The speed of the compression codec is only part of the story, you should also consider the support for the codec in different parts of the Hadoop stack. Gaining slightly faster compression at the expense of compatibility is probably not a good trade off. Snappy is supported by pretty much all of the stack for example, whereas LZ4 is not currently supported by Impala.
If in doubt I would stick with Snappy since it is a reasonably fast and splittable codec. If performance is an issue you're likely to find greater benefit focusing on other parts of the stack rather than data compression.