Venti is just a block storage server. It accepts blocks ranging in size from 0 bytes to 56kB. The application can use whatever block size it likes. 8k is typical. When I back up file systems I use the underlying file system block size.
The current Venti server (in Plan 9) does check when you write a block, if it hashes to an existing block, that the two are the same. My newer Venti server (in Plan 9 from User Space) does not do this, which is faster. All this discussion about things more likely than two random blocks having the same hash is amusing, but there is a serious point no one has brought up. All the math depends on blocks chosen randomly. An adversary might actually come up with two blocks with the same hash, not by random search but by being very clever. Some researchers in China recently claimed to have a program that generate two different blocks with the same MD5 hash (and I think there is one for SHA1 too, that takes longer). I ran the MD5 program for a while on my laptop but it did not finish. Maybe I just got unlucky (there is still a little randomness). It was supposed to be able to finish in something like 45 minutes and I ran it overnight. I don't know whether their approach generates two blocks of the same length. I do know that if they are trying to match a pre-existing hash they do so by adding padding in the form of some kind of comment. For example, they could start with a PDF that said "Buy 1M shares of X", change the Buy to Sell, and then insert some comments in the PDF to make the hash of the new document the same as the hash of the original document and thus the old signature would work for the new document. This is bad and will be worse as computers get faster, and various people are worried about how to switch to SHA256. I am not worried. If there is some adversary using your Venti system, there are simpler attacks they could use to render it inoperable (like fill it up). I am happy to assume that the Venti clients are playing nicely. If, at some point in the future, it was really a problem, the types that Venti stores with each block would allow a Venti server to be "transcoded" into a different hash function pretty easily. (Of course, the clients would have to be informed of the new hashes of their root blocks.) Russ
