Section: New Results

System support: System support for multicore machines

Participants : Vivien Quéma, Renaud Lachaize, Baptiste Lepers.

Multicore machines with Non-Uniform Memory Accesses (NUMA) are becoming commodity platforms. Efficiently exploiting their resources remains an open research problem. This line of work investigates system support to tackle various issues related to efficient resource management and programming support.

One of the key concerns in efficiently exploiting multicore NUMA architectures is to limit as much as possible the number of remote memory accesses (i.e., main memory accesses performed from a core to a memory bank that is not directly attached to it). However, in many cases, existing profilers do not provide enough information to help programmers achieve this goal. We have developed MemProf [24] , the first profiler that allows programmers to choose and implement efficient application-level optimizations for NUMA systems. MemProf achieves this goal by allowing programmers to (i) precisely understand which memory objects are accessed remotely in memory, and (ii) building temporal flows of interactions between threads and objects. We evaluated MemProf using four applications (FaceRec, Streamcluster, Psearchy, and Apache) on three different machines. In each case, we showed how MemProf helped us choose and implement efficient optimizations, unlike existing profilers. These optimizations provide significant performance gains on the studied applications (up to 161%), while requiring very lightweight modifications (10 lines of code or less).

State-machine replication is a well-known fault-tolerance technique. Unfortunately existing state-machine replication schemes do not scale well on multicore machines. In collaboration with U. Texas at Austin (L. Alvisi), we have developed a new state-machine replication scheme [23] , that departs from the standard agree-execute architecture of existing schemes, in favor of a more optimistic, and less deterministic, execute-verify replication scheme, which yields much better scalability. We have evaluated Eve's throughput gain compared with traditional sequential execution approaches, as well as Eve's overheads compared to unreplicated multithreaded execution and to alternative replication approaches.