[openssl-dev] [openssl-team] Discussion: design issue: async and -lpthread

[Alessandro Ghedini]

On Tue, Nov 24, 2015 at 07:56:15am +1000, Paul Dale wrote:
> Somewhat tangentially related to this is the how thread locking in OpenSSL is
> slowing things up.
> 
> We've been doing some connection establishment performance analysis recently
> and have discovered a lot of waiting on locks is occurring.  By far the worst
> culprit is CRYPTO_LOCK_EVP_PKEY in CRYPTO_add_lock calls.  Changing these to
> gcc's atomic add operations (__sync_add_and_fetch) improved things
> significantly:
> 
> base OpenSSL        11935 connections/s    85% CPU utilisation
> with Atomic Change  16465 connections/s    22% CPU utilisation
> 
> So fifty percent more connections for a quarter the CPU.

Is this TLS connections?

> At this point a number of other locks are causing the slow down.

I'd like to know more...

> Now, I'm not sure if such a change would be interesting to the community or
> not, but there definitely  is room for significant gains in the multi threaded
> locking.  Ignoring the atomic operations and moving to a separate lock per
> reference count would likely save a an amount of blocking -- is this a
> suitable use for dynamic locks?

As Matt mentioned, I'm curently working exactly on this. Would it be possible
for you to share your testing method and code? I'd like to verify that my
patches actually go in the right direction before having them merged (or maybe
you could do your tests on top of my patches, they should mostly work fine even
if the work is not complete yet, I think).

Thanks