[openssl-users] Session Ticket Support in Openssl TLS 1.2

[Matt Caswell]

On 16/06/17 23:51, Neetish Pathak wrote:
> Thanks Matt, Appreciate ur response and tips
> 
> On Fri, Jun 16, 2017 at 3:36 PM, Matt Caswell <matt at openssl.org
> <mailto:matt at openssl.org>> wrote:
> 
> 
> 
>     On 16/06/17 20:08, Benjamin Kaduk via openssl-users wrote:
>     > On 06/16/2017 01:58 PM, Neetish Pathak wrote:
>     >> Hello
>     >> Thanks
>     >> I tried reading  some content from the server side and I observed the
>     >> new_session_cb getting invoked in that case on the client side. I
>     >> understand that may be due to delayed NewSession info transfer from
>     >> server side to client side. But it is helpful for saving the session
>     >> info on the client side. (Thanks Matt for your input)
>     >>
>     >> However, now I have two concerns
>     >>
>     >> 1) I see that latency for handshake with session resumption in
>     TLS 1.3
>     >> has not improved as much as it improves for TLS 1.2
>     >> With TLS 1.3, I observed that resumption brings down the connection
>     >> time from 2.5 ms to 1.2-1.3 ms
>     >> while with TLS 1.2 (using either session IDs or tickets), the
>     >> connection time reduces from 2.5 ms to 0.5-0.6 ms.
>     >>
>     >> The whole code is similar for running the two experiments with only
>     >> max TLS version changed. Can someone comment on the latency
>     >> measurements for the two cases.
>     >> TLS 1.3 is supposed to be better than TLS 1.2 in my opinion. Please
>     >> comment.
>     >>
>     >
>     > Are you seeing a HelloRetryRequest in the resumption flow?  That would
>     > add an additional round trip.  (Your numbers in milliseconds are of
>     > course not transferrable to other systems; round-trips is an easier to
>     > understand number.)  RFC 5246 and draft-ietf-tls-tls13-20 both have
>     > message-flow diagrams that show the number of round trips in various
>     > handshake flows.
> 
>     Care should also be taken about when you are starting your "timer" and
>     when you stop it, e.g. if you stop your timer after the session ticket
>     data has been received by the client then this is not a "fair" test (the
>     connection is ready for data transfer earlier than the arrival of the
>     session ticket).
> 
>     I would expect to see the TLS1.3 latency for a full handshake to be
>     around the same as for a TLS1.2 resumption handshake. With a TLS1.3
>     resumption only marginally different. There are the same number of round
>     trips for a TLS1.3 full handshake as that there are for a resumption
>     one. The primary difference is that the Certificate message is not sent
>     (which can be quite a large message).
> 
>     Your timings suggest you are testing this over a LAN where the effects
>     of network latency are going to be relatively low. The real benefits
>     from fewer round trips will really be seen when network latency is more
>     significant.
> 
> 
> 
> In my application program, I put start and stop timer before and after
> SSL_accept on server side and before and after SSL_connect on the client
> side.

That should be fine (my worry was that you might also be including the
subsequent session ticket transfer).

> I am not sure how I can put timers at individual steps of Handshake
> using my client applications. I was assuming SSL and SSL_accept take
> care of the entire handshake process.
> 
> Yes, I am testing on local machine. I will migrate my test to remote
> machines. But I am not really able to understand why TLS 1.3 is slower
> on my machine.

If you are on a local machine I would not expect a significant speed up
in TLSv1.3 vs TLSv1.2. TLSv1.3 has been designed to reduce the number of
round-trips required in order to avoid unnecessary network latency. If
you are on a local machine then there isn't any significant network
latency anyway - so timings are presumably dominated by the CPU
intensive tasks. You should make sure that you are comparing like with
like, i.e. the same ciphers, key lengths, key exchange algorithms,
curves etc between TLSv1.2 and TLSv1.3. Differences in any one of these
could obviously have significant performance impacts.

Matt