In the original version of this article, the abbreviation ‘OWDM’ was incorrectly defined. The phrase ‘orthogonal wavelength division multiplexing’ should all be changed to ‘one-way wave depth migration’.

With supercomputers developing towards exascale, the number of compute cores increases dramatically, making more complex and larger-scale applications possible. The input/output (I/O) requirements of large-scale applications, workflow applications, and their checkpointing include substantial bandwidth and an extremely low latency, posing a serious challenge to high performance computing (HPC) storage systems. Current hard disk drive (HDD) based underlying storage systems are becoming more and more incompetent to meet the requirements of next-generation exascale supercomputers. To rise to the challenge, we propose a hierarchical hybrid storage system, on-line and near-line file system (ONFS). It leverages dynamic random access memory (DRAM) and solid state drive (SSD) in compute nodes, and HDD in storage servers to build a three-level storage system in a unified namespace. It supports portable operating system interface (POSIX) semantics, and provides high bandwidth, low latency, and huge storage capacity. In this paper, we present the technical details on distributed metadata management, the strategy of memory borrow and return, data consistency, parallel access control, and mechanisms guiding downward and upward migration in ONFS. We implement an ONFS prototype on the TH-1A supercomputer, and conduct experiments to test its I/O performance and scalability. The results show that the bandwidths of single-thread and multi-thread ‘read’/‘write’ are 6-fold and 5-fold better than HDD-based Lustre, respectively. The I/O bandwidth of data-intensive applications in ONFS can be 6.35 times that in Lustre.

Hard Disk Drive (HDD) based storage systems in high performance computing (HPC), such as Lustre, are being confronted with great challenges from massive parallel data-intensive applications, and becomes one of the most critical factors that influence applications performance. In this paper, we analyze the performance of Lustre, and statistically examine the distribution of file sizes, I/O patterns for typical data-intensive applications, as well as the utilization of CPU and memory resources in TH-1A supercomputer from the last five years. We present masFS, a novel file system for HPC that exploits available memory and SSD resources on compute nodes with little interference to applications running on the nodes. The masFS supports POSIX interface with an unique namespace and provides comprehensive file service for applications. We design and implement masFS with memory and SSD, and have deployed and evaluated it on TH-1A. Experimental results show that masFS works as general storage system and can meet the requirement of variety of I/O loads. It can achieve a speedup of 14.5x for read and 8.8x for write respectively in running benchmarks, and an average 7.6x faster I/O time in a real-world data-intensive application compared to Lustre.

Next-generation multiprocessor real-time systems consume less energy at the cost of increased power density. This increase in power density results in high heat density and may affect the reliability and performance of real-time systems. Thus, incorporating maximum temperature constraints in scheduling of real-time task sets is an important challenge. This paper investigates a novel algorithm for thermal-constrained energy-aware partitioning of periodic real-time tasks in heterogeneous multiprocessor systems. When designing our new algorithm, we have applied insights gained from a famous knapsack problem solution. Both simulation and experimental results show that our new branch-and-bound based partitioning algorithm can significantly reduce the total energy consumption of multiprocessor real-time systems.

Motivated by the increasing trend in embedded systems towards platform integration, there has been an increasing research interest in scheduling mixed-criticality systems. However, most existing efforts have concentrated on scheduling sequential tasks and ignored intra-task parallelism. In this paper, we study the scheduling of mixed-criticality parallel jobs on multiprocessor platforms. We propose a synchronous mixed-criticality job model, where each job consists of segments, each segment having an arbitrary number of parallel threads that synchronize at the end of the segment. A novel MinLoad algorithm is developed to decompose mixed-criticality parallel jobs into mixed-criticality sequential jobs. This decomposition enables us to leverage existing mixed-criticality scheduling algorithms and schedulability analysis to the multiprocessor scheduling of mixed-criticality parallel jobs. In addition, our MinLoad job decomposition algorithm is designed to make the decomposed mixed-criticality sequential tasks easier to schedule, and thus requires smaller-sized multiprocessor platforms for the mixed-criticality systems.

The Internet has recently been evolving from homogeneous congestion control to heterogeneous congestion control. Several years ago, Internet traffic was mainly controlled by the traditional RENO, whereas it is now controlled by multiple different TCP algorithms, such as RENO, CUBIC, and Compound TCP (CTCP). However, there is very little work on the performance and stability study of the Internet with heterogeneous congestion control. One fundamental reason is the lack of the deployment information of different TCP algorithms. In this paper, we first propose a tool called TCP Congestion Avoidance Algorithm Identification (CAAI) for actively identifying the TCP algorithm of a remote Web server. CAAI can identify all default TCP algorithms (e.g., RENO, CUBIC, and CTCP) and most non-default TCP algorithms of major operating system families. We then present the CAAI measurement result of about 30 000 Web servers. We found that only 3.31 % ~ 14.47 % of the Web servers still use RENO, 46.92% of the Web servers use BIC or CUBIC, and 14.5 % ~ 25.66 % of the Web servers use CTCP. Our measurement results show a strong sign that the majority of TCP flows are not controlled by RENO anymore, and a strong sign that the Internet congestion control has changed from homogeneous to heterogeneous.

