Search

Multivariate time series are used in many science and engineering domains, including health-care, astronomy, and high-performance computing. A recent trend is to use machine learning (ML) to process this complex data and these ML-based frameworks are starting to play a critical role for a variety of applications. However, barriers such as user distrust or difficulty of debugging need to be overcome to enable widespread adoption of such frameworks in production systems. To address this challenge, we propose a novel explainability technique, CoMTE, that provides counterfactual explanations for supervised machine learning frameworks on multivariate time series data. Using various machine learning frameworks and data sets, we compare CoMTE with several state-of-the-art explainability methods and show that we outperform existing methods in comprehensibility and robustness. We also show how CoMTE can be used to debug machine learning frameworks and gain a better understanding of the underlying multivariate time series data.

Effective power management in a data center is critical to ensure that power delivery constraints are met while maximizing the performance of users’ workloads. Power limiting is needed in order to respond to greater-than-expected power demand. HPC sites have generally tackled this by adopting one of two approaches: (1) a system-level power management approach that is aware of the facility or site-level power requirements, but is agnostic to the application demands; OR (2) a job-level power management solution that is aware of the application design patterns and requirements, but is agnostic to the site-level power constraints. Simultaneously incorporating solutions from both domains often leads to conflicts in power management mechanisms. This, in turn, affects system stability and leads to irreproducibility of performance. To avoid this irreproducibility, HPC sites have to choose between one of the two approaches, thereby leading to missed opportunities for efficiency gains.This paper demonstrates the need for the HPC community to collaborate towards seamless integration of system-aware and application-aware power management approaches. This is achieved by proposing a new dynamic policy that inherits the benefits of both approaches from tight integration of a resource manager and a performance-aware job runtime environment. An empirical comparison of this integrated management approach against state-of-the-art solutions exposes the benefits of investing in end-to-end solutions to optimize for system-wide performance or efficiency objectives. With our proposed system–application integrated policy, we observed up to 7% reduction in system time dedicated to jobs and up to 11% savings in compute energy, compared to a baseline that is agnostic to system power and application design constraints.

Demand response programs offer an opportunity for large power consumers to save on electricity costs by modulating their power consumption in response to demand changes in the electricity grid. Multiple types of such programs exist; for example, regulation service programs enable a consumer to bid for a sustainable amount of power draw over a time period, along with a reserve amount they are able to provide at request of the electricity service provider. Data centers offer unique capabilities to participate in these programs since they have significant capacity to modify their power consumption through workload scheduling and CPU power limiting. This paper proposes a novel power management policy and a bidding policy that enable data centers to participate in regulation service programs under real-world constraints. The power management policy schedules computing jobs and applies server power-capping under both the constraints of power programs and the constraints of job Quality-of-Service (QoS). Simulations with workload traces from a real data center show that the proposed policies enable data centers to meet both the requirement of regulation service programs and the QoS requirement of jobs. We demonstrate that, by applying our policies, data centers can save their electricity costs by 10% while abiding by all the QoS constraints in a real-world scenario.

Demand Response (DR) policies define the interactions between an energy supplier and its consumers and allow for customer energy regulation given a supplier request. Given the high flexibility and controllability of Data Centers (DC), they are promising candidates to participate in DR for power grid stabilization. In this work, we consider the setting where an energy supply deficit event occurs and must be addressed to avoid grid strain. We present two novel frameworks for DR, where a load aggregator offers price incentives to a set of consumer DCs so they can dynamically adjust their electricity consumption and provide DR to the grid via server usage reductions. Modeling DCs using realistic cost functions based on Quality of Service (QoS) requirements of the DC workloads, we present a data-driven inverse optimization method to estimate DC cost function parameters for precise and efficient pricing and provide an algorithm for solving the inverse problem. Experimental results on two test cases demonstrate the benefits of our proposed DR mechanisms for energy control.

