Advance reservations: a theoretical and practical comparison of GUR & HARC

ABSTRACT

Grid computing has provided new capabilities for connecting distributed computing nodes, experimental instruments, visualization facilities, data storage and people across high speed networks. Emerging communities of "e-Scientists" are now building cyber-infrastructures for inter-disciplinary scientists to exploit Grids for new modes of problem solving in science, engineering and the humanities.

One new need, that these cyber-infrastructures require, is to be able to co-allocate these varied resources. For example, multiple machines can be reserved at the same time for meta-computing with large scale tightly coupled simulations; visualization equipment and networks can be co-allocated for real time visualization and interaction; data management tools and high speed networks can be immediately reserved as soon as an experimental device, such as a particle detector, starts recording terabytes of data. As scientists start developing complex workflows for their work, the needs for sophisticated and flexible co-allocation will increase.

Traditionally, scientists submit their supercomputing tasks using batch scheduler, typically running in a backfill mode. Jobs are added to queues, and will run at some unknown time which may be in days or weeks time. Networks, visualization equipment, and data archives are typically scheduled through human reservations.

In this poster we examine two new tools which have the ability for advance reservations and co-allocation. HARC (Highly-Available Resource Co-allocator), developed at CCT LSU through the NSF EnLIGHTened Computing project, was motivated by the need to dynamically provision 10Gbps Ethernet connections for applications. GUR (Generic Universal Remote), deployed on the IA-64 clusters at NCSA (Mercury) and SDSC, is a metascheduler developed to automatically manage and create overlapping reservations at multiple sites. Both tools have been deployed on the NSF TeraGrid. And HARC is also deployed on subsets of LONI, UK NGS and NorthWest Grid. These two modern toolkits will be analyzed from both theoretical and practical points of view. For the theoretical comparison, we analyze them in terms of their software engineering and architecture based on their design philosophies. For the practical part, they are compared from the end user point of view. In total, the poster covers seven aspects of comparison: architecture; installation; usability; portability; reliability; security, and functionality. The analysis shows interesting differences between the service-based HARC and more intuitive designed GUR, which may shed some light on grid software design and applicability for real world applications. Advantages, disadvantages and problem-oriented nature of GUR and HARC are studied and exemplified.