The Sun BluePrints Online paper "Configuring Systems for High Bandwidth I/O" by Kevin Colwell and Carolyn Bumatay (login or registration required) describes a recipe for configuring systems to support the very high bandwidth I/O capabilities needed by such data-intensive applications as operational intelligence and surveillance, epidemic trend analysis and prediction, failure analysis of aircraft and ships, predictive traffic management, weather and ocean forecasting, virtual design, astronomy, human genomics, and other scientific disciplines that are data-intensive, and depend on streaming read/write I/O performance.
The subject configuration processes enable users to share massive data sets with a community of hundreds or thousands of users distributed worldwide who need to be able to transfer large subsets of these data sets to local sites or remote resources for processing. The success and continued proliferation of such advanced data sharing depends heavily on high-performance data acquisition, transfer, and storage for real-time data collection, processing, visualization and simulation, the authors write.
The extreme performance requirements of data-intensive applications have extreme performance requirements -- sustained transfer rates necessary for writing data from a server to disk storage at a rate of 10 Gb/sec for a sustained period of time, and simultaneously reading the data from disk storage to another server at 10 Gb/secare needed by many of these applications, the paper notes.
These rates simulate, at a high level, the I/O requirements for applications that must capture data to disk on a storage area network (SAN) at a sustained, high bandwidth, and immediately make that data available to consumers. Sun demonstrated a 10 Gb/sec transfer rate capability using the Sun Fire E25K server in 2006.
The authors report on their effort to demonstrate the same I/O rate at a fraction of the cost, power consumption, and footprint using the latest server, switch, and storage technologies. They explain that, since much has been written about techniques for configuration and performance tuning over local and wide area networks, they have chosen to focus on maximizing I/O bandwidth among the server and storage components of the high-performance architecture. They employed a proof of concept approach to determining the requirements for sustained high performance I/O environments.
Elements of the Sun solution included the following:
- Sun server platformsPerformance was demonstrated using both Sun Fire servers with CoolThreads technology and Sun SPARC Enterprise M5000 systems.
- Sun StorageTek controller modules and disk arraysPerformance and scalability was demonstrated using Sun StorageTek 6540 and Sun Storage 6780 arrays.
- Solaris 10 Operating SystemAll servers in the configuration ran the same version of the operating system.
- Cisco MDS 9124 Fibre Channel switchesThe Cisco switches provided SAN connectivity.
- QLogic QLE2462 Host Bus AdaptersThe QLogic HBAs gave the servers access to the storage resources.
- Sun QFS Shared File SystemThe file system was used to provide high-performance shared access to data.
The authors explain that because, in general, server and storage selection depends on requirements of the specific computational and I/O workload, their first goal in implementing the proof of concept was to demonstrate a minimum of 10 Gb/sec throughput to and from storage using the smallest available physical footprint, and with minimal hardware costs. They selected Sun SPARC Enterprise T5220 servers.
Availability considerations for storage required the use of two separate RAID subsystems. As a result, testing was performed using two Sun StorageTek 6540 controller modules, each supporting up to four disk arrays, the authors write. Once the 10 Gb/sec performance and availability objectives were met, testing aimed to demonstrate scalability of the software to deliver increasing performance with higher-end servers and storage as needed. Further testing has been done on Sun SPARC Enterprise T5240 and T5440 servers, showing excellent I/O scalability of servers with CoolThreads technology, they report. Tests results included in this article also demonstrate capabilities of Sun SPARC Enterprise M5000 servers connected to a SAN using the same switches and the Sun Storage 6780 controller. This configuration used Sun SPARC Enterprise M5000 servers and included eight disk arrays connected to a single Sun Storage 6780 controller module, the paper notes.
Details of the setup and configuration steps taken are provided. The paper also reports on the unit testing and end-to-end workload and availability testing of individual components as performed using an iterative approach. Tunable parameters were changed in the I/O stack or applications, and testing was repeated as needed to determine the optimal settings to maximize sustained throughput, the paper explains.
The authors summarize the results of their work, providing details of the 10 GB/sec SAN results and of the scalability testing. The paper concludes with a section outlining the lessons learned from this configuration exercise.
Sun QFS Installation and Upgrade Guide, and the Sun QFS 4.6 Documentation Collection
Sun SAM and Sun QFS Software Tunables
Sun Servers with CoolThreads Technology
Sun SPARC Enterprise T5220 Server and ILOM Documentation Collection
Sun SPARC Enterprise M5000 Server Documentation Collection
StorageTek 6540 Array and Common Array Manager Software Documentation
StorageTek 6780 Array and Common Array Manager Software Documentation
Read More ...