Jobs In Data Centers Research Articles

Component-distinguishable Co-location and Resource Reclamation for High-throughput Computing

Cloud service providers improve resource utilization by co-locating latency-critical (LC) workloads with best-effort batch (BE) jobs in datacenters. However, they usually treat multi-component LCs as monolithic applications and treat BEs as “second-class citizens” when allocating resources to them. Neglecting the inconsistent interference tolerance abilities of LC components and the inconsistent preemption loss of BE workloads can result in missed co-location opportunities for higher throughput. We present Rhythm , a co-location controller that deploys workloads and reclaims resources rhythmically for maximizing the system throughput while guaranteeing LC service’s tail latency requirement. The key idea is to differentiate the BE throughput launched with each LC component, that is, components with higher interference tolerance can be deployed together with more BE jobs. It also assigns different reclamation priority values to BEs by evaluating their preemption losses into a multi-level reclamation queue. We implement and evaluate Rhythm using workloads in the form of containerized processes and microservices. Experimental results show that it can improve the system throughput by 47.3%, CPU utilization by 38.6%, and memory bandwidth utilization by 45.4% while guaranteeing the tail latency requirement.

This site is a dead end? Employment uncertainties and labor in data centers

Would technological changes increase the need for human workers or eliminate them altogether? This uncertainty has produced an unresolved tension, from the industrial revolution to the rise of the information society. The data center industry has been largely invisible in public debates about this question. Yet the same tensions exist within the industry itself: Will automation create data center jobs or kill them? In this article, we work inside the “black box” – the data center, to examine uncertainties faced by those who work there. We do so through interviews and observations, first, of data center managers and executives at international trade expos, where anxieties about the shortage of data center workers but also their irrelevance were palpable. Then, we turn to a remote data center in Finland, where security guards and technical operators negotiate employment uncertainties through the biopolitics of their labor. In both sites, the uncertainties about data center employment are manifest and embodied, even if they are expressed and experienced in different ways. On both the top and bottom levels of data center hierarchies, people are discomfited by the possibility of their own redundancy. At the same time, they present the sunnier sides of data center work when they talked about their efforts to resolve ongoing issues of worker shortage, the lack of diversity in data centers, and the routines that could easily slide into boredom or anomie. We situate our findings on the long arc of capitalist transformations and discuss the insights they might provide for today’s data-driven economy in general.

Collaborative Accelerators for Streamlining MapReduce on Scale-up Machines With Incremental Data Aggregation

The MapReduce programming paradigm has been increasingly adopted to implement data-intensive applications processing both small and large scale datasets. As most jobs in data centers have a data footprint in the order of gigabytes, emerging high-end scale-up machines are capable of running most data center processing tasks, thus significantly improving power and server density. However, this approach provides limited performance and energy efficiency because of inefficient utilization of the memory subsystem and serial execution within the MapReduce programming model. Recent work has proposed a distributed hardware acceleration architecture, called CASM, which augments each core in a scale-up machine with a lightweight compute engine. The CASM's network of accelerators operates concurrently with the cores in executing MapReduce stages and reduces significantly traffic to/from storage. In this article, we study the benefits and applicability of CASM, by offering an extensive analysis of design parameters and of its scalable performance on a wide range of applications, and exploring its applicability to incremental data aggregation tasks. Our experimental evaluation indicates that CASM reduces off-chip traffic by four times on average over a chip multiprocessor solution, while scaling well with the number of cores in the system, and it is highly effective in providing incremental results that approximate final outcomes.

Shuffle Scheduling for MapReduce Jobs Based on Periodic Network Status

MapReduce jobs need to shuffle a large amount of data over the network between mapper and reducer nodes. The shuffle time accounts for a big part of the total running time of the MapReduce jobs. Therefore, optimizing the makespan of shuffle phase can greatly improve the performance of MapReduce jobs. A large fraction of production jobs in data centers are recurring with predictable characteristics, and the recurring jobs split the network into periodic busy and idle time slots, which allows us to better schedule the shuffle data in order to reduce the makespan of shuffle phase with the future predictable network status available. In this paper, we formulate the shuffle scheduling problem with the aim to minimize the makespan of MapReduce shuffle phase by leveraging the predictable periodic network status. We then propose a simple yet effective network-aware shuffle scheduling algorithm (NAS) to reduce the number of idle time slots required to transfer the shuffle data so as to reduce the shuffle makespan. We also prove that the proposed algorithm NAS is a $\frac {3}{2}$ -approximation algorithm to the shuffle scheduling problem when all the future idle time slots have the same duration. We finally conduct experiments through simulations. Experimental results demonstrate the proposed algorithm can effectively reduce the makespan of MapReduce shuffle phase and increase network utilization.

