ETHZ.12

Academic Compute Cloud

Long Title: Academic Compute Cloud Provisioning and Usage
Leading
Organization:
ETH Zürich
Participating
Organizations:
Universität Zürich
Domain: Grid
Status: finished
Start Date: 10.07.2012
End Date: 28.02.2013
Project Leader: P. Kunszt
Deputy Project Leader: S. Maffioletti

Is the cloud model suitable for an application or not? This project delivers the answers whether and how academic resource providers can expose their resources using cloud technologies to support scientific communities.

Results

Component Description
Academic Compute Cloud wiki Results, recommendations, comparisions, surveys and slides (public)
Workshop April 29 2013 Slides (public); detailed report for the ETH and UZH to be used as input to their respective IT Cloud strategies is available upon request and counts 85 pages
Octavius cloud resource Cloud resource at ETHZ (restricted access; accounts are given to test users on request)
Hobbes cloud resource Cloud resource at UZH (restricted access; accounts are given to test users on request)
Project wiki Project wiki (protected, open by request to project partners and universities)

This project was to build knowledge and to share this knowledge with all interested parties in Switzerland.
The team wanted to define the cloud, as there are already too many definition of the word "cloud computing". In this definition, it is a cloud if a user can request certain resources or services on demand at any time without any prior contracts or long-term commitments. The user can make use of these services for a certain amount of time and then he or she only has to pay for the actual usage of the resource. The usage is monitored and the user can see at any time how much is being used and what the cost is. Should the user need more or less resources, they can be dynamically scaled up or down. What's more, the user can drive all these services programmatically with his or her own software. If all this is possible, is is a cloud according to the project team definitions.
A ready-made solution from Hewlett Packard was rented and two open-source cloud systems built, based on a software called OpenStack, one at the ETH and one at the University of Zurich. All three analyzed systems (the commercial HP, and the two OpenStack systems) need professionals to operate them. The commercial system optimally needs expert support directly from the company. The open source solutions are not yet as mature but can be made to work if local expertise exists.
Some charging models for such pay-per-use facilities are proposed, as currently it is not easy for researchers to receive money for such infrastructures.

Both the ETH and the UZH take the output of this project as input to their respective cloud IT strategies. For the immediate future, both pilot infrastructures will be kept operational and will be part of the CRUS Bridge project for the Swiss Academic Compute Cloud (SwissACC), led by the UZH.


Goals

The project team will investigate

  • how university cluster services can be extended and improved by making use of cloud computing technology,
  • the performance implications, and
  • how to accommodate new application usage models like MapReduce.
The scalability and reliability of this approach will be tested with real workloads. A distributed testbed infrastructure will be deployed at the ETH Zurich and the UZH. This testbed will be designed to be flexible and capable enough to explore a variety of computing models in order to understand the impact on various scientific applications. The testbeds will also be used as a reference resource for selected application scientists.
The cloud computing model has performance implications due to virtualization and is not always suitable for parallel computations or for data intensive workloads. The intention is to learn from this project how to decide whether the cloud model is suitable for an application or not, especially since the rapidly evolving capabilities of clouds.

Benefit

Cloud approaches provide many advantages that are attractive to the scientific community. Through virtualization, applications have maximum flexibility in their deployment and new computing models can be accommodated without the need to reconfigure the cluster (MapReduce, BigTable, etc). On the infrastructure level it can allow a cluster to be extended dynamically to partners such as another private university cloud (making resource sharing possible on the infrastructure level) or to a public cloud (Amazon EC2, IBM SmartCloud, etc). This dynamic extension of a local private cloud is generally called Cloud Bursting. It will be tested between UZH and ETH Zurich, and also to the public clouds.

Description

1. Cloud Bursting batch cluster jobs to private and public clouds
The computing cluster be extended dynamically to another available private cloud only when needed and under certain conditions. This can be done also without virtualization. Cloud Bursting to the public cloud like Amazon EC2 will be explored, but also the SMSCG infrastructure, which can run virtual cluster image instances, extending the institutional private cloud dynamically. For that purpose, results of the VM-MAD project will be used.
Cloud Bursting the cluster has the advantage that end users do not need to change how they work with the system, their data analysis methods as the access patterns remain unchanged. Nevertheless, not all applications are suitable for such a model and the performance impact for various applications must be investigated. Applications that will be tested are:

  • Baseline Application: Rosetta has been used already successfully on clouds and grids, and will be used as a baseline reference for the other applications.
  • Compute Intensive Application: One or several of the currently used compute intensive applications on the Brutus cluster will be used to measure its performance in the Cloud Bursting scenario, both to partner and public clouds. It should be evaluated in the activity which applications are best suited for such an infrastructure. This could be computational chemistry applications (GAMESS or CP2K) or from computational fluid dynamics like Ansys.
  • Data Intensive Application: The systems biology image analysis pipeline of the SystemsX.ch InfectX project is highly customized and very data intensive. In the testbed the technology that will be developed by the RSNAS project will be used.

2. Virtualized Environments
In this model an EC2-like service on the institutional private cloud will be offered; end-users can dynamically instantiate virtual infrastructures for their scientific applications. They control and configure their own virtual instances. This has the advantage that virtual instances are fully customizable by end-users to their application-specific requirements that might otherwise not be met on a standardized cluster infrastructure. With this approach unsupported applications, like the applications used in VM-MAD by the Functional Genomic Center or full virtual clusters can be set up to operate on the institutional private cloud (for example a dedicated MapReduce/Hadoop cluster environments).

3. Funding Model and Usage Policies.
Usage policies and financial implications of shared private and public cloud usage will be investigated. Public clouds can be costly if used excessively and users need to be accounted and billed for cloud usage. Also private cloud sharing needs to be accounted for to assure fair usage on both sides. In terms of security the available layers and their suitability for research workloads will be tested. Information about legal aspects as far as such information is available will be gathered.

Back