Energy efficiency is now used as an important metric for evaluating a computing system. However, saving energy is a big challenge due to many constraints. For example, in one of the most popular distributed processing frameworks, Hadoop, three replicas of each data block are randomly distributed in order to improve performance and fault tolerance. But such a mechanism limits the largest number of machines that can be turned off to save energy without affecting the data availability. To overcome this limitation, previous research introduces a new mechanism called covering subset which maintains a set of active nodes to ensure the immediate availability of data, even when all other nodes are turned off. This covering subset based mechanism works smoothly if no failure happens. However, a node in the covering subset may fail. In this paper, we study the energy-efficient failure recovery in Hadoop clusters. Rather than only using the replication as adopted by a Hadoop system by default, we investigate both replication and erasure coding as possible redundancy mechanisms. We develop failure recovery algorithms for both systems and analytically compare their energy efficiency.

Next-generation multi-core multiprocessor real-time systems consume less energy at the cost of increased power density. This increase in power-density results in high heat density and may affect the reliability and performance of real-time systems. Thus, incorporating maximum temperature constraints in scheduling of real-time task sets is an important challenge. This paper investigates thermal-constrained energy-aware partitioning of periodic real-time tasks in heterogeneous multi-core multiprocessor systems. We adopt a power model which considers the impact of temperature and voltage on a processor's static power consumption. Two types of thermal models are used to respectively capture negligible and non-negligible amount of heat transfer among cores. We develop a novel genetic-algorithm based approach to solve the heterogeneous multi-core multiprocessor partitioning problem. Extensive simulations were performed to validate the effectiveness of the approach. Experimental results show that integrating a worst-fit based partitioning heuristic with the genetic algorithm can significantly reduce the total energy consumption of a heterogeneous multi-core multiprocessor real-time system.

The Internet has recently been evolving from homogeneous congestion control to heterogeneous congestion control. Several years ago, Internet traffic was mainly controlled by the traditional AIMD algorithm, whereas Internet traffic is now controlled by many different TCP algorithms, such as AIMD, BIC, CUBIC, and CTCP. However, there is very little work on the performance and stability study of the Internet with heterogeneous congestion control. One fundamental reason is the lack of the deployment information of different TCP algorithms. In this paper, we first propose a tool called TCP Congestion Avoidance Algorithm Identification (CAAI) for actively identifying the TCP algorithm of a remote web server. CAAI can identify all default TCP algorithms (i.e., AIMD, BIC, CUBIC, and CTCP) and most non-default TCP algorithms of major operating system families. We then present, for the first time, the CAAI measurement result of the 5000 most popular web servers. Among the web servers with valid traces, we found that only 16.85~25.58\% of web servers still use the traditional AIMD, 44.51\% of web servers use BIC or CUBIC, and 10.27$\sim$19\% of web servers use CTCP. In addition, we found that, for the first time, some web servers use non-default TCP algorithms, some web servers use some unknown TCP algorithms which are not available in any major operating system family, and some web servers use abnormal slow start algorithms. Our CAAI measurement results show a strong sign that the majority of TCP flows are not controlled by AIMD anymore, and a strong sign that the Internet congestion control has already changed from homogeneous to highly heterogeneous.

With growing cost of electricity, the power management of server clusters has become an important problem. Most existing research, however, either do not apply to virtualized environments or do not focus on power-efficient workload distribution. To fill in this research gap, we propose a workload distribution algorithm for virtualized server clusters to reduce their power consumptions and provide quality of service (QoS). Built upon optimization, queuing theory and control theory techniques, our approach achieves the design goal, where QoS is provided to a larger number of requests with a smaller amount of power consumption.

Providing QoS and performance guarantees to arbitrarily divisible loads has become a significant problem for many cluster-based research computing facilities. While progress is being made in scheduling arbitrarily divisible loads, current approaches are not efficient and do not scale well. In this paper, we propose a linear algorithm for real-time divisible load scheduling. Unlike existing approaches, the new algorithm relaxes the tight coupling between the task admission controller and the task dispatcher. By eliminating the need to generate exact schedules in the admission controller, the algorithm avoids high overhead. We experimentally evaluate the new algorithm. Simulation results demonstrate that the algorithm scales well, can schedule large numbers of tasks efficiently, and performs similarly to existing approaches in terms of providing real-time guarantees.

With growing cost of electricity, the power management (PM) of server clusters has become an important problem. However, most previous researchers only address the challenge in homogeneous environments. Considering the increasing popularity of heterogeneous systems, this paper proposes an efficient algorithm for PM of heterogeneous soft real-time clusters. It is built on simple but effective mathematical models. When deployed to a new platform, the software incurs low configuration cost because no extensive performance measurements and profiling are required. To strive for efficiency, a threshold-based approach is adopted. In this paper, we systematically study this approach and its design decisions.

Data grids provide geographically distributed storage and computing resources for large data-intensive applications. To support fast data access and processing, effective data management is essential. In this paper, we propose two novel algorithms on data placement and data replication. We evaluate the algorithms through intensive simulations. Results show that the new algorithms achieve significant improvement in facilitating load balancing, reducing job response time, and saving network bandwidth.

