To compete with Memcached's multi-core throughput, it's probably not a big deal for you to run a Redis process per core. Surely you are already running multiple memcached for higher availability, so I can't see multiple Redis procs adding much more complexity.
Of course, Redis has many more advantages. For one, I can't be bothered to ever again not have Redis data types, so Memcached just doesn't make sense in my stacks. And even a caching layer has advantages in being able to persist to disk, so I love that too about Redis.
If you don't know why you need one or the other, use redis until you can't anymore. Plenty of people still use memcache, although lots of the current users are forking their implementation to get specific behaviors from the internals.
Can you elaborate on what you mean by "trivial data sets" and why is redis only suitable for those?
Running redis as a cluster is much more complicated.
Redis rides a bit on its hype IMO, and I see that sometimes it is not used correctly.