Real-world Workloads Research Articles

Energy-efficient adaptive networked datacenters for the QoS support of real-time applications

In this paper, we develop the optimal minimum-energy scheduler for the adaptive joint allocation of the task sizes, computing rates, communication rates and communication powers in virtualized networked data centers (VNetDCs) that operate under hard per-job delay-constraints. The considered VNetDC platform works at the Middleware layer of the underlying protocol stack. It aims at supporting real-time stream service (such as, for example, the emerging big data stream computing (BDSC) services) by adopting the software-as-a-service (SaaS) computing model. Our objective is the minimization of the overall computing-plus-communication energy consumption. The main new contributions of the paper are the following ones: (i) the computing-plus-communication resources are jointly allotted in an adaptive fashion by accounting in real-time for both the (possibly, unpredictable) time fluctuations of the offered workload and the reconfiguration costs of the considered VNetDC platform; (ii) hard per-job delay-constraints on the overall allowed computing-plus-communication latencies are enforced; and, (iii) to deal with the inherently nonconvex nature of the resulting resource optimization problem, a novel solving approach is developed, that leads to the lossless decomposition of the afforded problem into the cascade of two simpler sub-problems. The sensitivity of the energy consumption of the proposed scheduler on the allowed processing latency, as well as the peak-to-mean ratio (PMR) and the correlation coefficient (i.e., the smoothness) of the offered workload is numerically tested under both synthetically generated and real-world workload traces. Finally, as an index of the attained energy efficiency, we compare the energy consumption of the proposed scheduler with the corresponding ones of some benchmark static, hybrid and sequential schedulers and numerically evaluate the resulting percent energy gaps.

FLAP: Flash-aware Prefetching for Improving SSD-based Disk Cache

In modern enterprise storage systems, there is a trend that using NAND flash based solid state disks (SSDs) as a second-level disk cache to reduce the slow access to hard disk drives (HDDs) by caching the hot data of HDDs with SSDs. However, using SSDs for both caching and prefetching has rarely been discussed due to the performance penalty caused by unsuccessful prefetching, including garbage collection cost and disk bandwidth wasting. In this paper, a FLash-Aware Prefetching (FLAP) scheme has been presented to provide aggressive prefetching for improving the performance of sequential disk accesses. FLAP has been evaluated using well known real-world workloads. Experiment results show that FLAP can offer performance improvement under sequential workloads compared with pure caching, while reducing the internal garbage collection of SSDs

Stochastic modeling of dynamic right-sizing for energy-efficiency in cloud data centers

Large data centers are usually built to support increasing computational and data storage demand of growing global business and industry, which consume an enormous amount of energy, at a huge cost to both business and the environment. However, much of that energy is wasted to maintain excess service capacity during periods of low load. In this paper, we investigate the problem of “right-sizing” data center for energy-efficiency through virtualization which allows consolidation of workloads into smaller number of servers while dynamically powering off the idle ones. In view of the dynamic nature of data centers, we propose a stochastic model based on Queueing theory to capture the main characteristics. Solving this model, we notice that there exists a tradeoff between the energy consumption and performance. We hereby develop a BFGS based algorithm to optimize the tradeoff by searching for the optimal system parameter values for the data center operators to “right-size” the data centers. We implement our Stochastic Right-sizing Model (SRM) and deploy it in the real-world cloud data center. Experiments with two real-world workload traces show that SRM can significantly reduce the energy consumption while maintaining high performance.

Improving Performance and Capacity of Flash Storage Devices by Exploiting Heterogeneity of MLC Flash Memory

The multi-level cell (MLC) NAND flash memory technology enables multiple bits of information to be stored in a memory cell, thus making it possible to increase the density of flash memory without increasing the die size. In MLC NAND flash memory, each memory cell can be programmed as a single-level cell or a multi-level cell at runtime because of its performance/capacity asymmetric programming property, which is called flexible programming in this paper. Therefore, MLC flash memory has a potential to achieve the high performance of SLC flash memory while preserving its maximum capacity. In this paper, we present a flexible flash file system, called FlexFS, which takes advantage of flexible programming. FlexFS divides a flash memory medium into SLC and MLC regions, and then dynamically changes two different types of regions to provide an optimal storage solution to end-users in terms of performance and capacity. FlexFS also provides a reasonable storage lifetime by managing the wearing rate of NAND flash memory, which is accelerated by the use of flexible programming. Our implementation of FlexFS in the Linux 2.6 kernel shows that it achieves the I/O performance comparable to SLC flash memory while guaranteeing the capacity of MLC flash memory in various real-world workloads.

