Brief benchmark docs

bench/README.md

@@ -0,0 +1,30 @@
+# snappy-c benchmarks
+
+There are five benchmark groups, all benchmarking compression and/or
+decompression on 2MB chunks of data:
+
+1. `random`: Pseudo-randomly generated bytes, the least compressible type of
+   data that we benchmark on.
+2. `zeros`: All zeros (`\NUL`). The most compressible type of data that we
+   benchmark on.
+3. `words`: A repeated sequence of 8 bytes, over and over. Also extremely
+   compressible.
+4. `interleaved_words`: Repeated sequences of 8 bytes, separated by 8 random
+   bytes in between. Meant to occupy a space on the "compressibility" spectrum
+   in between `random` and `words`.
+5. `combined`: All of the above types of data.
+
+`combined` is meant to capture performance on all the types of data meantioned
+above, so it's a good one to use if looking for regressions or more balanced
+performance metrics.
+
+Run all benchmarks with:
+```
+cabal run bench-snappy-c
+```
+
+Or just the combined with:
+```
+cabal run bench-snappy-c -- 'combined'
+```
+