The third part of this series presented PBKDF2 as a modern key derivation and password hashing algorithm. But PBKDF2 has its limitations; for best protection against password cracking the iteration count (defining the computing power needed to hash a password) should be chosen as high as possible. On the other hand, a higher iteration count also means that a login of a regular user will be slower. The maximal time users are prepared to wait for a successfull login will limit the maximal iteration count which you can choose for the available computing power.
For some time we could at least assume that all but the most resourcefull attackers will have roughly the same computing power at hand as the defenders have on their login servers. An attacker might be able to set up (and finance) hardware to hash passwords 100 or even 1000 times faster than a server, but this could be compensated for by a sufficiently high iteration count. However, by using hardware specialized towards massively parallel execution of hashing operations the relation of the average servers and the potential attackers “hashing power” shifted more and more to the advantage of the attacker. Hashing algorithms like the scrypt algorithm presented in this Blog article attempt to shift this relation back in favor of the defender.