Managing Energy Plus Performance in Data Centers and Battery-Based Devices Using an Online Non-Clairvoyant Speed-Bounded Multiprocessor Scheduling

An efficient scheduling reduces the time required to process the jobs, and energy management decreases the service cost as well as increases the lifetime of a battery. A balanced trade-off between the energy consumed and processing time gives an ideal objective for scheduling jobs in data centers and battery based devices. An online multiprocessor scheduling multiprocessor with bounded speed (MBS) is proposed in this paper. The objective of MBS is to minimize the importance-based flow time plus energy (IbFt+E), wherein the jobs arrive over time and the job’s sizes are known only at completion time. Every processor can execute at a different speed, to reduce the energy consumption. MBS is using the tradition power function and bounded speed model. The functioning of MBS is evaluated by utilizing potential function analysis against an offline adversary. For processors m ≥ 2, MBS is O(1)-competitive. The working of a set of jobs is simulated to compare MBS with the best known non-clairvoyant scheduling. The comparative analysis shows that the MBS outperforms other algorithms. The competitiveness of MBS is the least to date.

Sequence-to-sequence models for workload interference prediction on batch processing datacenters

Co-scheduling of jobs in data centers is a challenging scenario where jobs can compete for resources, leading to severe slowdowns or failed executions. Efficient job placement on environments where resources are shared requires awareness on how jobs interfere during execution, to go far beyond ineffective resource overbooking techniques.Current techniques, most of which already involve machine learning and job modeling, are based on workload behavior summarization over time, rather than focusing on effective job requirements at each instant of the execution. In this work, we propose a methodology for modeling co-scheduling of jobs on data centers, based on their behavior towards resources and execution time and using sequence-to-sequence models based on recurrent neural networks. The goal is to forecast co-executed jobs footprint on resources throughout their execution time, from the profile shown by the individual jobs, in order to enhance resource manager and scheduler placement decisions.The methods presented herein are validated by using High Performance Computing benchmarks based on different frameworks (such as Hadoop and Spark) and applications (CPU bound, IO bound, machine learning, SQL queries...). Experiments show that the model can correctly identify the resource usage trends from previously seen and even unseen co-scheduled jobs.

MILP formulations for spatio-temporal thermal-aware scheduling in Cloud and HPC datacenters

This paper focuses on scheduling problems related to the execution of computational jobs in datacenters with thermal constraints. Mixed integer linear programming (MILP) formulations are proposed that encompass both spatial and temporal aspects of the temperature evolution under a unified model. This model takes into account the dynamics of heat production and dissipation in order to schedule jobs at appropriate times on appropriate machines. The proposed MILP formulations are applicable to both high-performance computing (HPC) and Cloud settings, and can target several objectives including energy and makespan minimization, while incorporating the cooling costs and dynamic voltage and frequency scaling capabilities of servers. The applicability and usefulness of our formulations are demonstrated via several HPC and Cloud case-studies.

Low Complexity Hierarchical Scheduling for Diverse Datacenter Jobs

Datacenter jobs in a private cloud are often generated by different departments or groups of a business or other organization. In order to reflect the organizational structures of datacenter workloads, it is desirable for the datacenter scheduler to support hierarchical scheduling. The state-of-the-art hierarchical scheduling scheme, hierarchical dominant resource fairness (H-DRF), however, is expensive to implement. Specifically, the worst-case time complexity of H-DRF is $O(q^{2})$ , where $q$ is dominated by the number of jobs in the hierarchy, which is a large complexity in a practical large-scale datacenter. In this letter, we design a new online hierarchical fair scheduling scheme, multi-resource collapsed hierarchies (MCH), in the way of converting the hierarchy into a weighted flat tree. MCH takes at most $O(q)$ time and is simple enough to implement in practice. Simulations driven by Google cluster traces show that MCH consumes a nearly sqrt of the total run time of H-DRF, improving significantly in saving the computation time.

