This log contains several months worth of accounting records from the RICC installation in Japan. RICC (RIKEN Integrated Cluster of Clusters) is composed of 4 clusters, and was put into operation in August 2009. The data provided here is from the "massively parallel cluster", which has 1024 nodes, each with 12 GB of memory and two 4-core CPUs, for a total of 12 TB memory and 8192 cores. RIKEN is an independent scientific research and technology institution of the Japanese government. The workload log from the RICC cluster was graciously provided by Motoyoshi Kurokawa (motoyosi@riken.jp). If you use this log in your work, please use a similar acknowledgment.
Downloads:
System EnvironmentAllocation of cores and memory is combined, with a fixed ratio of 1.2 GB of memory per core. Thus jobs that need lots of memory may require more cores than they actually need for the computation.The scheduler is based on project priorities, as decided by a review committee on a quarterly basis. When there is contention for resources, the priority of each job reflects its fair-share value. The fair-share value is reduced as each job is run by the amount of computing time used, and increased continuously at a rate that reflects the project's priority. If there is no contention, the scheduler reverts to FCFS. |
|
In addition, using an SWF parser in conjunction with a general converter module the following anomalies were observed: