Work Smarter Not Harder: Platform Analytics Simplifies HPC Infrastructure Analysis
By: Gord Sissons, Product Marketing Manager, Platform Computing
High Performance Computing (HPC) data centers are crucibles of innovation and have pioneered advancements such as distributed cluster computing, parallel programming techniques and smart workload scheduling. While modern HPC data centers run at higher levels of efficiency than their commercial counterparts, there is always a need to process higher volumes of complex data in less time, resulting in additional challenges for data center managers. These challenges include dealing with rapid hardware and software advancements; tight or sometimes shrinking budgets; and the need to balance the demands of competing project teams with shifting priorities.
To boost efficiency, most HPC data centers turn to workload managers, which enable resources to be shared among users and project teams according to policy. However, while workload managers are good at enforcing policies; they can’t determine what those policies should be – and after delivering this “low-hanging fruit” in efficiency gains, further improvements become progressively more difficult.
The key to efficiency lies in providing senior managers and decision makers with better information, which in turn help them make better decisions. To analyze the effectiveness of an HPC environment, it is important to collect information about infrastructure (host models, capacities, networks, OS types); how infrastructure is used (application types, resource usage patterns); clusters and queue configurations (composition, scheduling policies); job-related statistics (run-times, pending times, failure rates, resource usage); projects, users and groups; and license inventories and usage patterns.
Turning Data into Knowledge
Workload managers simplify reporting by gathering and aggregating data into database tables, challenges remain. They include:
- Reporting systems may not incorporate all sources of data, making some questions impossible to answer;
- Reports and underlying data structures are fixed so users can only ask questions that the database schema is designed to readily answer; and
- Workload managers can also be costly to develop and maintain, and answering a new question may require significant development time.
When evaluating analysis and visualization tools for their HPC data centers, organizations should evaluate solutions using the following criteria:
Resource Optimization to Control Costs: By understanding exactly how resources are used, by whom and for what purpose, scheduling policies can be adjusted to provide better utilization and overall efficiency. By turning raw data into usable information, trends and changes in usage patterns become obvious quickly. By visualizing how the need for different applications and platforms are changing with time by project or department, planners can make better quality data-driven decisions more quickly. They can consolidate under-used assets and ensure that new spending is aligned optimally to the needs of the business.
Full Visibility into HPC Data Center Operations: Analysts are able to constantly test and validate planning assumptions and make mid-course corrections as needed. With proper analysis tools they can ensure that SLAs are being met and that business critical projects have ample resources. By analyzing key measures like pending time and license denials across different data dimensions, managers and analysts can be confident that users have access to critical resources when needed, but at minimum cost.
Ability to Identify Bottlenecks: By analyzing resource use and service levels together, administrators quickly spot delays impacting productivity. By understanding underlying causes rather than just symptoms, capacity and performance problems can be solved rapidly, often without incremental cost.
Usage reporting and chargeback accounting: Some organizations like to apportion costs between client departments based on measured resource usage. By combining resource, license and job level data, administrators can track and view resource usage by user, department or project. The rich capabilities of analytics software can make it possible to implement sophisticated chargeback accounting solutions tailored to the needs of the organization.
The Business Intelligence Advantage
A good approach for analyzing HPC infrastructure is the use of on-line analytics processing technology (OLAP) widely used in business intelligence applications. OLAP cubes store measures over multiple data dimensions enabling information to be analyzed and manipulated quickly from multiple perspectives.
The superiority of this analytical approach has led HPC vendors to offer OLAP-based infrastructure analysis solutions, including Platform Analytics from Platform Computing. The main challenge with OLAP is the sheer amount of data that needs to be collected, processed and analyzed. Depending on factors like data volumes and retention policies, data volumes can grow massively. Data sets of several terabytes are common.
Analyzing Efficiency with Rational OLAP Technology
While OLAP represents the best approach for analyzing the effectiveness of HPC environments, its use is usually limited to larger data centers due to the associated cost and complexity. Fortunately, recent innovations including Relational OLAP technology (ROLAP) and fast column-oriented databases now provide the means to address these limitations, making advanced analytics practical for smaller HPC environments as well.
ROLAP technology is an alternative to traditional multi-dimensional OLAP that avoids the pre-computation and storage of information in intermediate formats. Rather, it accomplishes the same functionally with standard SQL queries instead. This allows data center managers and analysts to perform full multi-dimensional analysis while avoiding the cost and complexity of pre-building cubes. With ROLAP-based solutions, users have access to their data immediately without waiting for intermediate data marts and cubes to be built involving multi-step time and resource intensive ETL process.
Parallel, Column-Oriented Databases
Another enabler is new types of grid-oriented databases that use column-based organizational strategies for storing data. Since this approach involves reading columns rather than rows, reads can be parallelized and distributed across multiple compute hosts on a cluster, which is made possible by the columns being independent of one another. With appropriate data replication to ensure integrity, columnar databases can be implemented using a “shared nothing” model and distributed on commodity compute hosts. Scaling the database performance becomes a matter of simply adding hosts.
A higher degree of data compression is also possible because data columns are of a homogeneous type and are stored together. Better compression reduces both data storage requirements and data transfer times. However, once a database server runs out of capacity, they become difficult and costly to enlarge. Database architects are often required to employ clustering technology or expensive SAN solutions to increase capacity.
To illustrate the performance gains, Platform Computing tested a traditional relational database compared to a column-oriented database and found loading 11 million records was measured to be 13 times faster using a column-oriented database. Even more significant, query performance in data sets ranging from 15 million to 1 billion records was measured to be between 78 and 100 times faster – a two orders-of-magnitude improvement.
By exploiting these advances and supplying a powerful new user interface, Platform Computing has developed an analysis and reporting platform that is simpler, more powerful and less costly to deploy and maintain than competing analysis solutions. This means that even smaller HPC environments can now benefit from the insights that advanced analysis tools can deliver.
Platform Analytics 8 is a next-generation analysis and visualization tool for Platform LSF. It enables analysts and managers to answer business-level questions quickly and easily while aggregating job, resource and license-usage data from multiple clusters boosting productivity and enabling data-driven decision-making.
Unlike analytics solutions that require extensive data manipulation to represent data in a usable form, Platform Analytics 8 incorporates a state-of-the-art ROLAP visualization tool. It also features several pre-built “dashboards” designed to cater information to various audiences, including users, project managers, IT personnel, administrators and line of business executives. With Platform Analytics, cluster administrators can “drill” into detailed data to examine the effectiveness of resource sharing policies, while executives can focus on key-performance indicators and relevant cost, productivity and utilization metrics.
As HPC environments increase in complexity they become progressively more difficult for analysts, managers and business planners to fully understand. Small inefficiencies tend to accumulate and multiply over time driving costs, slowing problem identification and resolution, and diminishing productivity.
By applying modern analytic methods pioneered in business intelligence, HPC managers and analysts can gain important new insights into their environments. Platform Analytics 8 leverages these recent advancements to provide rich analysis capabilities for Platform LSF. With better tools, managers and planners have access to higher quality information faster. With better information, they can “work smarter” by realizing gains in efficiency and productivity while simultaneously containing costs.
This article was based on the “Work Smarter Not Harder: Easier Said Than Done?” whitepaper. The full whitepaper is available for download here (registration is required).