diff options
Diffstat (limited to 'src/mongo/db/query/cost_model/cost_model_manager.cpp')
| -rw-r--r-- | src/mongo/db/query/cost_model/cost_model_manager.cpp | 88 |
1 files changed, 35 insertions, 53 deletions
diff --git a/src/mongo/db/query/cost_model/cost_model_manager.cpp b/src/mongo/db/query/cost_model/cost_model_manager.cpp index 917f84bc8b8..b013791f64f 100644 --- a/src/mongo/db/query/cost_model/cost_model_manager.cpp +++ b/src/mongo/db/query/cost_model/cost_model_manager.cpp @@ -38,76 +38,58 @@ namespace { */ void initializeCoefficients(CostModelCoefficients& coefficients) { // These cost should reflect estimated aggregated execution time in milliseconds. - constexpr double ms = 1.0e-3; + // The coeffeicient ns converts values from nanoseconds to milliseconds. + constexpr double nsToMs = 1.0e-6; - // Startup cost of an operator. This is the minimal cost of an operator since it is - // present even if it doesn't process any input. - // TODO: calibrate the cost individually for each operator - coefficients.setDefaultStartupCost(0.000001); + coefficients.setDefaultStartupCost(1.0 * nsToMs); - // TODO: collection scan should depend on the width of the doc. - // TODO: the actual measured cost is (0.4 * ms), however we increase it here because currently - // it is not possible to estimate the cost of a collection scan vs a full index scan. - coefficients.setScanIncrementalCost(0.6 * ms); - coefficients.setScanStartupCost(0.000001); + coefficients.setScanIncrementalCost(422.31145989 * nsToMs); + coefficients.setScanStartupCost(6175.527218993269 * nsToMs); - // TODO: cost(N fields) ~ (0.55 + 0.025 * N) - coefficients.setIndexScanIncrementalCost(0.5 * ms); - coefficients.setIndexScanStartupCost(0.000001); + coefficients.setIndexScanIncrementalCost(403.68075869 * nsToMs); + coefficients.setIndexScanStartupCost(14054.983953111061 * nsToMs); - // TODO: cost(N fields) ~ 0.7 + 0.19 * N - coefficients.setSeekCost(2.0 * ms); - coefficients.setSeekStartupCost(0.000001); + coefficients.setSeekCost(1174.84136356 * nsToMs); + coefficients.setSeekStartupCost(7488.662376624863 * nsToMs); - // TODO: take the expression into account. - // cost(N conditions) = 0.2 + N * ??? - coefficients.setFilterIncrementalCost(0.2 * ms); - coefficients.setFilterStartupCost(0.000001); + coefficients.setFilterIncrementalCost(83.7274685 * nsToMs); + coefficients.setFilterStartupCost(1461.3148783443378 * nsToMs); - // TODO: the cost of projection depends on number of fields: cost(N fields) ~ 0.1 + 0.2 * N - coefficients.setEvalIncrementalCost(2.0 * ms); - coefficients.setEvalStartupCost(0.000001); + coefficients.setEvalIncrementalCost(430.6176946 * nsToMs); + coefficients.setEvalStartupCost(1103.4048573163343 * nsToMs); - // TODO: cost(N fields) ~ 0.04 + 0.03*(N^2) - coefficients.setGroupByIncrementalCost(0.07 * ms); - coefficients.setGroupByStartupCost(0.000001); + coefficients.setGroupByIncrementalCost(413.07932374 * nsToMs); + coefficients.setGroupByStartupCost(1199.8878012735659 * nsToMs); - coefficients.setUnwindIncrementalCost(0.03 * ms); // TODO: not yet calibrated - coefficients.setUnwindStartupCost(0.000001); + coefficients.setUnwindIncrementalCost(586.57200195 * nsToMs); + coefficients.setUnwindStartupCost(1.0 * nsToMs); - // TODO: not yet calibrated, should be at least as expensive as a filter - coefficients.setBinaryJoinIncrementalCost(0.2 * ms); - coefficients.setBinaryJoinStartupCost(0.000001); + coefficients.setBinaryJoinIncrementalCost(161.62301944 * nsToMs); + coefficients.setBinaryJoinStartupCost(402.8455479458652 * nsToMs); - coefficients.setHashJoinIncrementalCost(0.05 * ms); // TODO: not yet calibrated - coefficients.setHashJoinStartupCost(0.000001); + coefficients.setHashJoinIncrementalCost(250.61365634 * nsToMs); + coefficients.setHashJoinStartupCost(1.0 * nsToMs); // Already calibrated. - coefficients.setMergeJoinIncrementalCost(0.02 * ms); // TODO: not yet calibrated - coefficients.setMergeJoinStartupCost(0.000001); + coefficients.setMergeJoinIncrementalCost(111.23423304 * nsToMs); + coefficients.setMergeJoinStartupCost(1517.7970800404169 * nsToMs); - coefficients.setUniqueIncrementalCost(0.7 * ms); - coefficients.setUniqueStartupCost(0.000001); + coefficients.setUniqueIncrementalCost(269.71368614 * nsToMs); + coefficients.setUniqueStartupCost(1.0 * nsToMs); // Already calibrated. - // TODO: implement collation cost that depends on number and size of sorted fields - // Based on a mix of int and str(64) fields: - // 1 sort field: sort_cost(N) = 1.0/10 * N * log(N) - // 5 sort fields: sort_cost(N) = 2.5/10 * N * log(N) - // 10 sort fields: sort_cost(N) = 3.0/10 * N * log(N) - // field_cost_coeff(F) ~ 0.75 + 0.2 * F - coefficients.setCollationIncrementalCost(2.5 * ms); // 5 fields avg - coefficients.setCollationStartupCost(0.000001); + coefficients.setCollationIncrementalCost(2500 * nsToMs); // TODO: not yet calibrated + coefficients.setCollationStartupCost(1.0 * nsToMs); // TODO: not yet calibrated - coefficients.setCollationWithLimitIncrementalCost(1.0 * ms); // TODO: not yet calibrated - coefficients.setCollationWithLimitStartupCost(0.000001); + coefficients.setCollationWithLimitIncrementalCost(1000 * nsToMs); // TODO: not yet calibrated + coefficients.setCollationWithLimitStartupCost(1.0 * nsToMs); // TODO: not yet calibrated - coefficients.setUnionIncrementalCost(0.02 * ms); - coefficients.setUnionStartupCost(0.000001); + coefficients.setUnionIncrementalCost(111.94945268 * nsToMs); + coefficients.setUnionStartupCost(69.88096657391543 * nsToMs); - coefficients.setExchangeIncrementalCost(0.1 * ms); // TODO: not yet calibrated - coefficients.setExchangeStartupCost(0.000001); + coefficients.setExchangeIncrementalCost(100 * nsToMs); // TODO: not yet calibrated + coefficients.setExchangeStartupCost(1.0 * nsToMs); // TODO: not yet calibrated - coefficients.setLimitSkipIncrementalCost(0.0 * ms); // TODO: not yet calibrated - coefficients.setLimitSkipStartupCost(0.000001); + coefficients.setLimitSkipIncrementalCost(62.42111111 * nsToMs); + coefficients.setLimitSkipStartupCost(655.1342592592522 * nsToMs); } } // namespace |