English (En)French (Fr)
Home  //  Recherche Research Results Optimizing environmental and financial costs of HPC systems
Erreur
  • Unable to load Cache Storage: database
  • Unable to load Cache Storage: database
  • Unable to load Cache Storage: database
  • Unable to load Cache Storage: database
  • Unable to load Cache Storage: database
  • Unable to load Cache Storage: database
  • Unable to load Cache Storage: database
  • Unable to load Cache Storage: database
  • Unable to load Cache Storage: database
  • Unable to load Cache Storage: database
Optimizing the environmental and financial costs of high-performance computing systems
PDF
Imprimer
Envoyer
There are no translations available.

Data centres containing high-performance computing (HPC) clusters may be able to coordinate with the operation of wind farms for mutual benefit.   Renewable energy sources such as wind, hydro, and solar exhibit significant variability and volatility of output, but there exists some flexibility in HPC systems to delay processing workload until energy is available at low cost.  Data centres consume megawatts of power, with this energy usage typically accounting for a majority of lifecycle carbon emissions as well as  a significant portion of the total cost of ownership.
We thus explored the potential of several different approaches to reduce the financial and environmental costs of computation: adapting to dynamic electrical pricing, variations in the carbon intensity of the electrical grid, or the availability of local renewables.  This was done by comparing an optimum allocation of workload across 1 day or 1 week time periods to randomly assigned time periods.  Adaptation to energy price and local energy availability were the most successful, with different portions of the workload may have different priorities, and thus we also explored how a scheduler might account for a portion of the workload marked as interactive.  We thus examined the impact of running interactive computing jobs immediately upon arrival, while delaying the remainder of the workload until low cost energy was available.
Testing was performed using workloads from the Parallel Workloads Archive1 along with power data for Alberta, Ontario, and Texas and wind generation data from the National Renewable Energy Laboratory’s wind integration dataset2.

 

Adapting to…


Dynamic electrical pricing

The price of power in real-time markets varies widely throughout the course of a typical day, responding patterns in energy production and demand, and unexpected situations may produce sharp spikes in prices that may account for a large fraction of the total energy cost.  We thus examined the potential savings in financial cost for electricity for several workloads.  If willing to defer jobs to later days greater savings might be seen but this comes at the cost of increased wait time, the extent of which we also evaluated.  Below an example of the extent of potential cost savings in Alberta can be seen.  In other regions in which the price is permitted to drop below zero, a workload with low utilization might even achieve a negative power bill.

Variations in carbon intensity

Variations in carbon intensity, here weighting each time period by the fraction of energy derived from each source type in Alberta produced only weak results as high-carbon-intensity coal plants are left at roughly constant output as they operate most efficiently in such a manner.  Several hypothetical scenarios involving low carbon grids were also examined.

 

Availability of local renewables

As data centres can be situated distant from population centres while still able to communicate effectively through networks, we also examined the results if situating a data centre adjacent to wind turbines and optimizing energy consumption based on the output of those turbines.  Generation data was taken from several Montana wind turbines from the NREL Western Wind Integration Dataset, and data centres of various levels of maximum power consumption were considered.
Below shows the potential percentage increase in use of locally generated renewables relative to a cap on the time that grid energy could be used during lulls in wind output.   Decreasing returns were seen as the size of installed wind capacity increased relative to the size of the data centre’s maximum power demand.

Capping the percentage of time that local energy can be used may be an effective strategy to point, but beyond a point the waiting time of the workload may begin to increase dramatically.

Incorporating interactive jobs

Different user requests have different priorities and thus we examined the impact of separately treating high priority jobs – in this case interactive job requests.  Below can be seen the average price of electricity when interactive jobs arrived in two scenarios, illustrating how such requests typically arrive when the price of electricity is above average.

 

Significant savings can still be seen for the non-interactive portion of workload as we demonstrate in our work, and we also show some preliminary results regarding how migration of jobs to a different time zone reduce power costs for the interactive portions of some but not all workloads.

 

More details can be found in:

  • David Aikema and Rob Simmonds, Electrical cost savings and clean-energy usage potential for HPC workloads, Proceedings of the 2011 IEEE ISSST International Symposium on Sustainable Systems and Technology, 2011.
  • David Aikema, Cameron Kiddle and Rob Simmonds, Energy-cost-aware Scheduling of HPC Workloads, Proceedings of the 1st International Workshop on Sustainable Internet and Internet for Sustainability (SustaInet), 2011.

 

Mise à jour le Vendredi, 26 Août 2011 18:02
 

Canarie Logo