Nash Bargaining Solution Allocation is Not Suitable for Datacenter Jobs

The Nash Bargaining Solution (NBS) has been broadly suggested as an effective solution for the problem of fair allocation of multiple resources, namely bandwidth allocation in datacenters. In spite of being thoroughly studied, and provably strategy-proof for most scenarios, NBS-based allocation methods lack research on the strategic behavior of tenants in the case of proportionality of resource demands, which is common in datacenter workloads. We found that misbehavior is beneficial: by lying about bandwidth demands tenants can improve their allocations. We show that a sequence of selfish improvements leads to trivial demand vectors for all tenants. It essentially removes sharing incentives which are very important for datacenter networks. In this paper, we analytically prove that tenants can misbehave in 2- and 3- tenants cases. We show that misbehavior is possible in one recently proposed NBS-based allocation system if proportionality of demands is taken into account. Monte Carlo simulations were done for 2–15 tenants to show a misbehavior possibility and its impact on aggregated bandwidth. We propose to use another game-theoretic approach, namely Dominant Resource Fairness (DRF) to allocate bandwidth in the case of proportional demands. We show that this method performs significantly better than NBS after misbehavior.

Understanding and improving JVM GC work stealing at the data center scale

Garbage collection (GC) is a critical part of performance in managed run-time systems such as the OpenJDK Java Virtual Machine (JVM). With a large number of latency sensitive applications written in Java the performance of the JVM is essential. Java application servers run in data centers on a large number of multi-core servers, thus load balancing in multi-threaded GC phases is critical. Dynamic load balancing in the JVM GC is achieved through work stealing, a well known and effective method to balance tasks across threads. This paper analyzes the JVM work stealing behaviour, and introduces a novel work stealing technique that improves performance, GC CPU utilization, scalability, and reduces the cost of jobs running on Google’s data-centers. We analyze both the DaCapo benchmark suite as well as Google’s data-center jobs. Our results show that the Gmail front-end server shows a 15-20% GC CPU reduction, and a 5% CPU performance improvement. Our analysis of a sample of ~59K jobs shows that GC CPU utilization improves by 38% geomean, 12% weighted geomean. GC pause time improves by 16% geomean, 20% weighted geomean. Full GC pause time improves by 34% geomean, 12% weighted geomean.

Towards Multi-Resource Fair Allocation with Placement Constraints

Multi-resource fair schedulers have been widely implemented in compute clusters to provide service isolation guarantees. Existing multi-resource sharing policies, notably Dominant Resource Fairness (DRF) and its variants, are designed for unconstrained jobs that can run on all machines in a cluster. However, an increasing number of datacenter jobs specify placement constraints and can only run on a particular class of machines meeting specific hardware/software requirements (e.g., GPUs or a particular kernel version). We show that directly extending existing policies to constrained jobs either compromises isolation guarantees or allows users to gain more resources by deceiving the scheduler. It remains unclear how multi-resource fair sharing is defined and achieved in the presence of placement constraints. We address this open problem by a new sharing policy, called Task Share Fairness (TSF), that provides provable isolation guarantees and is strategy-proof against gaming the allocation policy. TSF is shown to be envy-free and Pareto optimal as well.

Profiling a warehouse-scale computer

With the increasing prevalence of warehouse-scale (WSC) and cloud computing, understanding the interactions of server applications with the underlying microarchitecture becomes ever more important in order to extract maximum performance out of server hardware. To aid such understanding, this paper presents a detailed microarchitectural analysis of live datacenter jobs, measured on more than 20,000 Google machines over a three year period, and comprising thousands of different applications. We first find that WSC workloads are extremely diverse, breeding the need for architectures that can tolerate application variability without performance loss. However, some patterns emerge, offering opportunities for co-optimization of hardware and software. For example, we identify common building blocks in the lower levels of the software stack. This "datacenter tax" can comprise nearly 30% of cycles across jobs running in the fleet, which makes its constituents prime candidates for hardware specialization in future server systems-on-chips. We also uncover opportunities for classic microarchitectural optimizations for server processors, especially in the cache hierarchy. Typical workloads place significant stress on instruction caches and prefer memory latency over bandwidth. They also stall cores often, but compute heavily in bursts. These observations motivate several interesting directions for future warehouse-scale computers.

