Makalu is a failure-resilient, thread-safe memory allocator that can be used for the allocation of byte-addressable non-volatile memory (NVRAM). Makalu is designed to restart across execution cycles using durable metadata stored in NVRAM. While it takes advantage of transient DRAM, CPU caches, and registers, Makalu internal crash consistency mechanism is developed with the assumption that only the data in NVRAM survive a crash. It uses memory fences, and selective cache line flushing instructions to guarantee the timely visibility of updates to its persistent metadata.
Makalu is expected to be used together with other NVRAM programming libraries (NVMPLs) such Atlas, and Mnemosyne. In such use-case scenario, the NVMPL ensures the consistency of the user data stored in heap, while Makalu ensures the consistency of the heap structure and the absence of failure-induced memory leaks. Makalu can be used in the absence of NVMPL as a stand-alone allocator. However, in such use-case, it is upto the programmer to ensure the consistency of the data stored in persistent heap across failures and restarts. We only recommend this for experts in persistent programming.
Makalu has two distinct phases in which it operates. In an online phase, a user can allocate and deallocate persistent memory after proper initialization. After an ungraceful shutdown, Makalu is expected to start in offline phase for recovery and offline garbage collection.
##Makalu API##