(I'll try to do some perf measurements on BinIO vs. BinMem later today, that should give us a rough idea).
I did a rough test using GHC's current Binary library hacked to be standalone. The test was writing 100000 Ints, followed by reading back 100000 Ints. Results: using BinMem took about 0.25 secs, BinIO took about 1.1 secs. That is, BinMem is current 4-5 times faster than BinIO. I should stress that this isn't a realistic benchmark. I conjecture the difference is mostly due to the overhead of going through the IO library each time a character is written. When I wrote the Binary library I was pretty careful to make sure that putWord8/getWord8 on a BinMem was just straight-line code (that's what the FastMutInt stuff is for). The upshot is that if you want really good performance for BinIO, there will have to be a cache in the BinHandle. Cheers, Simon