We've migrated our port of Solaris umem from SourceForge to OmniTI Labs, and have finally gotten around to making a source release.
You can find the project page here, which includes a link to the downloadable source.
The tarball will build you libumem and libumem_malloc. The latter version can be used as an LD_PRELOAD to replace malloc for a target process, which is useful for debugging. The former is more useful when deliberately linked into an app that wants to take advantage of the slab allocator APIs directly.
I'm interested to see if using this in PHP will lead to performance gains, but currently lack the time to do any benchmarking. I doubt that performance will suddenly improve just by preloading libumem_malloc, but there might be a gain if the zend engine allocator were tied into libumem.
Perhaps someone reading this blog entry does have the time to investigate and could post the results?