Hardware Engineering |
![]() |
HERD Project - Highly Efficient Resource Distribution
The aim of this project is to design and build a cluster of workstations (COWs) from low-cost commodity computers for electronic structure calculations and potential energy surface studies using both ab initio (MO) and density functional theories (DFT). While DFT calculations mainly memory and CPU intensive, ab initio MO calculations require fast hard drives such as the Serial-ATA or more preferably SCSI devices.
In a small research group running limited number of software packages, high efficiency can be achieved by fine tuning the hardware resources to the particular software package. In my research, I use the Gaussian and Amsterdam Density Functional packages. The Gaussian is site licensed at the Chemistry and Biochemistry department, however the ADF licensing is node locked, i.e. only dedicated COWs can run ADF jobs, which also supports the idea of individually designed COWs for dedicated purposes.
The design idea is to start with a powerful master node (CALF), which is capable of efficiently executing input/output intensive ab initio jobs and also has fast processors for density functional calculations. Dual CPU boxes are excellent for this purpose, since the second CPU - if not in use - can handle the overheads from out-of-box parallelization, the extensive I/O generated system load and keep the workers busy. A worker node ideally consists of single CPU, diskless workstations, which are connected to the master node through a dedicated network for computation related activities. As the computational demand increases additional single CPU workers can be added to the growing COWs or new CALFs can be created. Depending on the computational demand and the number of users, the former expansion will speed up the computations for the same number of users, while the latter can accommodate additional user demands.
Each node in the HERD has an additional network card and connected to an end-user workstation for job monitoring and maintenance. The master nodes (COW heads) are connected to a keybord-monitor switchbox for quick troubleshooting.The Buffalo Project
Due to the availability of large amounts of after-hour computational power at the Department - it was quite tempting to configure and link the existing resources as part-time COWs. During the day, these computers are used in general, organic and physical chemical laboratory exercises as data acquisition workstations (MicroLab). Between 6pm and 7am and on holidays, they perform computational chemical calculations. Each group of part-time COWs is controlled from a dual AMD Athlon master node with MP2600+ (BISONs), which can provide access to the output files around the clock. A protocol for dynamic allocation of the number of workers is being developed, which could provide an on-availability based parallelization. Computational jobs could run faster overnight and on holidays and slower during the day, while the worker nodes (part-time COWs) are used for their primary purposes.
Configured part-time COWs:
- General Chemistry Lab: 44 Dell Optiplex 260 w/ P4 2260MHz
- Physical Chemistry Lab: 7 Dell Optiplex 270 w/ P4 2400MHz
- Organic Chemistry Lab: 6 Dell Optiplex 270 w/ P4 2400MHz
Other COWs in the field of computational chemistry:
- Solomon Group - Stanford University (80 CPU cluster)
- Morokuma Group - Emory University (150 CPU cluster)
News, headlines, good-to-knows:
- IDF Spring 2006: Will Intel's Core Architecture Close the Technology Gap?
- BIOS A-Z: All you ever wanted to know about BIOS.
- Oil cooling: Mark Reinholz' idea
- 4.1GHz PC?