面向 MemOS 中 NDP 的数据中心编程模型.pdf

1、Daniel BittmanCTO,Elephance Memory,Inc.Data-centric Programming with MemOSMemOSis a trademark of Elephance Memory,Inc.AI,HPC,Graph processing:high demands for throughput,parallelism,capacityNew interconnects promise memory disaggregationNew memory device promiseNear-Data Processing(NDP)orProcessing-

2、in-Memory(PIM)Applications demands are increasingComplexity:not just in management betweenheterogeneous devices,but also with data layout,optimization,and data movement and sharing.Wasted CPU time and energy due to frictionbetween application needs,hardware capabilities,and system software interface

3、s.Challenge:wide data sharing withoutsignificant coordination,for complexdata structures(graphs+vectors).New hardware is a panaceaData gravity implies the need for NDP,but dealing with complex data structures(e.g.graphs,pointer chasing)is challengingThe need to modify data as it moves betweencontext

4、s impacts data movement overheads(e.g.serialization)and data sharing.Data representation considerations forwidespread sharing combined with globalcompute metaphors and local optimizationsSharing is challengingBuild applications for full-stack parallelism,from large-scale to vector engines,enableshar

5、ing,and provide direct access to data.Open compute and data formats enable stablein-memory representations with industry supportfor standardization.Independent scaling of memory and compute.Parallelism,wide-sharing,data-centricMemOS provides a data-centric programmingmodel that gives applicationsdir

6、ect access to memory,including for complex structureslike graph and vector data.Domain-specific librariesprovide abstractions tobuild up data structures(e.g.an HNSW index)MemOS and Memory ObjectsMemory objects are regions of memory that contain datawith fluid movement and zero-copy sharing.Enable lo