ISCA'12 Tutorial: CloudSuite on Flexus
- Where? ISCA'12 (Hilton Portland & Executive Tower, Portland, Oregon.)
- When? Saturday, June 9th, 2012 (A full-day tutorial).
- Who? Academic/industrial researchers interested in scale-out datacenter applications, and their cycle-accurate performance evaluation via simulation.
- Keywords Emerging scale-out workloads, full-system simulation, sampling, measurement methodology.
- Requirement: The attendees should bring a 1 terabyte disk to get a copy of the CloudSuite released images.
The emergence of cloud computing as a dominant computing platform highlights the need for architectural evaluation of cloud server systems. In this tutorial, we will present CloudSuite, a new open-source benchmark suite for exploring the architectural behavior of scale-out cloud applications. We will highlight the key differences between cloud applications and traditional benchmarks used for architectural evaluation. Because evaluating cloud applications using current software simulators is prohibitively slow, we will also describe the Flexus simulation infrastructure and measurement methodology, which enables fast and accurate simulation of full-system benchmarks like CloudSuite. We will include a practical session to give attendees hands-on experience with evaluation of cloud applications using the CloudSuite disk images and checkpoints that will be publically released as part of the tutorial.
CloudSuite is a benchmark suite of representative scale-out applications, such as web search, media streaming, and data analytics. CloudSuite is based on real-world full-system software stacks and data sets and takes into account quality of service (QoS) performance metrics. CloudSuite benchmarks are compatible with the Flexus infrastructure, which achieves fast simulation turnaround while ensuring representative results by leveraging the SMARTS simulation sampling methodology. The Flexus component-based design allows for the composition of complex multi-core systems, enabling rapid exploration of cloud applications on a wide range of processor organizations and micro-architectures.
Our tutorial is organized as follows. First, we will introduce CloudSuite 1.0, a benchmark suite covering emerging cloud applications. We will present details on the software architecture of the benchmarks and steps to run them on real hardware. Second, we will make the case for sampling-based simulation of CloudSuite applications using Flexus. We will present relevant background and explain how the SMARTS methodology is extended to enable fast and statistically rigorous evaluation of systems running cloud workloads. Third, we will present an overview of the design of the Flexus components that enable rapid architectural exploration. Finally, we will provide a hands-on tutorial on running CloudSuite benchmarks on the latest Flexus distribution.
|8:30-10:00||CloudSuite 1.0||The structure of CloudSuite 1.0 benchmarks and details on running them on real hardware||Almutaz Adileh|
|10:30-12:00||Flexus part I||Applying the SMARTS statistical sampling methodology to measure processor performance with cloud workloads||Pejman Lotfi-Kamran|
|13:30-14:00||Flexus part II||Flexus simulator internals and tools||Stavros Volos|
|14:00-15:00||Flexus hands-on tutorial||The complete process of going from workload creation to checkpoint library and then to performance results||Stavros Volos|
|15:30-17:00||Hands-on tutorial: CloudSuite on Flexus||Demonstration of the CloudSuite images, and a how to use them correctly with Flexus for performance evaluation||Cansu Kaynak|
- Almutaz Adileh is a third-year PhD candidate in the Parallel Systems Architecture Laboratory at EPFL, where he is advised by Prof. Babak Falsafi. Almutaz's research interests target energy efficiency in datacenters, with emphasis on designing energy-efficient servers targeting cloud scale-out applications.
- Cansu Kaynak is a third-year PhD candidate in the Parallel Systems Architecture Laboratory at EPFL, advised by Prof. Babak Falsafi. Cansuís research focuses on high-performance memory systems to bridge the ever-increasing processor/memory performance gap, a key performance bottleneck for memory-intensive server applications.
- Pejman Lotfi-Kamran is a fourth-year PhD candidate in the Parallel Systems Architecture Laboratory at EPFL, where he is advised by Prof. Babak Falsafi. Pejman's research interests include various aspects of computer architecture, including multi-core architectures, power-efficient architectures, and interconnection networks.
- Stavros Volos is a third-year PhD candidate in the Parallel Systems Architecture Laboratory at EPFL, where he is advised by Prof. Babak Falsafi. Stavros' research interests target designing technology-scalable datacenter processors that match the requirements of emerging data-oriented services.