Photonic network-on-chips, despite their low-latency and high-bandwidth-density advantages in large manycore systems, suffer from high power overhead. This overhead is further exacerbated by the high bandwidth demands of data-centric applications. Prior works utilize bandwidth allocation policies at system-level to minimize photonic power and provide required bandwidth for applications. We present an approach to minimize the bandwidth requirements by instrumenting an application at the software level. This instrumented information is used to assist bandwidth allocation at system-level, thereby reducing the photonic power. We instrument PageRank application and demonstrate 35% lower power using instrumentation-assisted bandwidth allocation on PageRank running real-world graphs compared to bandwidth allocation on uninstrumented PageRank.

In modern high-performance computing (HPC) systems, network congestion is an important factor that contributes to performance degradation. However, how network congestion impacts application performance is not fully understood. As Aries network, a recent HPC network architecture featuring a dragonfly topology, is equipped with network counters measuring packet transmission statistics on each router, these network metrics can potentially be utilized to understand network performance. In this work, by experiments on a large HPC system, we quantify the impact of network congestion on various applications' performance in terms of execution time, and we correlate application performance with network metrics. Our results demonstrate diverse impacts of network congestion: while applications with intensive MPI operations (such as HACC and MILC) suffer from more than 40% extension in their execution times under network congestion, applications with less intensive MPI operations (such as Graph500 and HPCG) are mostly not affected. We also demonstrate that a stall-to-flit ratio metric derived from Aries network counters is positively correlated with performance degradation and, thus, this metric can serve as an indicator of network congestion in HPC systems.

© 2020 EDAA. Future high-performance chips will require new cooling technologies that can extract heat efficiently. Two-phase cooling is a promising processor cooling solution owing to its high heat transfer rate and potential benefits in cooling power. Two-phase cooling mechanisms, including microchannel-based two-phase cooling or two-phase vapor chambers (VCs), are typically modeled by computing the temperature-dependent heat transfer coefficient (HTC) of the evaporator or coolant using an iterative simulation framework. Precomputed HTC correlations are specific to a given cooling system design and cannot be applied to even the same cooling technology with different cooling parameters (such as different geometries). Another challenge is that HTC correlations are typically calculated with computational fluid dynamics (CFD) tools, which induce long design and simulation times. This paper introduces a learning-based temperature-dependent HTC simulation framework that is used to model a two-phase cooling solution with a wide range of cooling design parameters. In particular, the proposed framework includes a compact thermal model (CTM) of two-phase VCs with hybrid wick evaporators (of nanoporous membrane and microchannels). We build a new simulation tool to integrate the proposed simulation framework and CTM. We validate the proposed simulation framework as well as the new CTM through comparisons against a CFD model. Our simulation framework and CTM achieve a speedup of 21 × with an average error of 0.98° C (and a maximum error of 2.59° C). We design an optimization flow for hybrid wicks to select the most beneficial hybrid wick geometries. Our flow is capable of finding a geometry- coolant combination that results in a lower (or similar) maximum chip temperature compared to that of the best coolant-geometry pair selected by grid search, while providing a speedup of 9.4 x.

The ability to repeat the execution of a program is a fundamental requirement in many areas of computing from computer system evaluation to software engineering. Reproducing executions of mobile apps, in particular, has proven difficult under real-life scenarios due to multiple sources of external inputs and interactive nature of the apps. Previous works that provide record/replay functionality for mobile apps are restricted to particular input sources (e.g., touchscreen events) and present deployment challenges due to intrusive modifications to the underlying software stack. Moreover, due to their reliance on record and replay of device specific events, the recorded executions cannot be reliably reproduced across different platforms. In this paper, we present a new practical approach, RandR, for record and replay of Android applications. RandR captures and replays multiple sources of input (i.e., UI and network) without requiring source code (OS or app), administrative device privileges, or any special platform support. RandR achieves these qualities by instrumenting a select set of methods at runtime within an application's own sandbox. In addition, to enable portability of recorded executions across different platforms for replay, RandR contextualizes UI events as interactions with particular UI components (e.g., a button) as opposed to relying on platform specific features (e.g., screen coordinates). We demonstrate RandR's accurate cross-platform record and replay capabilities using over 30 real-world Android apps across a variety of platforms including emulators as well as commercial off-the-shelf mobile devices deployed in real life.