Providing QoS and performance guarantees to arbitrarily divisible loads has become a significant problem for many cluster-based research computing facilities. While progress is being made in scheduling arbitrarily divisible loads, previous approaches have no support for advance reservations. However, with the emergence of grid applications that require simultaneous access to multi-site resources, supporting advance reservations in a cluster has become increasingly important. In this paper we propose a new divisible load real-time scheduling algorithm that supports advance reservations in a cluster. Our approach not only enforces the real-time agreement but also addresses the under-utilization concerns raised by advance reservations. The impact of advance reservations on system performance is systematically studied. Simulation results show that, with the proposed algorithm and appropriate advance reservations, the system performance could be maintained at the same level as the no reservation case.

With growing cost of electricity, the power management of server clusters has become an important problem. However, most previous researchers only address the challenge in homogeneous environments. Considering the increasing popularity of heterogeneous systems, this paper proposes an efficient algorithm for power management of heterogeneous soft real-time clusters. It is built on simple but effective mathematical models. When deployed to a new platform, the software incurs low configuration cost because no extensive performance measurements and profiling are required. To strive for efficiency, a threshold-based approach is adopted. In this paper, we systematically study this approach and its design decisions.

Providing QoS and performance guarantees to arbitrarily divisible loads has become a significant problem for many cluster-based research computing facilities. While progress is being made in scheduling arbitrarily divisible loads, some of proposed approaches may cause inserted idle times (IITs) that are detrimental to system performance. In this paper we propose a new approach that utilizes IITs and thus enhances the system performance. The novelty of our approach is that, to simplify the analysis, a homogenous system with IITs is transformed to an equivalent heterogeneous system, and that our algorithms can schedule real-time divisible loads with different processor available times. Intensive simulations show that the new approach outperforms the previous approach in all configurations. We also compare the performance of our algorithm to the current practice of manually splitting workloads by users. Simulation results validate the advantages of our approach.

Cluster computing has emerged as a new paradigm for solving large-scale problems. To enhance QoS and provide performance guarantees in cluster computing environments, various real-time scheduling algorithms and workload models have been investigated. Computational loads that can be arbitrarily divided into independent pieces represent many real-world applications. Divisible load theory (DLT) provides insight into distribution strategies for such computations. However, the problem of providing performance guarantees to divisible load applications has not yet been systematically studied. This paper investigates such algorithms for a cluster environment. Design parameters that affect the performance of these algorithms and scenarios when the choice of these parameters have significant effects are studied. A novel algorithmic approach integrating DLT and EDF (earliest deadline first) scheduling is proposed. For comparison, we also propose a heuristic algorithm. Intensive experimental results show that the application of DLT to real-time cluster-based scheduling leads to significantly better scheduling approaches

This paper presents the design and implementation of an adaptive Web server architecture to provide relative and absolute connection delay guarantees for different service classes. The first contribution of this paper is an adaptive architecture based on feedback control loops that enforce desired connection delays via dynamic connection scheduling and process reallocation. The second contribution is the use of control theoretic techniques to model and design the feedback loops with desired dynamic performance. In contrast to heuristics-based approaches that rely on laborious hand-tuning and testing iteration, the control theoretic approach enables systematic design of an adaptive Web server with established analytical methods. The adaptive architecture has been implemented by modifying an Apache server. Experimental results demonstrate that the adaptive server provides robust delay guarantees even when workload varies significantly.

With the dramatic explosion of online information, the Internet is undergoing a transition from a data communication infrastructure to a global information utility. PDAs, wireless phones, Web-enabled vehicles, modem PCs, and high-end workstations can be viewed as appliances that "plug-in" to this utility for information. The increasing diversity of such appliances calls for an architecture for performance differentiation of information access. The key performance accelerator on the Internet is the caching and content distribution infrastructure. While many research efforts addressed performance differentiation in the network and on Web servers, providing multiple levels of service in the caching system has received much less attention. It has two main contributions. First, we describe, implement, and evaluate an architecture for differentiated content caching services as a key element of the Internet content distribution architecture. Second, we describe a control-theoretical approach that lays well-understood theoretical foundations for resource management to achieve performance differentiation in proxy caches. An experimental study using the Squid proxy cache shows that differentiated caching services provide significantly better performance to the premium content classes.

The increasing diversity of Internet appliances calls for an architecture for performance differentiation on information access. The World Wide Web is the dominant interface for information access today. Web proxy caching is the key performance accelerator in the Web infrastructure. While many research efforts have addressed performance differentiation in the network and on Web servers, providing multiple levels of service in proxy caches has received much less attention. This paper has two main contributions. First, we describe, implement and evaluate an architecture for differentiated content caching services as a key element of the Internet infrastructure. Second, we describe a control-theoretical approach that lays well-understood theoretical foundations for resource management to achieve performance differentiation in proxy caches. We describe our experiences with implementing the differentiated caching services scheme in Squid, a popular proxy cache used by many Internet service providers today. Experimental studies and analyses prove that differentiated caching services provide significant better performance to the premium content classes.