Georges Da Costa

Since the Paris Agreement, there has been increased focus on decreasing greenhouse gas (GHG) emissions related to power supply. Data centers are a huge energy consumer sector due to their continuous operational demand. A way to reduce their impact on GHG is by using renewable energy in their power supply. However, renewable energy introduces production intermittence since it depends on the climate conditions. Geo-distributed data centers can exploit diverse climate conditions to diminish the problem. This work presents a geo-distributed data center marketplace only supplied by renewable energies. In our marketplace model, the data centers are independent and compete for the users’ jobs. The winner data center is defined using a reverse auction process. We compare different aspects, such as centralized/distributed architectures and cooperative/non-cooperative marketplaces. Our results show that the cooperative and distributed data center model earns more money (5.39% more money), wastes less energy (3.81% less energy), and provides higher QoS (improvement of 4.31% for services and 2.71% for tasks).

With the increase of demand for computing resources and the struggle to provide the necessary energy, power-aware resource management is becoming a major issue for the High-performance computing (HPC) community. Including reliable energy management to a supercomputer's resource and job management system (RJMS) is not an easy task. The energy consumption of jobs is rarely known in advance and the workload of every machine is unique and different from the others. We argue that the first step towards properly managing power is to deeply understand the power consumption of the workload, which involves predicting the workload power consumption and exploiting it by using smart power-aware scheduling algorithms. Crucial questions are (i) how sophisticated a prediction method needs to be to provide accurate workload power predictions, and (ii) to what point an accurate workload's power prediction translates into efficient power management. In this work, we proposed a method to predict and exploit HPC workloads power consumption with the objective of reducing the supercomputers power consumption, while maintaining the management (scheduling) performance of the RJMS. Our method exploits workload submission logs with power monitoring data, and relies on a mix of lightweight power prediction methods and a classical EASY Backfillling inspired heuristic. Then, we model and solve the power capping scheduling as a greedy knapsack algorithm. This algorithm improves the Quality of Service and avoids starvation while keeping the solution lightweight. We base this study on logs of Marconi 100, a 980-node supercomputer. We show using simulation that a lightweight history-based prediction method can provide accurate enough power prediction to improve the energy management of a large scale supercomputer compared to energy-unaware scheduling algorithms. These improvements have no significant negative impact on performance.

Current cloud environments provide numerous storage functions to virtual machines such as disk encryption, snapshotting, compression and so on. These functions are implemented using software stacks inside the hypervisor's kernel, emulator, or as a userspace polling driver like SPDK. However, each stack brings its own limitations: Linux's kernel I/O stack cannot easily integrate proprietary technologies such as Intel SGX, while SPDK requires significant changes in software development and tooling yet lacks the rich feature set of existing solutions like Linux LVM. To remedy these limitations, we introduce NVMetro, a high-performance storage framework for virtual machines based on the NVMe protocol. NVMetro provides multiple I/O paths that can be dynamically combined to fit the needs of each storage function. It links these paths together with an eBPF-based I/O router/classifier framework, as well as a userspace software stack for out-of-kernel I/O processing. We implemented three different storage functions with NVMetro and evaluated them under various workloads. Our results show that NVMetro approaches the performance of kernel-bypass solutions like SPDK while maintaining the compatibility and ease of use of in-kernel storage stacks.

With the increase of demand for computing resources and the struggle to provide the necessary energy, power-aware resource management is becoming a major issue for the High-performance computing (HPC) community.Including reliable energy management to a supercomputer's resource and job management system (RJMS) is not an easy task. The energy consumption of jobs is rarely known in advance and the workload of every machine is unique and different from the others. We argue that the first step toward properly managing energy is to deeply understand the energy consumption of the workload, which involves predicting the workload's power consumption and exploiting it by using smart power-aware scheduling algorithms. Crucial questions are (i) how sophisticated a prediction method needs to be to provide accurate workload power predictions, and (ii) to what point an accurate workload's power prediction translates into efficient energy management. In this work, we propose a method to predict and exploit HPC workloads' power consumption, with the objective of reducing the supercomputer's power consumption while maintaining the management (scheduling) performance of the RJMS. Our method exploits workload submission logs with power monitoring data, and relies on a mix of light-weight power prediction methods and a classical EASY Backfillling inspired heuristic. We base this study on logs of Marconi 100, a 980 servers supercomputer. We show using simulation that a light-weight history-based prediction method can provide accurate enough power prediction to improve the energy management of a large scale supercomputer compared to energy-unaware scheduling algorithms. These improvements have no significant negative impact on performance.

With the increase of demand for computing resources and the struggle to provide the necessary energy, power-aware resource management is becoming a major issue for the High-performance computing (HPC) community.Including reliable energy management to a supercomputer's resource and job management system (RJMS) is not an easy task. The energy consumption of jobs is rarely known in advance and the workload of every machine is unique and different from the others. We argue that the first step toward properly managing energy is to deeply understand the energy consumption of the workload, which involves predicting the workload's power consumption and exploiting it by using smart power-aware scheduling algorithms. Crucial questions are (i) how sophisticated a prediction method needs to be to provide accurate workload power predictions, and (ii) to what point an accurate workload's power prediction translates into efficient energy management. In this work, we propose a method to predict and exploit HPC workloads' power consumption, with the objective of reducing the supercomputer's power consumption while maintaining the management (scheduling) performance of the RJMS. Our method exploits workload submission logs with power monitoring data, and relies on a mix of light-weight power prediction methods and a classical EASY Backfillling inspired heuristic. We base this study on logs of Marconi 100, a 980 servers supercomputer. We show using simulation that a light-weight history-based prediction method can provide accurate enough power prediction to improve the energy management of a large scale supercomputer compared to energy-unaware scheduling algorithms. These improvements have no significant negative impact on performance.

Billions smartphones and smart objects battery-powered use Android, i.e., on the Linux kernel. To save energy, the main kernel leverage is to put processors in a low power state as soon as they are idle. It predicts the next event to estimate the sleep duration and choose a sleep state accordingly. Several wake-up sources (interrupts, events...) impact this prediction which is usually done considering them as a single source. The resulting signal is nearly random and difficult to predict. Processors are recently supporting deeper idle states but the prediction paradigm was never challenged. We propose to predict the next event by splitting the wake-up source signal into simpler event patterns. We describe a fast and efficient algorithm and its kernel-level performance evaluation. We compare our approach with multiple reference sleep state selection algorithms on actual ARM and x86 boards using classical mobile workloads. Our proposal detects correctly (up to 20% improved correctness leading to 5% reduced energy consumption) the time of next interrupt, and thus the right sleep level for the processor. We show and discuss the energy impact of the tested prediction algorithm and we compare it with the different generations of sleep level managers in the Linux kernel.