Ultra-high power densities that are expected in future processors cannot be efficiently mitigated by conventional cooling solutions. Using two-phase vapor chambers (VCs) with micropillar wick evaporators is an emerging cooling technique that can effectively remove high heat fluxes through the evaporation process of a coolant. Two-phase VCs with micropillar wicks offer high cooling efficiency by leveraging a capillary-driven flow, where the coolant is passively driven by the wicking structure that eliminates the need for an external pump. Thermal models for such emerging cooling technologies are essential to evaluate their impact on future processors. Existing thermal models for two-phase VCs use computational fluid dynamics (CFD) modules, which incur long design and simulation times. This paper presents a fast and accurate compact thermal model for two-phase VCs with micropillar wicks. Our model achieves a maximum error of 1.25°C with a speedup of 214x in comparison to a CFD model. Using our proposed thermal model, we build an optimization flow that selects the best cooling solution and its cooling parameters to minimize the cooling power under a temperature constraint for a given processor and power profile. We then demonstrate our optimization flow on different chip sizes and hot spot distributions to choose the optimal cooling technique among VCs, microchannel-based two-phase cooling, liquid cooling via microchannels, and a hybrid cooling technique with thermoelectric coolers and liquid cooling with microchannels.

© 2019 IEEE High power densities lead to thermal hot spots in modern processors. These power densities are expected to reach kW/cm2 scale in future high-performance chips and this increase may significantly degrade performance and reliability, if not handled efficiently. Using two-phase vapor chambers (VCs) with micropillar wick evaporators is an emerging technique that removes heat through the evaporation process of a coolant and has the potential to remove high heat fluxes. In this cooling system, the coolant is supplied passively to the micropillar wick via capillary pumping, eliminating the need for an external pump and ensuring stable thin-film flow. Evaluation of such an emerging cooling technique on realistic chip power densities and micropillar geometries necessitates accurate and fast thermal models. Although multi-physics simulators based on either finite-element or finite-volume methods are highly accurate, they have long design and simulation times. This paper introduces a novel compact thermal model capable of simulating two-phase vapor chambers with micropillar wick evaporators. In comparison to COMSOL, our model shows a competitively low error of 1.25°C and a 214x speedup. We also present a comparison of the cooling performance of different cooling techniques such as a conventional heat sink, liquid cooling via microchannels, hybrid cooling using thermoelectric coolers and liquid cooling via microchannels, and two-phase VCs with micropillar wick evaporators for the first time. Based on our observations, two-phase VCs and microchannel-based two-phase cooling show better cooling performance for hot spot power densities of less than 1500 W/cm2, while hybrid cooling achieves lower hot spot temperature and thermal gradients for hot spot power densities between 1500 and 2000 W/cm2

Photonic Network-on-Chips (PNoCs) offer promising benefits over Electrical Network-on-Chips (ENoCs) in many-core systems owing to their lower latencies, higher bandwidth, and lower energy-per-bit communication with negligible data-dependent power. These benefits, however, are limited by a number of challenges. Microring resonators (MRRs) that are used for photonic communication have high sensitivity to process variations and on-chip thermal variations, giving rise to possible resonant wavelength mismatches. State-of-the-art microheaters, which are used to tune the resonant wavelength of MRRs, have poor efficiency resulting in high thermal tuning power. In addition, laser power and high static power consumption of drivers, serializers, comparators, and arbitration logic partially negate the benefits of the sub-pJ operating regime that can be obtained with PNoCs. To reduce PNoC power consumption, this paper introduces WAVES, a wavelength selection technique to identify and activate the minimum number of laser wavelengths needed, depending on an application’s bandwidth requirement. Our results on a simulated 2.5D manycore system with PNoC demonstrate an average of 23% (resp. 38%) reduction in PNoC power with only

