Did you try benchmarking the serialization in memory without any disk I/O?

In my experience Storable is amazingly fast and it was the disk that quickly became the bottleneck.
But then I was only using ~ 5k structures.