Near-Optimal Scheduling Mechanisms for Deadline-Sensitive Jobs in Large Computing Clusters

We consider a market-based resource allocation model for batch jobs in cloud computing clusters. In our model, we incorporate the importance of the due date of a job rather than the number of servers allocated to it at any given time. Each batch job is characterized by the work volume of total computing units (e.g., CPU hours) along with a bound on maximum degree of parallelism. Users specify, along with these job characteristics, their desired due date and a value for finishing the job by its deadline. Given this specification, the primary goal is to determine the scheduling of cloud computing instances under capacity constraints in order to maximize the social welfare (i.e., sum of values gained by allocated users). Our main result is a new (CC-kċss-1)-approximation algorithm for this objective, where C denotes cloud capacity, k is the maximal bound on parallelized execution (in practical settings, k < C ) and s is the slackness on the job completion time, that is, the minimal ratio between a specified deadline and the earliest finish time of a job. Our algorithm is based on utilizing dual fitting arguments over a strengthened linear program to the problem. Based on the new approximation algorithm, we construct truthful allocation and pricing mechanisms, in which reporting the true value and other properties of the job (deadline, work volume, and the parallelism bound) is a dominant strategy for all users. To that end, we extend known results for single-value settings to provide a general framework for transforming allocation algorithms into truthful mechanisms in domains of single-value and multi-properties. We then show that the basic mechanism can be extended under proper Bayesian assumptions to the objective of maximizing revenues, which is important for public clouds. We empirically evaluate the benefits of our approach through simulations on data-center job traces, and show that the revenues obtained under our mechanism are comparable with an ideal fixed-price mechanism, which sets an on-demand price using oracle knowledge of users’ valuations. Finally, we discuss how our model can be extended to accommodate uncertainties in job work volumes, which is a practical challenge in cloud settings.

SEED: solar energy‐aware efficient scheduling for data centers

SUMMARYIt is well known that data centers are consuming a large amount of energy that incurs significant financial and environmental costs. Recently, there has been an increasing interest in utilizing green energy for data centers, where green energy sources include solar and wind. This paper studies the crucial problem of maximizing the utilization of green energy through scheduling complex jobs in data centers in order to reduce the use of traditional brown energy. However, it is highly challenging for data centers to make use of green energy. First, the availability of typical green energy is variable to dynamic changes of natural environments, for example, weather. Second, although predictions can be made for the future availability of green energy, it is inevitable that such predictions have errors. Third, jobs are associated with strict deadlines, and it is required that jobs are completed before their deadlines. Finally, because the reliability in a data center relies upon temperature, the awareness of temperature should be taken into account while maximizing the green energy. In this paper, we consider online scheduling of jobs whose arrivals to the data center system dynamically. In addition, we explicitly take the power consumption of switches into account when scheduling jobs onto computing nodes. Two solar energy‐aware algorithms called SEEDMin and SEEDMax have been proposed. Then, we extend SEED to RSEED with the awareness of reliability. To evaluate the effectiveness of the proposed algorithms, comprehensive simulations have been conducted, and the proposed algorithms are compared with other state‐of‐art algorithms. Experimental results demonstrate that both SEEDMin and SEEDMax can significantly increase the utilization of solar energy without violating job deadlines and overall energy budget. The amount of solar energy utilized by SEEDMin and SEEDMax is 33.4%and35.3% larger than that of two traditional scheduling algorithms, MinMin and MinMax, respectively. Also, it can be seen that RSEED greatly improves the reliability by decreasing the temperature. Copyright © 2013 John Wiley & Sons, Ltd.

Utilizing green energy prediction to schedule mixed batch and service jobs in data centers

As brown energy costs grow, renewable energy becomes more widely used. Previous work focused on using immediately available green energy to supplement the non-renewable, or brown energy at the cost of canceling and rescheduling jobs whenever the green energy availability is too low [16]. In this paper we design an adaptive data center job scheduler which utilizes short term prediction of solar and wind energy production. This enables us to scale the number of jobs to the expected energy availability, thus reducing the number of cancelled jobs by 4x and improving green energy usage efficiency by 3x over just utilizing the immediately available green energy.