In a cluster that is shared by many users, jobs often need to wait in the queue for a significant amount of time. Much research has been done to reduce this time with scheduling, including aggressive back-filling strategies and sharing nodes among different jobs. Although most resources are shared to some extent in HPC clusters, it is somewhat surprising that a well-known technique used on commercial clouds, i.e., oversubscribing nodes so that CPU cores are shared among jobs, is rather rare. This is partially due to concerns about interference. This paper presents Tangram, a framework for colocating applications in HPC clusters. Tangram uses prior knowledge of applications, such as whether they are I/O or CPU intensive, to predict whether potential colocations improve overall performance. To predict with sufficient accuracy, Tangram uses a combination of performance counter measurements, knowledge of past colocation performance, and machine learning. We show that Tangram can choose colocations to reduce makespan by 19% on average and by 55% in the best case, while limiting the performance degradation caused by colocation from 1598% to 26% in the worst case.

While state-of-the-art system-level simulators can deliver swift estimation of power dissipation for microprocessor designs, they do so at the expense of reduced accuracy. On the other hand, RTL simulators are typically cycle-accurate but overwhelmingly time consuming for real-life workloads. Consequently, the design community often has to make a compromise between accuracy and speed. In this work, we propose a novel cross-layer approach that can enable accurate power estimation by carefully integrating components from system-level and RTL simulation of the target design. We first leverage the concept of simulation points to transform the workload application and isolate its most critical segments. We then profile the highest weighted simulation point (HWSP) with a RTL simulator (AnyCore) for maximum accuracy, while the rest are simulated with a system-level simulator (gem5) for ensuring fast evaluation. Finally, we combine the integrated set of profiling data as input to the power simulator (McPAT). Our evaluation results for three different SPEC2006 benchmark applications demonstrate that our proposed cross-layer framework can improve the power estimation accuracy by up to 15% for individual simulation points and by ~9% for the full application, compared to that of a conventional system-level simulation scheme.

Modern cloud applications are designed in a highly configurable way to ensure increased reusability and portability. With the growing complexity of these applications, configuration errors (i.e., misconfigurations) have become major sources of service outages and disruptions. While some research has so far focused on detecting errors in configurations that are represented as well-structured key-value pairs, the configurations of cloud applications are typically stored in text files with application-specific syntax and in unlabeled file system locations, limiting the use of existing error detection tools. This paper introduces ConfEx, a framework that enables discovery and extraction of text-based configurations in multi-tenant cloud platforms and cloud image repositories for configuration analysis and validation. ConfEx uses a novel vocabulary-based technique to identify text-based configuration files in cloud system instances with unlabeled content, and leverages existing configuration parsers to extract the information in these files. We show that ConfEx achieves over 98% precision and recall in identifying configuration files on 3893 popular Docker Hub images and we also demonstrate a use case of ConfEx for detecting injected misconfigurations via outlier analysis.

The dragonfly network topology has attracted attention in recent years owing to its high radix and constant diameter. However, the influence of job allocation on communication time in dragonfly networks is not fully understood. Recent studies have shown that random allocation is better at balancing the network traffic, while compact allocation is better at harnessing the locality in dragonfly groups. Based on these observations, this paper introduces a novel allocation policy called Level-Spread for dragonfly networks. This policy spreads jobs within the smallest network level that a given job can fit in at the time of its allocation. In this way, it simultaneously harnesses node adjacency and balances link congestion. To evaluate the performance of Level-Spread, we run packet-level network simulations using a diverse set of application communication patterns, job sizes, and communication intensities. We also explore the impact of network properties such as the number of groups, number of routers per group, machine utilization level, and global link bandwidth. Level-Spread reduces the communication overhead by 16% on average (and up to 71%) compared to the state-of-the-art allocation policies.