Using Worker Quality Scores to Improve Stopping Rules

We consider the crowdsourcing task of learning the answer to simple multiple-choice microtasks. In order to provide statistically significant results, one often needs to ask multiple workers to answer the same microtask. A stopping rule is an algorithm that for a given microtask decides for any given set of worker answers if the system should stop and output an answer or iterate and ask one more worker. A quality score for a worker is a score that reflects the historic performance of that worker. In this paper we investigate how to devise better stopping rules given such quality scores. We conduct a data analysis on a large-scale industrial crowdsourcing platform, and use the observations from this analysis to design new stopping rules that use the workers’ quality scores in a non-trivial manner. We then conduct a simulation based on a real-world workload, showing that our algorithm performs better than the more naive approaches.

Simulation and optimization of HPC job allocation for jointly reducing communication and cooling costs

Performance and energy are critical aspects in high performance computing (HPC) data centers. Highly parallel HPC applications that require multiple nodes usually run for long durations in the range of minutes, hours or days. As the threads of parallel applications communicate with each other intensively, the communication cost of these applications has a significant impact on data center performance. Energy consumption has also become a first-order constraint of HPC data centers. Nearly half of the energy in the computing clusters today is consumed by the cooling infrastructure. Existing job allocation policies either target improving the system performance or reducing the cooling energy cost of the server nodes. How to optimize the system performance while minimizing the cooling energy consumption is still an open question. This paper proposes a job allocation methodology aimed at jointly reducing the communication cost and the cooling energy of HPC data centers. In order to evaluate and validate our optimization algorithm, we implement our joint job allocation methodology in the structural simulation toolkit (SST) – a simulation framework for large-scale data centers. We evaluate our joint optimization algorithm using traces extracted from real-world workloads. Experimental results show that, in comparison to performance-aware job allocation algorithms, our algorithm achieves comparable running times and reduces the cooling power by up to 42.21% across all the jobs.

Phase change memory in enterprise storage systems

Storage devices based on Phase Change Memory (PCM) devices are beginning to generate considerable attention in both industry and academic communities. But whether the technology in its current state will be a commercially and technically viable alternative to entrenched technologies such as flash-based SSDs still remains unanswered. To address this it is important to consider PCM SSD devices not just from a device standpoint, but also from a holistic perspective. This paper presents the results of our performance measurement study of a recent all-PCM SSD prototype. The average latency for 4 KB random read is 6.7 ?s, which is about 16x faster than a comparable eMLC flash SSD. The distribution of I/O response times is also much narrower than the flash SSD for both reads and writes. Based on real-world workload traces, we model a hypothetical storage device which consists of flash, HDD, and PCM to identify the combinations of device types that offer the best performance within cost constraints. Our results show that - even at current price points - PCM storage devices show promise as a new component in multi-tiered enterprise storage systems.

Reducing excessive journaling overhead with small-sized NVRAM for mobile devices

Journaling techniques are widely used to guarantee file system consistency of battery-powered mobile devices such as smartphones and tablets. In a journaling file system, system recovery is facilitated by first writing updated data blocks to a journal area and then periodically writing them to their home locations. However, these duplicated writes degrade the performance and shorten the lifetime of NAND flash storage in mobile devices. In particular, a lightweight database library, which is mainly used to manage application data in mobile devices, is a major cause of excessive journaling because it frequently triggers the costly file synchronization to guarantee the atomicity of transactional execution and thus generates a significant amount of synchronous random write traffic. This paper presents a novel journaling scheme, called Delta Journaling (DJ), to resolve this problem efficiently by using small-sized nonvolatile random access memory (NVRAM). DJ is based on a unique update pattern found in mobile devices, where file system updates are mostly very small. By exploiting the byteaddressable and the nonvolatile characteristics of NVRAM, DJ stores a journal block as a compressed delta in the smallsized NVRAM only when the compressed delta is small enough. Experimental results show that DJ outperforms a traditional journaling file system by up to 16.8 times for synthetic workloads. For a real-world workload, it enhances transaction throughput by 25.5% and 29.2% in ordered and journal modes, respectively, with only 16 MB NVRAM. Also, DJ enhances the lifetime of NAND flash storage by eliminating almost all journal writes without any loss of reliability.

