The rapid increases in model parameter sizes introduces new challenges in pre-trained model loading. Currently, machine learning code often deserializes each parameter as a tensor object in host memory before copying it to device memory. We found that this approach underutilized storage throughput and significantly slowed down loading large models with a widely-used model file formats, safetensors. In this work, we present fastsafetensors, a Python library designed to optimize the deserialization of tensors in safetensors files. Our approach first copies groups of on-disk parameters to device memory, where they are directly instantiated as tensor objects. This design enables further optimization in low-level I/O and high-level tensor preprocessing, including parallelized copying, peer-to-peer DMA, and GPU offloading. Experimental results show performance improvements of 4.8x to 7.5x in loading models such as Llama (7, 13, and 70 billion parameters), Falcon (40 billion parameters), and the Bloom (176 billion parameters).

In this paper, we design and implement the first-ever decentralized replicated relational database with blockchain properties that we term blockchain relational database . We highlight several similarities between features provided by blockchain platforms and a replicated relational database, although they are conceptually different, primarily in their trust model. Motivated by this, we leverage the rich features, decades of research and optimization, and available tooling in relational databases to build a blockchain relational database. We consider a permissioned blockchain model of known, but mutually distrustful organizations each operating their own database instance that are replicas of one another. The replicas execute transactions independently and engage in decentralized consensus to determine the commit order for transactions. We design two approaches, the first where the commit order for transactions is agreed upon prior to executing them, and the second where transactions are executed without prior knowledge of the commit order while the ordering happens in parallel. We leverage serializable snapshot isolation (SSI) to guarantee that the replicas across nodes remain consistent and respect the ordering determined by consensus, and devise a new variant of SSI based on block height for the latter approach. We implement our system on PostgreSQL and present detailed performance experiments analyzing both approaches.

Fabric is a modular and extensible open-source system for deploying and operating permissioned blockchains and one of the Hyperledger projects hosted by the Linux Foundation (www.hyperledger.org).

There is an emerging class of enterprise applications in areas such as log data analysis, information discovery, and social media marketing that involve analytics over large volumes of unstructured and semi-structured data. These applications are leveraging new analytics platforms based on the MapReduce framework and its open source Hadoop implementation. While this trend has engendered work on high-level data analysis languages, NoSQL data stores, workflow engines etc., there has been very little attention to the challenges of deploying analytic workflows into production for continuous operation. In this paper, we argue that an essential platform component for enabling continuous production analytic workflows is an analytics store. We highlight five key requirements that impact the design of such a store: (i) efficient incremental operations, (ii) flexible storage model for hierarchical data, (iii) snapshot support (iv) object-level incremental updates, and (v) support for handling change sets. We describe the design of SASH, a scalable analytics store that we have developed on top of HBase to address these requirements. Using the workload from a production workflow that powers search within IBM's intranet and extranet, we demonstrate orders of magnitude improvement in IO performance using SASH.

Migrating production applications from physical datacenters to a virtualized environment is becoming essential to reduce operational costs. In order to avoid direct access to production systems, migration from disk snapshots is preferable. Model based migration approaches are not suitable for this purpose since they need to perform configuration discovery on production systems. Recent approaches that rely upon an isolated network require special setup and are limited to handling network configurations only. In this paper we present a framework for instantiating an application from disk snapshots in a virtualized environment. Our approach comprises of formally specifying the knowledge about product configurations and compiling the knowledge into a plan. The plan is capable of discovering application configurations in an inconsistent configuration setup. A key feature of our framework is that knowledge is specified once per product and re-used across applications that use the product. We have implemented a prototype of our approach and evaluated it on a real world application. The evaluation demonstrates the feasibility of our approach for migration using disk snapshots.

Heterogeneity has been the curse of the IT industry and systems management is no exception. As organizations are expanding their global business, and new technology and platforms are being introduced, the complexity of systems management is increasing rapidly due to heterogeneity and getting unified view of IT environments is becoming more difficult. To simplify and expedite information integration, this paper brings best practices of model-driven system and information integration to the problems of information integration in systems management. Specifically, we present an open framework for describing meta-model of a new component in isolation using UML and relating it with meta-models of other components through semantic relationships. The resulting integrated meta-model together with querying mechanisms for various sources can federate automatically and on-demand, integrated model of IT environment without relying on a central database repository.