mirror of
https://github.com/scylladb/scylladb.git
synced 2026-04-24 10:30:38 +00:00
cdb1499898
The interval class's memory footprint isn't important for single objects, but intervals are frequently held in moderately sized collections. In #3335 this caused a stall. Therefore reducing interval's memory footprint and reduce allocation pressure. This series does this by consolidating badly-padded booleans in the object tree spanned by interval into 5 booleans that are consecutive in memory. This reduces the space required by these booleans from 40 bytes to 8 bytes. perf-simple-query report (with refresh-pgo-profiles.sh for each measurement): before: 252127.60 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.1 tasks/op, 37128 insns/op, 18147 cycles/op, 0 errors) INFO 2025-06-07 21:00:34,010 [shard 0:main] group0_tombstone_gc_handler - Setting reconcile time to 1749319231 (min id=4dbed2f4-43c9-11f0-cbc6-87d1a08b4ca4) 246492.37 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.1 tasks/op, 37153 insns/op, 18411 cycles/op, 0 errors) 253633.11 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.1 tasks/op, 37127 insns/op, 17941 cycles/op, 0 errors) 254029.93 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.1 tasks/op, 37155 insns/op, 17951 cycles/op, 0 errors) 254465.76 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.1 tasks/op, 37123 insns/op, 17906 cycles/op, 0 errors) throughput: mean= 252149.75 standard-deviation=3282.75 median= 253633.11 median-absolute-deviation=1880.17 maximum=254465.76 minimum=246492.37 instructions_per_op: mean= 37137.24 standard-deviation=15.71 median= 37127.54 median-absolute-deviation=14.45 maximum=37155.24 minimum=37122.79 cpu_cycles_per_op: mean= 18071.19 standard-deviation=212.25 median= 17950.62 median-absolute-deviation=130.10 maximum=18411.50 minimum=17906.13 after: 252561.26 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.1 tasks/op, 37039 insns/op, 18075 cycles/op, 0 errors) 256876.44 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.1 tasks/op, 37022 insns/op, 17785 cycles/op, 0 errors) 257084.38 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.1 tasks/op, 37030 insns/op, 17840 cycles/op, 0 errors) 257305.35 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.1 tasks/op, 37042 insns/op, 17804 cycles/op, 0 errors) 258088.53 tps ( 66.1 allocs/op, 0.0 logallocs/op, 14.1 tasks/op, 37028 insns/op, 17778 cycles/op, 0 errors) throughput: mean= 256383.19 standard-deviation=2185.22 median= 257084.38 median-absolute-deviation=922.16 maximum=258088.53 minimum=252561.26 instructions_per_op: mean= 37032.17 standard-deviation=8.06 median= 37030.46 median-absolute-deviation=6.44 maximum=37041.83 minimum=37021.93 cpu_cycles_per_op: mean= 17856.60 standard-deviation=124.70 median= 17804.16 median-absolute-deviation=71.24 maximum=18075.50 minimum=17777.95 A small improvement is observed in instructions_per_op. It could be random fluctuations in the compiler performance, or maybe the default constructor/destructor of interval are meaningful even in this simple test. Small performance improvement, so not a backport candidate. Closes scylladb/scylladb#24232 * github.com:scylladb/scylladb: interval: reduce sizeof interval: change start()/end() not to return references to data members interval: rename start_ref() back to start() (and end_ref() etc). interval: rename start() to start_ref() (and end() etc). test: wrapping_interval_test: add more tests for intervals