For the HPC community, the week of Supercomputing (SC) is the most active and exciting time of the year. A colleague of mind postulated earlier today that half of the year’s news cycle is squeezed into SC, while the other half is distributed among the other 51 weeks of the year. (Anyone want to graph that distribution problem?)
Speaking of distribution, distributed computing and the concepts ensconced in grid and cloud computing have been a big part of the SC show for…well since it started 24 years ago. But in the past few years, it’s “cloud” that has gone from buzz word to ubiquity. While cloud computing has gained credibility and acceptance in business circles, the science and research community has been a bit more circumspect. The cloud (however you define it) is still not for the most-demanding HPC workloads, the ones at the top of the performance pyramid, and likely never will be, but that doesn’t matter because there are more and more scenarios that do make sense. If you want to know where that needle is and even get to watch it move, then SC12 is the place to be.
Taking place this week in Provo, Utah, SC12 has attracted the largest number of technical submissions in the event’s history, and the Technical Forum includes an impressive lineup of invited talks, research papers, tutorials, panels, workshops, research posters, and Birds of a Feather (BoF) sessions. More than 20 of these sessions are devoted to some branch of the cloud tree and I’m sure we’ll all be hearing these key phrases (and big data as well) all over the show floor and into the hallway meetings. Without further ado, here’s a rundown on the cloudy side of the Technical Program.
Data Intensive Computing in the Clouds
“As scientific applications become more data intensive, the management of data resources and dataflow between storage and compute resources is becoming the main bottleneck. Analyzing, visualizing, and disseminating these large data sets has become a major challenge,” writes the organizers for the The Third International Workshop on Data-Intensive Computing in the Clouds – DataCloud (DataCloud 2012) in their call for papers. The group came together for an all-day meeting on Sunday to discuss the challenges associated with running data-intensive workloads in the cloud. The event was organized by Tevfir Kosar, associate professor and director of the Data Intensive Distributed Computing Laboratory at the University of Buffalo, Ioan Raicu, assistant professor in the Department of Computer Science (CS) at Illinois Institute of Technology, and Roger Barga, architect and group lead in the eXtreme Computing Group (XCG), part of Microsoft Research. When it comes to handling data-intensive workloads, they are are concerned that current supercomputers, grids and clouds fall short. Their agenda is focused on identifying current issues and setting goals for cloud-based technologies that will be better matched for the data deluge.
Securing Your Middleware
Taking place Monday morning is a tutorial on Secure Coding Practices for Grid and Cloud Middleware and Services, presented by Barton Miller, Elisa Heymann. In this new cloud era, security is more important than ever. In survey after survey, security is cited as a top concern among would-be cloud users. This tutorial brings awareness to some of the best practices to help minimize security flaws in software at the coding level. The presenters share what they have learned from their experience assessing real-life cloud and middleware systems, including Google Chrome, Wireshark, Condor, SDSC Storage Resource Broker, NCSA MyProxy, INFN VOMS Admin and Core, among others. Examples are in Java, C, C++, Perl and Python.
Helping Science Tap Cloud’s Potential
Another Monday cloud tutorial examines Infrastructure Clouds and Elastic Services for Science. The outsourcing of IT shows no signs of stopping. It seems like every week another infrastructure-as-a-service provider is launching a cloud for the business world, but what about the science and research community? Increasingly, academia is looking to cloud as a way around either insufficient resources or congested queue times, but information is lacking on whether the resource used is the most suitable one. The presenters of this session – John Bresnahan and Kate Keahey of Argonne National Laborator, and Patrick Armstrong and Pierre Riteau of the University of Chicago – emphasize that the research world will benefit from a better understanding of the cloud paradigm. Their goal is to design a cloud-based infrastructure with elastic tools that specifically targets the requirements of the scientific community. Key issues to the panel are performance, elasticity and privacy. This is an all-day event.
Also taking place on Monday is the The 5th Workshop on Many-Task Computing on Grids and Supercomputers (MTAGS) 2012 with Ioan Raicu of the Illinois Institute of Technology, Ian Foster from Argonne National Laboratory, and Yong Zhao of the University of Electronic Science and Technology of China. The all-day session will highlight the development, and deployment efforts of large-scale many-task computing (MTC) applications on a wide range system types, including large scale clusters, grids, supercomputers, and clouds. Loosely coupled applications will be a key theme, as will the common obstacles to running applications on large-scale systems. These can involve local resource manager scalability and granularity, efficient utilization of raw hardware, parallel file-system contention and scalability, data management, I/O management, reliability at scale, and application scalability. The challenges of MTAGS in relation to petascale and exascale systems is another key topic.
