Energy-aware scheduling of real time applications over multiprocessor systems is considered in this paper. Early research reports that while various energysaving policies, for instance Dynamic Power Management (DPM) and Dynamic Voltage & Frequency scaling (DVFS) policies, perform well individually for a specific set of operating conditions, they often outperform each other under different workload and/or architecture configuration. Thus, no single policy fits perfectly all operating conditions. Instead of designing new policies for specific operating conditions, this paper proposes a generic power/energy management scheme that takes a set of wellknown existing (DPM and DVFS) policies, each of which performs well for a set of conditions, and adapts at runtime to the best-performing policy for any given workload. Experiments are performed using state-of the-art DPM and DVFS policies and the results show that our proposed scheme adapts well to the changing workload and always achieves overall energy savings comparable to that of best-performing policy at any point in time.

This paper presents a run-time detection mechanism for access-driven cache-based Side-Channel Attacks (CSCAs) on Intel's x86 architecture. We demonstrate the detection capability and effectiveness of proposed mechanism on Prime+Probe attcks. The mechanism comprises of multiple machine learning models, which use real-time data from the HPCs for detection. Experiments are performed with two different implementations of AES cryptosystem while under Prime+Probe attack. We provide results under stringent design constraints such as: realistic system load conditions, real-time detection accuracy, speed, system-wide performance overhead and distribution of error (i.e., false positives and negatives) for the used machine learning models. Our results show detection accuracy of > 99% for Prime+Probe attack with performance overhead of 3 − 4% at the highest detection speed, i.e., within 1−2% completion of 4800 AES encryption rounds needed to complete a successful attack.

High resolution and stealthy attacks and their variants such as Flush+Reload, Flush+Flush, Prime+Probe, Spectre and Meltdown have completely exposed the vulnerabilities in Intel's computing architecture over the past few years. Mitigation techniques against such attacks are not very effective for two reasons: 1) Most mitigation techniques protect against a specific vulnerability and do not take a system-wide approach, and 2) they either completely remove or greatly reduce the performance benefits of resource sharing. In this work, we argue in favor of detection-based protection, which would help apply mitigation only after successful detection of the attack at runtime. As such, detection would serve as the first line of defense against such attacks. However, for a detection based protection strategy to be effective, detection needs to be highly accurate, to incur minimum system overhead at runtime, should cover a large set of attacks and be capable of early stage detection, i.e., at the very least before the attack is completed. We propose a machine learning based side-channel attack (SCA) detection tool, called WHISPER that satisfies the above mentioned design constraints. WHISPER uses multiple machine learning models in an Ensemble fashion to detect SCAs at runtime using behavioral data of concurrent processes, that are collected through hardware performance counters (HPCs). Through extensive experiments with different variants of state-of-the-art attacks, we demonstrate that the proposed tool is capable of detecting a large set of known attacks that target both computational and storage parts in computing systems. We present experimental evaluation of WHISPER against Flush+Reload, Flush+Flush, Prime+Probe, Spectre and Meltdown attacks. The results are provided under variable system load conditions and stringent evaluation metrics comprising detection accuracy, speed, system-wide performance overhead and distribution of error (i.e., False Positives & False Negatives). Our experiments show that WHISPER can detect a large and diverse attack vector with more than 99% accuracy at a reasonably low performance overhead.

No abstract