I don't think it's possible to gather information like this. Memory isn't really used to process requests. There may only be one request to map a page in memory and then it may remain used for long periods of time without the system knowing how much it's used except for the periodic probing to look for stale memory.
Also, the reasoning behind the request is misguided. Even if most of the cache is never used, the hit rate of the part that is used is still determined by how much cache you have.
Let me give you analogy. Say there are 20 different kinds of balls and you can keep 18 of them in stock. If you cache randomly, there's an 18/20 chance you'll have the ball someone wants in stock. So if someone asks you for a particular ball and you have it in stock, only 1/18 of your balls will have been used. But you'll have a 18/20 chance of having that ball because of the other 17 balls sitting there unused.
So amount used to serve requests isn't really the right measurement.
The system doesn't really keep enough information to allow you to figure out how well the cache would have done had it been a different size.
Update: Let me try one more time to explain why this won't work. You're trying to infer from the fact that 5GB of RAM was used for cache that wasn't accessed that the system would have performed roughly the same if it had had 5GB less RAM. But that's totally false.
Say you run a book store. You notice that in a given month, you only sell 10% of your inventory. You think: "What a waste. 90% of my books are just sitting there unused. I don't need to keep so much inventory." So you reduce your inventory by 90%. What do you think will happen?
Yes, after the fact you see that lots of your inventory didn't sell. But with a much smaller inventory, most of your customers wouldn't have found the book they wanted in inventory. Knowing after the fact which books didn't sell doesn't mean you could have made do with a smaller inventory before the fact -- before you knew which books people would want and which they wouldn't.
So even if the information you want was available, it wouldn't let you draw the conclusion you want to draw. You would have to keep enough information to run a simulation and say -- had I kept fewer things in memory, which would I have kept? And thus, would I still have had the information that later got used?
So even if only a small amount of the data in the cache winds up being used, unless you could have predicted which information would be used, you can't infer that the rest of the cache didn't have a significant performance impact by allowing you to keep in memory the data that was used.