The Remote Technical Cloud
On Tuesday, The Technical Cloud: When Remote 3D Visualization meets HPC promises to be an interesting exhibitor forum for workstation-based users (or their administrators), who face multiple challenges in regard to their post-processing work. Data access speed, rapid obsolescence, heat, noise, and application availability are commonplace in this setting. While virtualization and remote desktop solutions are helping many users address these pain points in other spaces, these technologies are still fairly rare in the technical computing space because VDI did not support the sharing of GPUs. Andrea Rodolico of NICE will show how NICE Desktop Cloud Visualization (DCV) can provide access to fully accelerated, OpenGL applications in a remote capacity on Windows or Linux. When pixels are transferred instead of data, the result is an increase in application performance, security and manageability, according to NICE. Multiple usage scenarios to be demonstrated include both physical and virtual deployments with dedicated GPU, shared GPU as well as acceleration using an external GPU appliance.
Vendor Perspective: AMD & Adaptive
In another exhibitor forum, John Gustafson gives an AMD HPC Technology Update, delivering a 30-minute presentation on HPC in the Cloud and advanced technologies. Following that session, Adaptive Computing’s Brady Kimball explores the HPC Cloud ROI and Opportunities Cloud Brings to HPC in another exhibitor forum. The message is that HPC system managers can leverage the benefits of cloud for their traditional HPC environments, especially for surge and peak loads.
In this one-hour Birds-of-a-Feather session, Robert Grossman from the University of Chicago and Heidi Alvarez from the Florida International University will speak about Collaborative Opportunities with the Open Science Data Cloud. The Open Science Data Cloud (OSDC) was formed to help scientists field increasingly-complex data management challenges.
In another BOF, researchers from Argonne and INRIA will address the importance of Interoperability in Scientific Cloud Federations. They believe that federation can be achieved through a combination of open standards and protocols or through the use of a specific middleware to connect services to cloud providers. These topics will be addressed in relation to both IaaS and PaaS models.
Yet another BOF examines the important topic: HPC Cloud: Can Infrastructure Clouds Provide a Viable Platform for HPC? Led by Argonne National Laboratory’s Kate Keahey, the presenters are bringing together key HPC stakeholders to explore what cloud viability concerns still remain despite the increased popularity of cloud services. The group’s primary objective is to locate the cutting-edge of HPC cloud and to identify the next steps.
Pay as You Go, One Watt at a Time
Pay as You Go in the Cloud: One Watt at a Time is the title of an ACM Student Research Competition Poster Session, from Yale University’s Kayo Teramoto. The session, chaired by Torsten Hoefler of ETH Zurich looks at the practice of selling virtual machines to consumers on an hourly basis. Teramoto believes the current pricing scheme employed by cloud computing fails to reflect variances from user to user, and as an alternative proposes “a new pricing model based on the liable power consumption of the VM, which correlates to the VM’s proportion of CPU and disk I/O usage.” The poster sets out to prove the fairness and practicality of their model using four consumer examples.
Cloud Storage Brokering
In a paper session chaired by Manish Parashar of Rutgers University, the authors, who hail from EPFL, introduce Scalia: An Adaptive Scheme for Efficient Multi-Cloud Storage. Scalia is described as “a cloud storage brokerage solution that continuously adapts the placement of data based on its access pattern and subject to optimization objectives, such as storage costs.” The service is said to achieve significant cost reductions by re-positioning only selected objects.
On Wednesday, Manish Parashar of Rutgers University will be the primary session leader for a BOF on Science-as-a-Service: Exploring Clouds for Computational and Data-Enabled Science and Engineering. The presenters, which also include Geoffrey Fox of Indiana University, Kate Keahey of Argonne National Laboratory, and David Lifka of Cornell University, will address the use of clusters and grids for scientific exploration and discovery. These researchers believe that understanding a hybrid infrastructure and how application workflows can effectively utilize it is key to progress. This session will explore how clouds can best be implemented to support real-world science and engineering applications. Research challenges – from both a computer science and an applications perspective – will be included, as will the community research agenda.
