I confess I don't see what the big deal is about 77 meg, I routinely use a 1 gig buffer to test linear speed of algos (a 1 gig VOB file does fine) and as you would expect, the first time you have disk read time, the next is straight memory speed. Just isolate what you are testing, if its disk IO, clock the first instance before memory operations, if its algo speed on data in memory, clock the time AFTER the disk IO.
If you are designing a task that must read very large files (larger than memory) work out a tiling scheme, disk IO in one thread and memory ops in one or more other thread(s). If you are running Win7 64 bit with 32 bit apps, you can run a larger buffer size and with inter-app communication, you can schedule multiple instances and synchronise them to perform much larger tasks using multi-core, multi-thread designs.
