Re: [rtcweb] Require/Suggest AEAD GCM for SRTP

["Martin Thomson" <[e-mail]>]

On Thu, Jul 11, 2019, at 09:03, Justin Uberti wrote:
> We looked into this in Chrome in 
> https://bugs.chromium.org/p/chromium/issues/detail?id=713701, but we 
> decided not to proceed because of the resultant blowup from using the 
> non-truncatable AEAD MAC (16 bytes per packet vs 4/10 for HMAC-SHA1).

I don't find the extra bytes that big of a deal, across an entire PMTU-sized packet, it's not significant.  I'm sure you can say that it makes a difference in terms of responsiveness or number of packets, but it's <0.5% of a modest MTU for the 10 vs. 16 comparison.

32-bit auth tags are pretty ridiculous if you think that authenticity is a valid goal.  We're considering turning them off entirely.  That said, if we had a cipher with a shorter tag that wasn't awful, I'd probably concede that maybe we could have fewer than 128 bits of tag for time-sensitive applications.
 
> I think we'd be open to revisiting this if there were obvious 
> performance benefits, but your numbers for HMAC-SHA1 seem unusually 
> bad. For example, "openssl speed sha1" yields 500 MB/s for 256-byte 
> packets on my MacBook Pro, compared to the 28 MB/s that you noted in 
> the bug. "openssl speed aes-128-gcm" does yield 1500 MB/s, so there's 
> clearly some upside here, but it's hard to see this as a must-have.

Keep in mind that you are running AES in counter mode and HMAC-SHA1 will add an extra invocation per packet (which is small, but noticeable), so a better comparison would be GHASH with SHA1.  With acceleration GHASH is much faster.  Without, it's is slower, but most modern machines have carryless multiply instructions.  AES-GCM runs at about a cycle per byte with native instruction support.  SHA-1 can get close to that on the very latest CPUs, but even CPUs from last year are twice as slow.  So add AES-CTR and that is consistent with your measurements.

Only for the oldest machines (like a 2012 ARM chip, for instance) will AES-CTR+HMAC-SHA-1 start to be comparable to the performance of a constant-time AES-GCM.