Use Cases

CLUES is currently being used in different equipments. In the following, we show some singular cases, detailing the advantages provided by the use of this tool.

ODIN (parallel cluster)

Odin is a HPC cluster used as a parallel computing resource by the GRyCAP research group. As it is often the case with this kind of clusters, it undergoes periods of high utilization (e.g. to perform tests for the publication of a paper) and underutilization (e.g. holidays).

Singular features:

• SCI Network with a 10x5 2D Torus topology
• Torque/PBS batch-queuing system

One of the peculiarities of this cluster is the fact that the SCI network requires all the nodes to be powered on in order to achieve maximum performance in the case of parallel job submissions. CLUES is configurable and can cope with this kind of situations easily by simply adjusting its configuration parameters.

After a six-month cluster usage analysis, it was observed that the cluster was inactive during more than 80% of time. Details about the evaluation of CLUES usage impact on energy consumption are provided below:

* Cost: 0,091 €/kw. Data obtained from the Ministerio de Industria, Turismo y Comercio of Spanish Government.

With the results shown, the accumulated saving during the 6 months would be 1,629.82 €, which is nearly 50% of the total energy consumption cost of the cluster.

El Superordenador Caléndula

Calendula is the supercomputer of the Castilla y León Supercomputing Center, installed in the city of León, with the following features:

• 304 blade servers.
• A total of 2432 cores for computation.
• Infiniband interconnection network.
• Open Grid Scheduler/Grid Engine for task management.

As in many supercomputers, the periods in which the full power of the system is used are quite short. There is a combination of short periods with intensive use of many cores, with long periods of execution of jobs that only use a fraction of the full power. CLUES provides the flexibility to have the nodes available only when they are demanded by the jobs, and keep them powered off for the rest of the time. Given the large number of servers available in Calendula, this reduces energetic consumption to a great extent.

The nodes are switched on by means of WoL when user jobs enter the task manager, and are powered off 10 minutes after the task has been completed.

Weather forecasting is one of main areas of computing that Calendula is currently used for, with a part of the resources reserved exclusively for this task. Since there are four executions each day, CLUES enables to use the servers only at the hours at which the simulations are running.

NGI UPV-GRyCAP Node

This cluster is a node of the Spanish National Grid Initiative (NGI-ES) and it uses the CLUES system for its normal operation. Given the diversity of resources available in NGI-ES, the amount of jobs submitted to the cluster is normally low, and the cluster could be partially off.

Singular features:

• Blade-type nodes with 16 cores per node.
• System used: gLite 3.2 and Torque/PBS

This cluster has two interesting aspects, which are the multiple cores per node, and the requirement of keeping several nodes powered on in order to improve the integration with gLite.

This is a gLite Computing Element and is monitored by NGI-ES systems. These monitoring systems periodically submit execution tests to the cluster and, given the frequency of these tests (around one test every 10 minutes) and in order to avoid any of these tests having to wait, it was decided to keep two of the nodes always on.

On the other hand, CLUES can also take into account the availability of several cores per node, in order to reduce the number of actual nodes to switch on.

Details about the evaluation of CLUES usage impact on energy consumption are provided below:

* Cost: 0,091 €/kw. Data obtained from the Ministerio de Industria, Turismo y Comercio of Spanish Government.

With the results shown, the accumulated saving during the 5 months would be 62,77 €, which is nearly 20% of the total energy consumption cost of the cluster.

In this case the accumulated saving is lower than the previous case because this cluster only is composed by 7 nodes, and 2 of them must be always on. Furthermore the blade enclosure has relatively high power consumption regardless of whether the nodes are on or not.

Kefren (virtualization platform)

This cluster is being used as a virtualization platform, based on OpenNebula 2.2 and VMWare Server 2.0. The system is also fully integrated with CLUES, so that those nodes that are not hosting any virtual machine are powered off.

Singular features:

• Cluster-type nodes with 2 processors per node.
• System used: OpenNebula 2.2 and VMWare Server 2.0.

In this case the policy defined for the cluster usage considers one virtual processor for each real core, and requires that the sum of memory of the virtual machines running in a node must not exceed the actual node memory. Whenever CLUES detects that a virtual machine is to be deployed, it verifies the requirements, switching on nodes if necessary.

Due to OpenNebula features, we are currently considering green-aware scheduling policies. Since a simple versatile API is available to access the cloud infrastructure control system, development has been started of a scheduler that takes into account CLUES subsystem during the assignment of virtual machines to real machines.

Details about the evaluation of CLUES usage impact on energy consumption are provided below:

* Cost: 0,091 €/kw. Data obtained from the Ministerio de Industria, Turismo y Comercio of Spanish Government.

With the results shown, the accumulated saving during the 5 months would be 527,51 €, which is more than 40% of the total energy consumption cost of the cluster.

Hyades (research cluster)

Hyades is a cluster used by different research groups of the Universitat Politècnica de València. It is, however, a relatively old resource, for which reason it is being less and less used, although there are still researchers that use it in different periods (algorithm testing, data collection, etc.)

Singular features:

• Old cluster with 60 nodes
• Power management: proprietary system

The lack of continuous usage of the cluster made the administrator consider switching it off definitely, so as to avoid the energy consumed by a resource of this kind. However, the integration with CLUES system enabled the cluster to be in use normally, while reducing the consumption from that of 60 nodes to only one front-end node, and keeping the resource available for the user community.

Due to the fact that the cluster is old, it does not support features such as Wake-on-LAN, nor the complete switching off by means of ACPI. Because of this, it was necessary to integrate the switching on/off processes with the proprietary power management subsystem previously installed in the cluster.