Contributed by Jonathan Pool on 2009/09/01.
In “PanLex Performance Recommendations”, Quinn Weaver predicted that an increase in the amount of RAM in the server from the then current 4 GB to at least 16 GB would substantially decrease some execution times.
This document reports the results of an increase in RAM from 4 GB to 8 GB. My intent was to further increment the RAM to 12 GB and then 16 GB so that the functional relationship between RAM and execution time can be determined more precisely than if RAM were increased to 16 GB all at once.
The outputs from the “free” command during the execution of the “sosreport” command with 4 GB and with 8 GB of RAM are as follows:
total used free shared buffers cached
Mem: 4047108 4006364 40744 0 179488 3265232
-/+ buffers/cache: 561644 3485464
Swap: 2031608 628 2030980
total used free shared buffers cached
Mem: 8177908 826936 7350972 0 85780 436936
-/+ buffers/cache: 304220 7873688
Swap: 2031608 0 2031608
After the increase of memory from 4 GB to 8 GB, the PostgreSQL configuration was amended in accord with the recommendations in “PanLex Performance Recommendations”. Then some of the most time-consuming tasks in “UI Tasks 1” were replicated, with the improved character-tabulation algorithm of “UI Tasks 1b” in effect. In all cases, no decrease in execution time was found except decreases attributable to the character-tabulation algorithm revision. A few tasks that were very time-consuming when both hot and cold were also replicated. No substantial decrease in execution time in the hot condition, relative to the cold condition, was observed. Thus, it appears that none of the tasks in “UI Tasks 1” had been constrained by limited RAM. It does not seem beneficial to implement the plan to perform this same set of benchmarks with incrementally increased RAM.