On the other hand, you might not want to be hashing a 20meg password. It is fast on my computer but it's fair to limit at something more reasonable.

    $ python -c 'print "8 bytes\n" * (20 * 1024 * 1024 / 8)' > 20meg.txt; time shasum -a 512 20meg.txt 
    59cb7f88ad8d6229e6d3a74ee422dff57e17f168c6e6fa44ef32c3f07a73a6e455d8b55c1265d5212b9ed5475b6d9364286645200dada59aa16905a9ce748561  20meg.txt

    real	0m0.289s
    user	0m0.284s
    sys	0m0.004s

    $ python -c 'print "8 bytes\n" * (16 / 8)' > 16byte.txt; time shasum -a 512 16byte.txt
    b6043d3a520424d5ec17dc0c23ba3b591d74517e2b9faa0df2d69d13c89a5f372d6dc35f95836687ee05be18433277e1c4b67393eb2771b475d655a832b16654  16byte.txt

    real	0m0.045s
    user	0m0.040s
    sys	0m0.004s

Fast password hashing is bad anyway. Slow schemes are better (assuming they come from a thoroughly tested library written by someone that actually knows what they're doing).

I'm admittedly not familiar with the details of the hashing process, but it could be done client-side, no? Then the compute power required falls on the user, PLUS the 20 megs never gets sent over the wire.

No. Actually we would not want the hashing technique to be exposed in the source code.

That turns out not to be the whole story.

Hiding the technique to compute the hash is relying on security by obscurity. Ideally you want a system that is secure even if potential attackers know what methods you're using.

If you do the hashing on the client side, then the hash itself effectively becomes the password.

That would assume your users have JS enabled. I've seen something like that done before, but always with a fallback in case user has JS disabled.

If you're not sending 20 megs of data, you're not getting 20 megs of security. So why allow it if it doesn't add anything?

It doesn't add to the security, but I might find it easier to remember 20 MB of redundant and meaningful stuff than to remember 384 bits of literally random stuff. The entropy might be the same, but my memory is not a computer. I can remember vast amounts of material that is meaningful and use it as a password. I can't remember 384 bits that have no meaning.

The benefit isn't in the entropy, it's in the abilities of your users to remember their passwords/passphrases in the first place.

Maybe 1 or 2 KB, max. 20 million characters is a ridiculous password to remember.

That's not the point - the difference between 2 KB and 20 MB is purely a detail. You said:

  > If you're not sending 20 megs of data,
  > you're not getting 20 megs of security.
  > So why allow it if it doesn't add anything?

You could just as equally say:

  > If you're not sending 2 KB of data,
  > you're not getting 2 KB of security.
  > So why allow it if it doesn't add anything?

Your point is the same, and it's still wrong. What you're getting is not the security - that's only half the story. My point is that is does add something, it's just that the something it adds isn't the entropy for the purpose of security.

When there will be multiple shorter passwords that hash to the same value, is there a point to a 20mb pass?

Yes. Ideally you want users to be able to remember their pass-phrases. To do so usually implies significant internal structure and/or correlations, so to get the necessary entropy they will be large. The fact that they hash to the same value as other things is effectively irrelevant.

I misspoke, I meant size.

Depends. Can you guess them?

If I'm an attacker who is running through hashes...yes. Faster than the 20mb one.

There is a slight exception to this. If they are using an older statically typed language like C, it might make sense for them to have a limit on the buffer ready to store your unhashed password. Yes it seems crazy these days, but it might apply to some of the older systems which have password length limits.