Maximizing the revenues of data centers in regulation market by coordinating with electric vehicles

Frequency regulation is a major market service to reduce the undesired imbalance between power supply and demand in the power market. In order to participate in the regulation market, both the supply and demand sides need to be capable of flexibly adjusting their power generation and consumption, respectively. As the scale of Internet data centers is increasing rapidly, their significant power consumption has enabled them to become an important player in the regulation market for maximized profits and thus minimized operating expenses. On the other side, Plug-in Hybrid Electric Vehicles (PHEVs) have also recently been identified as a major participant in the regulation market, due to their large power demand for battery charging.In this paper, we propose a novel power management scheme that jointly leverages a data center and its employees’ PHEVs to (1) maximize the revenues that the data center receives from the regulation market and (2) get the PHEVs charged at no expense to their owners. Our scheme features a two-level hierarchical power control design. At the first level, our scheme interacts with the regulation market to provide information about the data center power consumption on an hourly basis. At the second level, the scheme decides the power budgets for the servers and UPS in the data center, as well as PHEVs, in real time, to follow the given regulation signal. In addition, we show how to leverage the thermal energy storage (TES) tanks available in many data centers to adapt the cooling power consumption for better management of the data center power demand and further increased regulation revenues. We evaluate the proposed scheme with real-world workload and regulation traces. The results show that our scheme performs a high-quality regulation service. As a result, the proposed scheme outperforms several commonly used baselines by having higher regulation revenues, and so lower operating expenses, for the data center. Finally, we analyze the cost savings of the PHEV owners, throughout the lifetime of the PHEVs, by getting their batteries charged at no expense.

Analytic Models of SSD Write Performance

Solid-state drives (SSDs) update data by writing a new copy, rather than overwriting old data, causing prior copies of the same data to be invalidated . These writes are performed in units of pages , while space is reclaimed in units of multipage erase blocks , necessitating copying of any remaining valid pages in the block before reclamation. The efficiency of this cleaning process greatly affects performance under random workloads; in particular, in SSDs, the write bottleneck is typically internal media throughput, and write amplification due to additional internal copying directly reduces application throughput. We present the first nearly-exact closed-form solution for write amplification under greedy cleaning for uniformly-distributed random traffic, validate its accuracy via simulation, and show that its inaccuracies are negligible for reasonable block sizes and overprovisioning ratios. In addition, we also present the first models which predict performance degradation for both LRW (least-recently-written) cleaning and greedy cleaning under simple nonuniform traffic conditions; simulation results show the first model to be exact and the second to be accurate within 2%. We extend the LRW model to arbitrary combinations of random traffic and demonstrate its use in predicting cleaning performance for real-world workloads. Using these analytic models, we examine the strategy of separating “hot” and “cold” data, showing that for our traffic model, such separation eliminates any loss in performance due to nonuniform traffic. We then show how a system which segregates hot and cold data into different block pools may shift free space between these pools in order to achieve improved performance, and how numeric methods may be used with our model to find the optimum operating point, which approaches a write amplification of 1.0 for increasingly skewed traffic. We examine online methods for achieving this optimal operating point and show a control strategy based on our model which achieves high performance for a number of real-world block traces.

Resource Demand Prediction and Carbon Emission Estimation for Data Centers

The energy consumption of data centers has become a key issue in today's ICT sector and a significant factor of green environment. A substantial reduction in energy consumption can be made by powering down servers when they are not in use. In cloud data center, it is very hard to manage and allocate their resource to incoming dynamic workload demands. Predicting the required resource demand, it can save the data center's resource wasting and achieve max-profit and min-risk. Without proper prediction, data center can have overprovision and underprovision, which can cause resource waste and significant financial penalties. So an efficient resource management scheme is needed to reduce energy consumption and carbon dioxide (CO2 ) emission. The aim of the present study is to develop model for predicting the future resource demand and estimation of CO2 emission by comparatively assessing the suitability of several machine learning techniques. In order to reduce processing overheads, feature selection is conducted in prediction model. To estimate the CO2 emission, Power model and Carbon model are also developed. Experiment is conducted on real world workload traces and results show that prediction model can predict future resource demand with acceptable accuracy.

