Overclocking SSL

RSA ciphersuites involve the client generating a random session key, encrypting it using the server's public key and sending it to the server. The session key is then used to key the MAC and bulk cipher.

If an attacker records the traffic, then factors or steals the public key, they can decrypt the session key and read the traffic.

With ephemeral Diffie-Hellman, a random DH public value is generated for each connection and the server signs it with its public key. Each side then performs key agreement to generate the session key. Now, an attacker can solve the discreet logarithm and decrypt the traffic, but they have to break each connection separately.

The trick in the future is to perform Diffie-Hellman much faster (answer: write tuned code to work in an elliptic curve group) and then to perform EDH without the latency hit (longer term).

Just to add to what agl said:

With RSA key exchange, if somebody ever gets your RSA private key, they can decrypt all past and future traffic that was protected with that key. With ephemeral Diffie-Hellman key exchange, even with the RSA private key, they cannot decrypt that past and future traffic, because each session has its own unique encryption keys that are independent of the RSA keys. They can only impersonate you to intercept future traffic, and only if they set up a MITM attack, but that kind of active attack is MUCH harder to execute (especially undetected) than the passive attacks that RSA key exchange allows. And, again, it doesn't allow them to decrypt past traffic at all.

This advantage of ephemeral key exchange is called "perfect forward secrecy."

For example, let's say you are a dictator and you want to decrypt some encrypted GMail sessions you suspect were done by dissidents in your country last month. If those sessions were done with RSA key exchange, you could steal the RSA private key from Google today (admittedly a very difficult task) and then decrypt those past sessions. If (EC)DHE key exchange was used, it would be practically impossible (no exaggeration) to ever decrypt them.

thanks for the replies.

If you can afford 4096-bit certs, or you need them, then you should use them. However, on my laptop, I can only handle 34 connections/second/core with a 4096-bit cert. Such performance is likely to be unacceptable for many uses.

Bigger numbers take much longer to crunch.

They specifically mention that the certificate size (and that of intermediates, among other things) impacts the message size and thus the TCP packet sizes used during handshakes. They actually explain all of this quite well, just read the entire article [specifically the "Certificates" section].