Meeting the Needs of Science Apps
In this Doctoral Showcase, Towards Scalable and Efficient Scientific Cloud Computing, Wojtek James Goscinski of Monash University and Iman Sadooghi of Illinois Institute of Technology investigate the use of public clouds as a source of computing infrastructure for researchers who don’t have access to the biggest and best systems. They explore the ability of clouds to support the unique needs of scientific applications, including software requirements like batch scheduling, parallel and distributed POSIX-compliant file systems, and hardware requirements like fast interconnects, and so on. They’re also looking at low-overhead virtualization techniques.
Learn About HPC Cloud and Big Data Analytics from IBM
On Thursday, the half-hour IBM exhibitor forum, HPC Cloud and Big Data Analytics – Transforming High Performance Technical Computing, will be led by IBM Platform’s Chris Porter and Scott Campbell. The presenters make the case that big data and cloud computing have transitioned from trending topics to key-enabling technologies. Cloud and workload management platforms can turn silo’d computing resources into a dynamic shared resource pool for both compute- and data-intensive applications, providing benefits such as flexibility, ease of access, reduced management overhead, and optimal resource utilization.
Learn About Windows Azure
In another half-hour exhibitor forum, Alex Sutton from Microsoft explores the business case for Azure and Windows HPC in the Cloud. Like other public cloud platforms, Azure offers on-demand, pay-as-you-go access to scalable compute and storage resources. The company says that what sets them apart is the ability to use Microsoft HPC Pack 2012 on Windows Azure, which provides “a comprehensive and cost-effective solution that delivers high performance while providing a unified solution for running compute and data intensive HPC applications on Windows Azure and on premises.”
Self-Management for Volunteer Grids
In a Thursday grid computing paper session, Trilce Estrada and Michela Taufer of the University of Delaware highlight the drawbacks of manual intervention. On the Effectiveness of Application-Aware Self-Management for Scientific Discovery in Volunteer Computing Systems will showcase “an application-aware modular framework that provides self-management for computational multiscale applications in volunteer computing (VC).” Jim Basney from the National Center for Supercomputing Applications is the session chair. The authors’ algorithm KOTree provides application-aware self-management in VC system, and what’s more, they claim it is capable of accurately predicting the expected length of new jobs, leading to an average of 85 percent increased throughput with respect to other algorithms.
Sustainable HPC Cloud Computing
A four-hour workshop takes place on Friday to tackle the important topic of Sustainable HPC Cloud Computing 2012. On the commercial side, there is a robust business case for cloud, however the success of cloud in the HPC sphere has been more limited. This workshop, led by Justin Y. Shi of Temple University, Pavan Balaji of Argonne National Laboratory, and Abdallah Khreishah from the New Jersey Institute of Technology, investigates the technologies that will help cloud meet the special requirements of HPCers. The workshop presenters have brought together a group of researchers to report on performance studies involving a range of computing systems – from the traditional HPC cluster to GPUs for the cloud.
Will Cloud Transform the Film Industry?
Did you know that cloud computing is increasingly the enabling-technology behind your favorite animated films? In this Friday morning panel, Is the Cloud a Game Changer for the Film Industry, Pixar and GreenButton discuss their recently launched cloud-based rendering service. Aimed at the film industry, the service is based on the Windows Azure platform, and connects studios to the cloud, where they can render large production movies. The presenters say this is a game changer for the industry because it gives smaller studios a way to compete with major production companies because it doesn’t involve an expensive capital outlay. One day soon an Oscar will be won by a studio with a globally-distributed workforce but without a major production system.
Cloud started to gain a visible show presence at SC10, but we also saw resistance and pushback. As the use cases for cloud started to accumulate, it gained greater relevancy and acceptance, and this was reflected at last year’s show. The hard work of the research community and the growing ubiquity of the Amazon EC2 cloud helped the subject gain visibility, as did popular stories like Cycle’s 50,000 core virtual supercomputer.
In the past year, HPC cloud has continued to gain momentum as the community reached past the “what”and “why” stages and onto the more difficult “how” stage. Cloud, with or without virtualization, is not for every HPC workload, but for a segment of applications and segment of researchers, it’s growing into a more viable option every year. We’ve seen middleware advances and several IaaS providers have added HPC-friendly hardware, complete with GPUs and InfiniBand. Two events this past September, the EGI Technical Forum and the ISC Cloud event in Mannheim, especially reflected the progress in the HPC cloud space. This year’s SC12 is shaping up to have the same kind of energy and excitement around the enabling qualities of cloud – from public to private and that best-of-both worlds hybrid space.