illumos gate

libumem does a good job at providing a scalable malloc implementation for heavily threaded applications. However, in the case of uncontended allocations, libumem still has to always grab a single lock. In small object malloc heavy code, this adds a noticeable amount of overhead, especially when compared to other systems where they have lock-free ways of returning small objects. This work investigates adding a per-thread cache to libumem where recently freed allocations are stored in the thread itself.