From IDGF-SP Financial and energetic analysis of volunteer computing and desktop grids
Jump to: navigation, search

The attached spreadsheet is a model created a Sony CSL Paris to compare the costs and energy consumption between a volunteer computing network and a computing cluster: File:Financial-comparison.xls

Sony CSL evaluated the energy requirements of volunteer computing compared to a dedicated computing cluster in its study on the commercial opportunities for volunteer computing services. The comparison takes as target audience, a small research group that has to invest in a computing infrastructure to run CPU-intensive scientific models. Two solutions are compared, a dedicated computing cluster and a volunteer computing network. The comparison starts with a given computing performance that is the same for both solutions. This performance is the desired, sustained performance of the infrastructure. It uses the CFP2006 benchmark as the unit of reference.

The evaluation includes:

  • The purchase cost of the equipment,
  • The installation cost,
  • The cost of the energy consumption, and
  • The salaries of the staff that maintains the infrastructure.

Several factors are also missing in this estimate, including:

  • the cost of adapting the building to house the cluster and the cooling system. These costs are too site specific but can be very significant.
  • the cost of the Internet connection. The bandwidth required to run a large volunteer computing network may be substantial.

For each of the two solutions, the comparison makes a low-end estimate and high-end estimate. Both solutions present many configuration options (best-of-class machines versus best-bargain machines, cost of salaries, and so on). We are therefore interested in the range of possibilities.



In our study, we estimated the cost for a fairly ambitious computing project. The size of this project is, however, very realistic. The performance of the infrastructure was set at 600000 CFP2006 benchmarks.

The computing cluster requires between 8800 and 10400 computing cores (1100 and 2600 computing nodes). The volunteer computing network, to obtain a similar performance, required between 2200 and 64000 participants.

Using this very simple model we estimate that for the cost:

  • a computing cluster would cost between 1.5 and 4 million euros PER YEAR.
  • an equivalent volunteer computing network requires between 195.000 and 490.000 Euro per year.

For the energy consumption, we obtained the following:

  • the computing cluster would consume between 2.9 and 4.4 GWh (1 GWh = 1 million kWh) per year.
  • the volunteer computing network would consume 2.9 and 12.2 GWh per year.


What we see is that the volunteer computing network is always cheaper. However, from these results, it doesn't seem likely that volunteer computing networks consume less than computing clusters. An easy analysis of the energy consumption would be that data centres consume more due to the need of a cooling infrastructure. If the cooling equipment requires 50% of the thermal dissipation power of the computing nodes (a rough estimate), than we would conclude that data centres would requite 150% more energy than volunteer computing networks, who don't require the extra cooling. However, other factors come into play. Some of the most important are the performance/energy of the computing nodes, the efficiency of the cooling, and the amount of duplication of the work-units in volunteer computing. Scientific projects will send some of the work-units to several volunteers, for example, to verify the results or to restart a work-unit when one of the participants is no longer responsive.

Until we have more detailed information about the participating machines in a volunteer computing network, we would estimate that volunteer computing and an equivalent computing cluster consume roughly equal amount of energy.

Another result is that the volunteers contribute between 12 and 140 euros per year through increases in their electricity bill, with a likely average of 40 euros/person/year.

Global parameters

Parameter Description
CFP2006 Benchmark The total requested computing performance, expressed using the CFP2006 benchmark as a unit.
CPUs The lists of CPUs consider in the comparison, together with their CFP2006 benchmark values, and an estimate of their power consumption. The CPU benchmark can be estimated from official results list at [1].
Cost of electricity The average cost of 1 kWh of electricity.

Cluster parameters

The list of parameters for the cluster are:

Parameter Description
The computing node The system that is selected as the basic building block for the cluster. Together with its price and the number of CPUs (the CPU should be in the global list of selected CPUs). The base node can be selected in the catalog of major hardware vendors.
Installation costs The cost to install a large cluster. It is hard to evaluate this figure. The installation can be performed by trained personnel that is already employed or by the hardware vendor through a service contract. Prices will vary.
Depreciation time How many years is the cluster expected to be used before being replaced.
Residual value The value of the equipment at the end of its use, in percentage of its initial cost.
Cooling/TDP ratio The amount of energy used for cooling as a ratio of the energy used by the computing equipment.
Number of network switches Additional network equipment to connect the servers
Price of network switch Estimated price of a network switch.
Number of engineers How many people does it require to maintain the cluster
Net salary The average salary (net) of the system engineers.

Volunteer computing parameters

Parameter Description
Active/inactive ratio The PCs of volunteer are rarely used continuously for running the scientific applications. This ratio indicates on average across the community how many hours a day a machine is available.
Throttle ratio The BOINC software allows to run the computation periodically, with short breaks in between. This settings reflects the average throttling ratio used by the volunteer community.
Work duplication To verify the results, BOINC can schedule to same work unit to several volunteers. This multiplication reduces the risk of erroneous results. It also reduces to total performance of the network.
Extra energy consumption The amount of energy that is required per machine to run the computation.
Price per server The purchase price of the servers needed to run the network (distribute work units, collect results).
Number of network switches Additional network equipment to connect the servers
Price of network switch Estimated price of a network switch.
Depreciation time How long is the network infrastructure used before being replaced.
Residual value The value of the equipment at the end of its use, in percentage of its initial cost.
Number of employees The number of staff needed to maintain the infrastructure and to provide support to the volunteer community.
Net salary The average salary (net) of the system engineers.

Download & copyright


Disclaimer: This model has not been reviewed by peers.

Copyright: Sony Computer Science Laboratory

License: GPLv3

Author: Peter Hanappe