Search

Storing and streaming high dimensional data for foundation model training became a critical requirement with the rise of foundation models beyond natural language. In this paper we introduce TensorBank, a petabyte scale tensor lakehouse capable of streaming tensors from Cloud Object Store (COS) to GPU memory at wire speed based on complex relational queries. We use Hierarchical Statistical Indices (HSI) for query acceleration. Our architecture allows to directly address tensors on block level using HTTP range reads. Once in GPU memory, data can be transformed using PyTorch transforms. We provide a generic PyTorch dataset type with a corresponding dataset factory translating relational queries and requested transformations as an instance. By making use of the HSI, irrelevant blocks can be skipped without reading them as those indices contain statistics on their content at different hierarchical resolution levels. This is an opinionated architecture powered by open standards and making heavy use of open-source technology. Although, hardened for production use using geospatial-temporal data, this architecture generalizes to other use case like computer vision, computational neuroscience, biological sequence analysis and more.

Within the next 10 years, advances on resource disaggregation will enable full transparency for most Cloud applications: to run unmodified single-machine applications over effectively unlimited remote computing resources. In this article, we present five serverless predictions for the next decade that will realize this vision of transparency -- equivalent to Tim Wagner's Serverless SuperComputer or AnyScale's Infinite Laptop proposals., Comment: arXiv admin note: text overlap with arXiv:2006.01251

For many years, the distributed systems community has struggled to smooth the transition from local to remote computing. Transparency means concealing the complexities of distributed programming like remote locations, failures or scaling. For us, full transparency implies that we can compile, debug and run unmodified single-machine code over effectively unlimited compute, storage, and memory resources. We elaborate in this article why resource disaggregation in serverless computing is the definitive catalyst to enable full transparency in the Cloud. We demonstrate with two experiments that we can achieve transparency today over disaggregated serverless resources and obtain comparable performance to local executions. We also show that locality cannot be neglected for many problems and we present five open research challenges: granular middleware and locality, memory disaggregation, virtualization, elastic programming models, and optimized deployment. If full transparency is possible, who needs explicit use of middleware if you can treat remote entities as local ones? Can we close the curtains of distributed systems complexity for the majority of users?