NUMA-Aware Memory Management with In-Memory Databases

Abstract

Writing enterprise grade software for multi-processor systems is an interesting challenge since such a system primarily involves a multitude of hardware components that exhibit conflict due to simultaneous access by unorganized software threads of user applications. The problem is particularly compounded with In-Memory paradigm that includes potential applications like Data Management in the modern era. With an emergence of distributed hardware trends like Non-Uniform Memory Access (NUMA), where access times to a system’s physical address space depend on relative location of Memory w.r.t CPU, it is crucial to rethink about placement of a user process’ workable memory with respect to executing threads. We present a few novel techniques from our Heap management work with SAP HANA as part of our goal towards building a strong NUMA awareness with in-memory databases. Our work primarily focuses on providing a robust and well-performant Memory Management framework on Linux OS by handling the associated complexity and challenges seen with enabling enterprise software to live on a distributed memory landscape. One of the important outcomes of our approach is to build a rich set of kernel APIs that provide fine-granular control to higher DBMS layers like Store and Query for educated placements of their relational data structures. However the generality of our techniques allows them to be readily applied to other domains that need to deal with NUMA performance penalty.