Q: Design a system to calculate the number of unique words in a file
1) What if the file is huge? (i.e. cannot fit in the main memory)
2) Assuming that you have more than one computer available, how can you distribute the problem?
Constraints are the key to such an optimization. Let’s make it more realistic but hopefully without loss of generality. Say the file is 2 TB ascii of purely alphabetical words of any language in a unified alphabet, with natural distribution such as text from world newspapers. Word length is typically below 20.
I’d assume regular 100GB network with dedicated sockets between machines. The machines have roughly equal memory, and the combined memory is enough to hold the file.
I’d minimize disk and network access since these are slower than memory access and require serialization.
Q: is the network transfer such a bottle neck that I’m better off processing entire file in one machine?
— one-machine solution —
Assuming my memory (2GB) can only hold 1% of the unique words. I’d select only those words “below” ad* — i.e. aa*, ab*, ac* only. Save the unique words to a temp file, then rescan the input file looking for ad*, ae*…ak* to produce a 2nd temp file… Finally Combine the temp files.
— multi-machine solution —
Don’t bother to have one machine scanning the file and tcp the words to other machines. Just copy the entire input file by CD or file transfer to each machine. Each machine would ignore words outside its target range.
How do we divide the task. Say we have 50 machines. We don’t know the exact distribution, so if we assume aa-ak to Not have too many unique words to fit into one machine (2GB), assumption might be wrong. Instead, we’d divide the entire universe into 50 * 10 ranges. We assume even if we are underestimating, still each range should fit into one machine. Every time a machine finishes one range, it sends a tiny signal to a controller and waits for controller to give it next range.
— hashing on words —
Hash table should be sized to minimize rehash. We need superfast hashCode and compression. hashcode should use all the characters, perhaps except the first, since it tends to be the same within a range.