Wednesday, 13 December 2017

Session 4: Benchmarking 1

Session Chair: David Kaeli (Northeastern University)

Experimental Study of Protocol-independent Redundancy Elimination Algorithms

Authors:

Maxim Martynov (Cisco Systems)

Abstract:

The idea of identifying and removing repetitive patterns in the network data transfers, also known as protocol-independent redundancy elimination, and its benefits have received thorough consideration. However, actual implementation of such systems received much less attention. The intention of the redundancy elimination is to increase capacity of low-bandwidth network connections, when searching for redundancies and replacing them is faster than transmitting unprocessed redundant data. As long as network is slow, any reasonable implementation is beneficial. But as network capacities grow, the maximal throughput that system can provide becomes critical for its deployment. Thus, an appropriate choice of redundancy eliminating algorithm and its parameters becomes very important. This work addresses the problem of algorithm and parameter selection. We describe possible variations of the basic scheme, and demonstrate experiments that we have conducted for each variation. We discuss the trends observed in the results and explain their nature. We then propose a methodology to make a choice of the algorithm and its parameters, based on the obtained measurements.

DOI: 10.1145/1712605.1712628

Full text: PDF

[#][]

A Power Consumption Analysis of Decision Support Systems

Authors:

Meikel Poess (Oracle Corporation)
Raghunath Othayoth Nambiar (Hewlett-Packard Company)

Abstract:

Enterprise data warehouses have been doubling every three years, demanding high compute power and storage capacities. The in-dustry is expected to meet such compute demands, but dealing with the dramatic increase in energy requirements will be challenging. Energy efficiency has already become the top priority for system developers and data center managers. While system vendors focus on developing energy efficient systems there is a huge demand for industry-standard workloads and processes to measure and analyze energy consumption for enterprise data warehouses. SPEC has developed a power benchmark for single servers (SPECpower_ssj2008), but so far, no benchmark exists that measures the power consumption of large, complex systems. In this paper, we present a simple power consumption model for enterprise data warehouses based on the industry standard TPC-H benchmark. By applying our model to a subset of 7 years of TPC-H publications, we identify the most power-intensive components where research and development should focus and also analyze existing power consumption trends over time. This paper com-plements a similar study conducted for enterprise OLTP systems published by the same authors at VLDB 2008 and the Transaction Processing Performance Council's initiative of energy metric to its benchmarks.

DOI: 10.1145/1712605.1712629

Full text: PDF

[#][]

Black-box Performance Models for Virtualized Web Service Applications

Authors:

Danilo Ardagna (Politecnico di Milano)
Mara Tanelli (Politecnico di Milano)
Marco Lovera (Politecnico di Milano)
Li Zhang (IBM T.J. Watson Research Center)

Abstract:

In order to reduce the operating costs of IT systems, nowadays service applications are executed in virtualized infrastructures and a time varying fraction of the physical servers' capacity is shared among running applications. The performance modelling of a virtualized server is very challenging as the impact of the choice of the Virtual Machine Monitor (VMM) scheduler, its parameters and I/O management overhead is still only partially understood. In this paper, black-box models based on the Linear Parameter Varying (LPV) framework are proposed for the run-time modelling and performance control of Web services in virtualized hosting environments. As the behavior of the application response time is highly time varying and the workload conditions substantially change within the same business day, LPV models seem very promising for predicting the performance of such systems. Specifically, the suitability of subspace LPV identification methods for multi-variable systems is investigated and their performance assessed on experimental data gathered on Xen environments.

DOI: 10.1145/1712605.1712630

Full text: PDF

[#][]