I'm a dev guy from the Munin project(*). We have some heavy RRD among our users.
Some hints from the trenches:
* I discovered that having multiple rrdcached daemon is actually very bad for performance. The only way to shard them efficiently is to have them handling a different FS with different disk subsystem. (I didn't try with SSD btw)
* For some workloads tmpfs+sync (note that a simple cp might be more efficient than a rsync) is simply the best.
* You have to avoid doing reads on rrds, as it flushes. (1.4.x, don't know about 1.5)
* A very big -w and -z setting (think 3600) helps, as it give rrdcached the most flexibility to reorder writes
* Always flush manually before closing (FLUSHALL on the socket), to avoid too long restart times. The restarts journal read is way too slow, a you noticed.