2º WCloud-HPC (Online)

Abstract: Cloud computing has emerged as a flexible support for high-performance computing activities. There are several reasons or arguments for choosing this paradigm that mainly come from the standpoint of costs and technical safety. However, in any case, the main concern in high-performance computing is the runtime efficiency, which includes both the efficiency of the tasks and that of the corresponding data accesses. The latter is critical in the context of geographically distributed data centers because they are typically located in different places far from each other, which might imply a heavy network activity to have the data available at the right location whenever they are needed by the running tasks. One of the advantages of Cloud computing from the user standpoint is that, through the virtualization, explicit scheduling details are seamlessly managed on the provider’s side. This is also the case of data management, which includes physical storage strategies and explicit migrations. In order to avoid the severe penalty of an inadequate default data management mechanism in the context of geo-distributed centers, it is worth considering a more skillful organization. We first propose an efficient data placement strategy by considering the characteristics of the tasks and their allocation onto the data centers. Then, we suggest a dynamic procedure for data provisioning using explicit redundancy and efficient choice of the providers. This talk will present the context and the problem, followed by our solutions and related thoughts.

Abstract: Data Stream Processing is an emerging computing paradigm characterized by the continuous analysis of data streams. Several application domains are of special interest for stream processing, such as financial trading, sensor networks, environmenal monitoring, network analysis and many others. Modern Stream Processing Systems are designed to exploit at best scale-out environments such as a Cloud, where horizontal scalability is addressed by spreading the applications in multiple nodes. However, there is the urgent need to target different computing environments like scale-up servers and machines equipped with co-processors (e.g., GPUs and FPGAs), both in the form of traditional hosts as well as resource-constrained edge and IoT embedded devices. This talk will try to explain the most promising research perspectives in this stream processing evolution, with a special focus on the ongoing research conducted for this purpose at the University of Pisa.