This unit introduces the fundamental building blocks of grid computing, explaining the components and how they interact to form a functional system.
1. Core Grid Resources
A grid is a collection of machines (nodes) that contribute resources. The three primary resources are:
- Computing Cycles (Processing Power): The most common resource, provided by the machine's processors. It can be exploited in three main ways:
- Running an existing application on a remote grid machine.
- Running a parallel application designed to split work across multiple processors.
- Running the same application many times on different machines.
- Scalability measures how efficiently adding processors improves performance. Perfect scalability means doubling the processors halves the computation time.
- Unifying File Systems (like AFS, NFS, GPFS) provide a single, uniform view of all storage, increasing capacity and making data access location-independent.
- Data Striping spreads data across multiple disks to enable faster read/write speeds.
- Other features include journaling for reliability and synchronization for managing shared data.
- Data Communication (Network Capacity): The bandwidth for moving data and jobs within and outside the grid.
- It's a critical resource, as moving data to the machine where a job runs can create bottlenecks.
- Redundant and high-speed networks are used to improve reliability and performance for data-intensive tasks.
2. Specialized and Logical Resources
Beyond the physical resources, grids manage other important assets:
- Software and Licenses: Expensive or licensed software can be installed on only a few machines. The grid scheduler then sends jobs requiring that software to those specific machines. License management software ensures the number of concurrent uses doesn't exceed the purchased limit.
- Platform Attributes: Machines have different architectures (e.g., PowerPC vs. x86), operating systems, and hardware. Jobs must be assigned to machines that match their specific platform requirements.
- Policy-Based Attributes: Administrators can create "artificial" resources, like a medical_research flag, to enforce policies. This ensures that certain jobs (e.g., medical research) only run on designated machines.
3. Workload Management: Jobs, Applications, and Scheduling
Work is organized and executed on the grid through a structured process.
- Jobs and Applications: An application is a larger piece of work, which is broken down into one or more jobs. Jobs are the actual programs executed on grid machines. They can have dependencies, meaning some jobs cannot start until others have finished or until specific data is available.
- Scheduling and Resource Brokering: A scheduler (or resource broker) automatically finds the most appropriate machine for a pending job based on resource availability and job requirements.
- Execution Models:
- Scavenging: Jobs run only on idle machines at the lowest priority. They are suspended or delayed if the machine's owner needs it. This is unobtrusive but leads to unpredictable completion times.
- Dedicated: Machines are reserved exclusively for grid computing, providing more predictable performance.
- Reservation: Resources can be booked in advance (like a meeting room) to guarantee quality of service and meet deadlines. Idle reserved resources can sometimes be scavenged for lower-priority jobs.
4. Grid Typologies and Evolution
Grids can be built at different scales, with increasing complexity in management, policy, and security.
- Simple/Homogeneous Grid: A few machines of the same type in a single department (sometimes called a cluster).
- Intragrid: A larger grid within a single organization, often involving heterogeneous machines, multiple departments, and requiring policies for usage and security.
- Intergrid: A grid that spans multiple organizations, used for collaboration. This requires the highest levels of security and may involve trading or brokering resources.