This paper describes the issues confronted by the climateprediction.net project in creating a volunteer computing project using a large legacy climate model application. This application typically takes from weeks to months to complete one simulation, and has large memory and disk usage requirements. We describe issues in porting the climate model to a single-processor PC platform, checkpointing, computer resources and workunit size for a simulation, and the volunteer computing infrastructures used (a project-specific system and BOINC). We also describe the methods used to obtain and retain users, and examine the retention/attrition rate of users running the lengthy modelling simulations.