I generate an md5 hash from the filename, and use that to create a file hierarchy. I split the hash into chunks, which I then use to create subdirectories. This keeps the number of entries per level within filesystem limits, and lookup stays relatively fast. The distributive properties of MD5 garantuee a good spread among buckets.

For example, a filename of "foo12345.txt" might map to "/4d9/a43/9b6/eb3/f21/600/448/a03/b78/5a3/17". Each subdirectory has a maximum of 4096 subdirectories, which is well within the limit of ext3. By varying where you place the slashes, you control the breadth and depth of your directory tree.

I personally use a mapping of "/11/222/333/rest_of_hash" to archive (currently) over 10 million files. That scheme allows for 256*4096*4096 buckets, which is quite enough for my purposes. Speed and server load are excellent, even under highly concurrent usage.