.31k   70731.22k    70714.02k

It's interesting how -evp run w/o AES-NI got capped at ~67 GB/s, while
the baseline had sustained at ~91 GB/s.  AES-NI run had reached pretty
solid ~230 GB/s.

./danfe



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20181221091639.GA53513>