Enabling code porting to HPC systems and their optimization
IT4Innovations could offer support and assistance for porting of existing codes to HPC systems such as Anselm and other systems under PRACE project. This support could be offered to commercial and in-house codes of our partners. Within this support, profiling and optimization for HPC platforms could be offered as well.
Access with or without assistance
IT4Innovations offers processing time on Anselm supercomputer to our perspective customers. The cluster consist of 207 compute nodes, where each node contains two 8-core CPUs (x86 compatible, SandyBridge architecture); 64GB of RAM; and 500 GB local HDD. The nodes are interconnected using Infiniband QDR network with non-blocking fat tree topology.
Several compute nodes are equipped with heterogeneous accelerators: 23 nodes are accelerated with GPU accelerators Tesla K20m and 4 nodes contains Intel Xeon Phi 5110P. The cluster operating system is Linux. Each node of the cluster has access to shared storage for user data (HOME – 320TB) and temporary data (SCRATCH – 146 TB). More information about the cluster hardware can be found here: http://support.it4i.cz/docs/anselm-cluster-documentation/hardware-overview
Customers interested in IT4I computational resources can use one of the following modes:
- Dynamic allocation without a quote
Using this model, a user pays only for core-hours that are actually used. This is the most affordable way to access and use the cluster. In this case, we do not guarantee the availability of the resources at a time of a request nor the time when jobs are executed. The users with pre-negotiated quotes (model B and C) have higher priority.
- Dynamic allocation for specific number of core-hours (shared access to a hardware)
In this model, a user shares access to the cluster resources with other active users through PBS manager. For every project a quote with total number of core-hours is negotiated. This quote is allocated for the user and it is guaranteed by Supercomputing Center that the user will be able to use all allocated resources if the workload distribution is uniform in time over the project period. In this model, the first half of the allocated core-hours is prepaid, while from the second half the user pays only for used core-hours. This model is recommended for most of the serious customers.
- Static allocation for specific time period (exclusive access to hardware)
In this case, allocated hardware is dedicated to a single user and no other users have access to the same resources. The allocation is done for a certain number of general compute nodes, GPU accelerated compute nodes or MIC accelerated compute nodes for predefined period of time. This is the most expensive option designed for users requiring fastest possible delivery.
Please note that the processing time is charged in core-hours, where 1 core-hour means that 1 CPU core is allocated for 1 hour. Allocation of 1 node for 1 hour is equal to 16 core-hours, since one node has 2×8 CPU cores. The GPU or MIC accelerated nodes are charged in increments of 16 core-hours.
Examples of Assistance:
3D map of Český Krumlov