Memory efficient and scalable address mapping for flash storage devices

Flash memory devices commonly rely upon traditional address mapping schemes such as page mapping, block mapping or a hybrid of the two. Page mapping is more flexible than block or hybrid mapping without being restricted by block boundaries. However, its mapping table tends to grow large quickly as the capacity of flash memory devices does. To overcome this limitation, we propose novel mapping schemes that are fundamentally different from the existing mapping strategies. We call these new schemes Virtual Extent Trie (VET) and Extent Mapping Tree (EMT), as they manage mapping information by treating each I/O request as an extent and by using extents as basic mapping units rather than pages or blocks. By storing extents instead of individual addresses, our extent mapping schemes consume much less memory to store mapping information and still remain as flexible as page mapping. We observed in our experiments that our schemes reduced memory consumption by up to an order of magnitude in com- parison with the traditional mapping schemes for several real world workloads. Our extent mapping schemes also scaled well with increasing address spaces by synthetic workloads. Even though the asymp- totic mapping cost of VET and EMT is higher than the O(1) time of a page mapping scheme, the amount of increased overhead was almost negligible or low enough to be hidden by an accompanying I/O operation.

Framework for Analyzing Android I/O Stack Behavior: From Generating the Workload to Analyzing the Trace

The existing I/O workload generators and trace capturing tools are not adaptable to generating nor capturing the I/O requests of Android apps. The Android platform needs proper tools to capture and replay real world workload in the Android platform to verify the result of benchmark tools. This paper introduces Android Storage Performance Analysis Tool, AndroStep, which is specifically designed for characterizing and analyzing the behavior of the I/O subsystem in Android based devices. The AndroStep consists of Mobibench (workload generator), MOST (Mobile Storage Analyzer), and Mobigen (workload replayer). Mobibench is an Android app that generates a filesystem as well as SQLite database operations. Mobibench can also vary the number of concurrent threads to examining the filesystem scalability to support concurrency, e.g., metadata updates, journal file creation/deletion. MOST captures the trace and extracts key filesystem access characteristics such as access pattern with respect to file types, ratio between random vs. sequential access, ratio between buffered vs. synchronous I/O, fraction of metadata accesses, etc. MOST implements reverse mapping feature (finding an inode for a given block) and retrospective reverse mapping (finding an inode for a deleted file). Mobigen is a trace capturing and replaying tool that is specifically designed to perform the user experiment without actual human intervention. Mobigen records the system calls generated from the user behavior and sanitizes the trace for replayable form. Mobigen can replay this trace on different Android platforms or with different I/O stack configurations. As an example of using AndroStep, we analyzed the performances of twelve Android smartphones and the SQLite performances on five different filesystems. AndroStep makes otherwise time consuming I/O stack analysis extremely versatile. AndroStep makes a significant contribution in terms of shedding light on internal behavior of the Android I/O stack.

Anchor: A Versatile and Efficient Framework for Resource Management in the Cloud

We present Anchor, a general resource management architecture that uses the stable matching framework to decouple policies from mechanisms when mapping virtual machines to physical servers. In Anchor, clients and operators are able to express a variety of distinct resource management policies as they deem fit, and these policies are captured as preferences in the stable matching framework. The highlight of Anchor is a new many-to-one stable matching theory that efficiently matches VMs with heterogeneous resource needs to servers, using both offline and online algorithms. Our theoretical analyses show the convergence and optimality of the algorithm. Our experiments with a prototype implementation on a 20-node server cluster, as well as large-scale simulations based on real-world workload traces, demonstrate that the architecture is able to realize a diverse set of policy objectives with good performance and practicality.

A study of pricing for cloud resources

We present a study of pricing cloud resources in this position paper. Our objective is to explore and understand the interplay between economics and systems designs proposed by recent research. We develop a general model that captures the resource needs of various applications and usage pricing of cloud computing. We show that a uniform price does not suffer any revenue loss compared to first-order price discrimination. We then consider alternative strategies that a provider can use to improve revenue, including resource throttling and performance guarantees, enabled by recent technical developments. We prove that throttling achieves the maximum revenue at the expense of tenant surplus, while providing performance guarantees with an extra fee is a fairer solution for both parties. We further extend the model to incorporate the cost aspect of the problem, and the possibility of right-sizing capacity. We reveal another interesting insight that in some cases, instead of focusing on right-sizing, the provider should work on the demand and revenue side of the equation, and pricing is a more feasible and simpler solution. Our claims are evaluated through extensive trace-driven simulations with real-world workloads.

SCALLA

Today’s one-pass analytics applications tend to be data-intensive in nature and require the ability to process high volumes of data efficiently. MapReduce is a popular programming model for processing large datasets using a cluster of machines. However, the traditional MapReduce model is not well-suited for one-pass analytics, since it is geared towards batch processing and requires the dataset to be fully loaded into the cluster before running analytical queries. This article examines, from a systems standpoint, what architectural design changes are necessary to bring the benefits of the MapReduce model to incremental one-pass analytics. Our empirical and theoretical analyses of Hadoop-based MapReduce systems show that the widely used sort-merge implementation for partitioning and parallel processing poses a fundamental barrier to incremental one-pass analytics, despite various optimizations. To address these limitations, we propose a new data analysis platform that employs hash techniques to enable fast in-memory processing, and a new frequent key based technique to extend such processing to workloads that require a large key-state space. Evaluation of our Hadoop-based prototype using real-world workloads shows that our new platform significantly improves the progress of map tasks, allows the reduce progress to keep up with the map progress, with up to 3 orders of magnitude reduction of internal data spills, and enables results to be returned continuously during the job.

Migration-Resistant Policies for Probe-Wear Leveling in MEMS Storage Devices

Probes (read/write heads) in a MEMS storage device are susceptible to wear. We study probe wear, and analyze the causes of uneven wear. We show that under real-world workloads some probes can wear one order of magnitude faster than others. This premature expiry has severe consequences for reliability, timing performance, energy efficiency, and lifetime. Wear leveling precludes premature expiry and is thus necessary. We discuss the fundamental differences between probe wear in MEMS storage and medium wear in Flash, calling for a different treatment. We devise three policies to level probe wear. The policies provide a spectrum between best lifetime and least influence on the response time and energy efficiency of a MEMS storage device. We make the case that data migration can be prevented by augmenting the policies with a simple rule. We study the influence of the data layout configuration on the leveling performance of the policies.

An adaptive model-free resource and power management approach for multi-tier cloud environments

With the development of cloud environments serving as a unified infrastructure, the resource management and energy consumption issues become more important in the operations of such systems. In this paper, we investigate adaptive model-free approaches for resource allocation and energy management under time-varying workloads and heterogeneous multi-tier applications. Specifically, we make use of measurable metrics, including throughput, rejection amount, queuing state, and so on, to design resource adjustment schemes and to make control decisions adaptively. The ultimate objective is to guarantee the summarized revenue of the resource provider while saving energy and operational costs. To validate the effectiveness, performance evaluation experiments are performed in a simulated environment, with realistic workloads considered. Results show that with the combination of long-term adaptation and short-term adaptation, the fluctuation of unpredictable workloads can be captured, and thus the total revenue can be preserved while balancing the power consumption as needed. Furthermore, the proposed approach can achieve better effect and efficiency than the model-based approaches in dealing with real-world workloads.

ZFTL: power-efficient data compression support for NAND flash-based consumer electronics devices

Flash translation layers play an important role in determining the storage performance and lifetime of NAND flash-based consumer electronics devices. In this paper, we present a flash translation layer called zFTL, which reduces the amount of data written to NAND flash memory by supporting on-line, transparent data compression based on the X-Match algorithm. To minimize compression overhead and power consumption, we also propose a novel prediction scheme that identifies incompressible data in advance without going through full compression. Our evaluations with five real-world workloads show that zFTL successfully enhances storage performance and lifetime by improving the write amplification factor (WAF) by a factor of 2.6 (geometric mean) compared to the case without compression support. In addition, we find that the proposed prediction scheme is effective in reducing power consumption by skipping compression for incompressible data.

Request Bridging and Interleaving

Write-through caching in modern disk drives enables the protection of data in the event of power failures as well as from certain disk errors when the write-back cache does not. Host system can achieve these benefits at the price of significant performance degradation, especially for small disk writes. We present new block-level techniques to address the performance problem of write-through caching disks. Our techniques are strongly motivated by some interesting results when the disk-level caching is turned off. By extending the conventional request merging, request bridging increases the request size and amortizes the inherent delays in the disk drive across more bytes of data. Like sector interleaving, request interleaving rearranges requests to prevent the disk head from missing the target sector position in close proximity, and thus reduces disk latency. We have evaluated our block-level approach using a variety of I/O workloads and shown that it increases disk I/O throughput by up to about 50%. For some real-world workloads, the disk performance is comparable or even superior to that of using the write-back disk cache. In practice, our simple yet effective solutions achieve better tradeoffs between data reliability and disk performance when applied to write-through caching disks.