While the growth of cloud-based technologies has benefited the society tremendously, it has also increased the surface area for cyber attacks. Given that cloud services are prevalent today, it is critical to devise systems that detect intrusions. One form of security breach in the cloud is when cyber-criminals compromise Virtual Machines (VMs) of unwitting users and, then, utilize user resources to run time-consuming, malicious, or illegal applications for their own benefit. This work proposes a method to detect unusual resource usage trends and alert the user and the administrator in real time. We experiment with three categories of methods: simple statistical techniques, unsupervised classification, and regression. So far, our approach successfully detects anomalous resource usage when experimenting with typical trends synthesized from published real-world web server logs and cluster traces. We observe the best results with unsupervised classification, which gives an average F1-score of 0.83 for web server logs and 0.95 for the cluster traces.

Network messaging delay historically constitutes a large portion of the wall-clock time for High Performance Computing (HPC) applications, as these applications run on many nodes and involve intensive communication among their tasks. Dragonfly network topology has emerged as a promising solution for building exascale HPC systems owing to its low network diameter and large bisection bandwidth. Dragonfly includes local links that form groups and global links that connect these groups via high bandwidth optical links. Many aspects of the dragonfly network design are yet to be explored, such as the performance impact of the connectivity of the global links, i.e., global link arrangements, the bandwidth of the local and global links, or the job allocation algorithm. This paper first introduces a packet-level simulation framework to model the performance of HPC applications in detail. The proposed framework is able to simulate known MPI (message passing interface) routines as well as applications with custom-defined communication patterns for a given job placement algorithm and network topology. Using this simulation framework, we investigate the coupling between global link bandwidth and arrangements, communication pattern and intensity, job allocation and task mapping algorithms, and routing mechanisms in dragonfly topologies. We demonstrate that by choosing the right combination of system settings and workload allocation algorithms, communication overhead can be decreased by up to 44%. We also show that circulant arrangement provides up to 15% higher bisection bandwidth compared to the other arrangements, but for realistic workloads, the performance impact of link arrangements is less than 3%.

To track security and compliance requirements and perform problem diagnosis, administrators of cloud computing systems need to monitor significant system changes occurring on the set of cloud instances under their supervision. Considering the large number of instances (virtual machines, containers) possibly operating under multiple configurations, this is a difficult-to-track process. Standard solutions to this problem rely on manually-created rules to identify changes. These techniques suffer from a limited scope, rely on domain expertise, and are time-consuming and error-prone. Recently, more streamlined approaches that automatically determine the type of individual system changes have been proposed, but these techniques assume that system states right before and after each individual change can be captured, a rather difficult requirement to enforce in real world usage. This paper proposes DeltaSherlock, a practical system change discovery framework that can capture system states on-demand and detect multiple system changes between them. We evaluate DeltaSherlock over 25,000 system changes caused by software installations collected from virtual machines (VMs) deployed over a commercial cloud. DeltaSherlock can accurately identify multiple software installations with 96.8% accuracy when supplied with a non-overlapping record of system changes and with 77.8% accuracy when supplied with random irregular observations possibly containing overlapping or incomplete changes.

Data centers have shown great opportunities to participate in extensive demand response programs in recently years. This paper specifically focuses on data centers as participants in regulation service reserves (RSR) power market. We propose a novel approach to model the dynamics of the job processing Quality of Service (QoS) in data centers that offer RSR, and use stochastic dynamic programming (DP) to solve for the optimal reserve deployment policies. We show that the job QoS degradation can be modeled as a time varying probability distribution function (PDF) whose mean and variance evolve as functions of recent control statistics. The mean and variance are in fact additional state variables or sufficient statistics of the stochastic DP whose solution provides the data center operator (DCO) decision supports to minimize the average operating costs associated with RSR signal tracking error and job processing QoS degradation. Simulation results show that the feedback control policy obtained from the stochastic DP solution can reduce the DCO's operating costs compared to heuristic operating protocols reported in the literature. In addition, the DP value function can assist the DCO to bid optimally into the hour-ahead joint energy and reserve market.