Design and implementation of a framework for efficient remote reconfigurable accelerator deployment in disaggregated environment

This work has been supported in part by EU H2020 ICT project dRedBox, contract #687632Summarization: Cloud computing usage has drastically increased over the years, providing data security and privacy which are prime concerns these days. The scalability of the cloud capacity, as well as the accessibility of the provided services, constitute relevant factors for the cloud computing evolution. Data centers are mainly deployed as cloud computing resources to deal with large storage and computation requirements. The need for specialized hardware acceleration in this domain is well established and intensified by the insatiable demand for compute power. Hardware accelerators can also provide high energy efficiency for many application domains in comparison with current architectures based on general purpose processors. The fixed amount of the available resources constitutes the major disadvantage of traditional data centers. Resource disaggregation alleviates this issue while offering the opportunity to manage resources more efficiently. In disaggregated computing environments, where all data transfers between remote nodes are realized via packet exchanges over a rack-scale network, reducing communication and synchronization is a prerequisite to the effective employment of remote acceleration. To this end, this thesis presents ReFiRe (Remote Fine-grained Reconfigurable acceleration), a generic deployment framework with native support for partial reconfiguration that allows considerable reduction of communication needs between a processor and remote accelerators. Custom instructions that encapsulate complex sequences of operations and their respective synchronization requirements deployed for shifting control flow and partial reconfiguration decisions to the remote side. Considering the high complexity of the instruction-initialization procedure, a source-to-source transformation framework based on the ReFiRe infrastructure was further implemented. Through this framework, these instructions are automatically generated according to application requirements transparently to the user level. To evaluate ReFiRe, three benchmark applications were employed. A 2D-FFT algorithm , a genomics application that detects positive selection in genomes and a Binarized Neural Network, demonstrate that offloading computations to remote accelerators using ReFiRe leads to superior aggregate performance on the same specialized hardware platform compared to using dedicated accelerator calls on a per-operation basis