System monitoring is an important basis for system modelling and improvement. For achieving higher efficiency in performance and lower energy consumption in cluster systems, we design a monitoring system that tracks the performance and temperature of clusters for education and research purposes. The total energy consumption of clusters can be estimated by using the performance and temperature data. Specifically, in our proposed system, all real-time data (including temperature and activities of components of computing nodes) is collected and stored in Round-Robin Databases (RRDs). These data can be visualized or downloaded through a friendly user interface for further analysis. Moreover, our system also provides users with a powerful runtime comparison feature, which allows users to compare the performance of a running experiment with historical experimental results without waiting for the completion of experiments. The data visualization and user interfaces in the monitoring system are demonstrated by using an experiment on our cluster system.

Publication:

[1] X.-F. Jiang, T. Baigalmaa, L. Nguyen, D. Akiyoshi, E. Ramthun, N. Parajuli, and C. Peck, "High Performance Cluster Monitoring System", 2018 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), Honolulu, HI, USA, 12-15 Nov, 2018.

Disks have not been paid fully attention in their impacts on outlet temperatures of data nodes. In modern data centers, Teradata equipments could support more than 100 disks in a single data node. We expect large affect from the disks' temperature.

We proposed a thermal modeling approach to estimate temperatures of CPUs and disks (including hard drives and solid state disks). This approach can be applied to investigate thermal behaviours of storage systems under various utilizations patterns.

With the thermal modeling approach, outlet temperatures of data nodes can be estimated under particular workloads without deploying temperature sensors in storage systems. In addition, differences between inlet and outlet temperatures of data nodes can be modeled based on CPU and disk temperatures. Combining these models enables administrators to set up an appropriate inlet temperature of storage systems, which protects computing facilities from working in high temperature. Moreover, the cooling costs of storage systems can be calculated by using the proposed thermal model and the COP (Coefficient of Performance) model.

Publications:

[1] X.-F. Jiang,Y.-H. Liu, X.-J. Ruan, T. Baigalmaa, L. Nguyen, D. Akiyoshi, and C. Peck, "Energy Modeling of Cluster System", 2018 IEEE International Workshop on Signal Processing Systems (SiPS),Cape Town, South Africa, 21-24 Oct, 2018.

[2] X.-F. Jiang, M. I. Alghamdi, J. Zhang, M. Al Assaf, X.-J. Ruan, T. Muzaffar, and X. Qin, "Thermal Modeling and Analysis of Storage Systems", Proc. the 31th IEEE International Performance Computing and Communications Conference (IPCCC), 1-3 Dec. 2012. (Acceptance Rate: 27.8%, 32/115)

[3] X.-F. Jiang, M. M. Al Assaf, J. Zhang, M. I. Alghamdi, X.-J. Ruan, T. Muzaffar, and X. Qin, "Thermal modeling of hybrid storage clusters", Journal of Signal Processing Systems, 2013.

Dispatching tasks plays a significant important role in balancing the workload and reducing the energy consumption of data storage systems. Conventional task scheduling strategies distribute tasks for decreasing the computing cost of data nodes in storage systems. New trends are brought up by considering the reduction of cooling cost of data nodes.

With energy consumption of data nodes estimated by the thermal models, we proposed a task scheduling strategy, which keeps the outlet temperature of data nodes balanced across data storage systems. By considering the temperature distribution in data storage systems, a task scheduling strategy was proposed to dispatch tasks. The strategy not only selects the best data node that the task should be assigned to in terms of total energy costs of storage systems, but also ensures that the outlet temperatures of data nodes are not over a pre-determined threshold, which protects computing resources from working in high temperature environment.

Publication:

[1] X.-F. Jiang, J. Zhang, X. Qin, M.-H. Jiang, and J.-F. Zhang, "Thermal Modeling and Management of Storage Systems in Data Centers", Handbook on Data Centers (Springer), P915-944, 2015.

We built an energy efficient data transmission framework for storage systems. Giving a large amount of transferred data, PEAM judiciously selects the most energy-efficient data transmission method by predicting energy consumptions of all the candidate methods. Giving large amount of data and a group of data transmission methods, PEAM is able to select the most energy-efficient method by evaluating energy consumptions of all potential methods.

The PEAM is composed of three components: an energy cost predictor, a method selector, and a monitor. The energy cost predictor estimates energy consumption of data transmission through a particular method. The method selector chooses the best data transmission method in terms of energy efficiency. The monitor collects run-time information of each data node.

Publication:

[1] X.-F. Jiang, J. Zhang, M. I. Alghamdi, X. Qin, M.-H. Jiang, J.-F. Zhang, "PEAM: Predictive Energy-Aware Management for Storage Systems", Proc. 8th IEEE International Conference on Networking, Architecture, and Storage (NAS 2013), Xi'an, China, 17-19 July 2013.