Policy |
 |
Pension Projection Models |
Key Feature: Fast Running
While the input specification and output analysis ease-of-use features make the PSG models fast to use, they are also fast when executing a run (that is, reading the input databases, performing the requested simulation, and writing the output files). The execution time of a run can vary substantially depending on the nature of the run and on the kind of computer running the model.
The PSG models are designed to take advantage of multi-core computers when operating in OLC mode, so run times will fall substantially in coming years as standard computers have increasingly many cores.
Execution times for several types of runs are described below, followed by notes on the feasibility of achieving massive throughput by running the Linux version of the PSG models on many cloud-computing servers that can be rented inexpensively by the hour.
The following execution times are for PSG model runs that use 2007 Trustees Report intermediate-cost and other baseline assumptions and that execute under Windows XP on an Alienware workstation equipped with two dual-core AMD Opteron 280 chips running at 2.4 gigahertz. This computer is equipped with a SATA disk drive rotating at 7200 RPM and with two gigabytes of RAM, which is far more memory than needed by the PSG models even when eight threads of execution are being used in an OLC mode run.
A single-scenario, non-stochastic CBA mode run executes in less than one second.
A stochastic CBA mode run, which simulates 1000 macro scenarios and uses the standard 75 year social security actuarial test period executes in about 0.8 minutes.
A single-scenario, non-stochastic OLC mode run, which generates a one-in-a-thousand sample for each of 150 cohorts born since 1935 and uses the standard 75 year social security actuarial test period, executes in about 1.8 minutes. This execution time is for an OLC run that simulates eight birth cohorts simultaneously, which suggests that moving up to a computer with two quad-core chips of the same speed would cut the execution time in half. If the disk drive became a bottleneck with eight cores, it would be easy and inexpensive to add another disk drive and upgrade to a RAID 0 disk configuration, which would most likely eliminate the bottleneck.
A single-scenario ESP mode run, which generates a two percent sample of the 1990 birth cohort, simulates employer-sponsored pensions but not social security, and writes detailed pension output, executes in about 1.9 minutes. Note that the two percent sample excludes the spouses of sample individuals, so the total number of life histories simulated is more than twice as many as the 104,000 sample individuals included in the output files.
A single-scenario RCS mode run, which generates a two percent sample of the 1990 birth cohort, simulates current-law-policy social security but not pensions, and writes detailed social security output, executes in about 2.0 minutes.
A single-scenario RCS+ESP mode run, which generates a two percent sample of the 1990 birth cohort, simulates both pensions and social security, and writes detailed output, executes in about 3.1 minutes.
Stochastic runs take proportionately longer, so stochastic OLC mode runs will require many hours to execute even on an eight-core computer. Experience has shown that analytically useful stochastic ESP and RCS mode runs can be specified using a modest per-scenario sample size that implies run times of only about an hour.
These run execution times are very fast relative to other models that simulate U.S. pensions and social security. Most users will find run execution speed to be no problem when using the PSG models.
The problem of long stochastic OLC mode run times, plus the unusual need of one model user to execute from several hundred to a few thousand runs at a time, led to the development in early 2008 of a Linux version of the dbrun-create utility and the three models. The idea is to position the PSG models to take advantage of grid computing opportunities offered by cloud-computing vendors who rent Linux servers by the hour.
Other key features.
This page was last revised on July 30, 2008.