perf improvement - Summary metric

[mdebjit]

CKMS Quantiles/Summary Performance Improvements

• Optimized insertion: Replaced linear search with binary search in the insertBatch algorithm, improving insertion speed.

• Enhanced readability & maintainability: Streamlined index management in both the compress and insertBatch algorithms, resulting in cleaner, more maintainable code.

• Prevent undefined behavior: Addressed potential issues in the compress method by eliminating vector element erasure during iteration, preventing iterator invalidation and out-of-bounds errors.

• Pre-allocation: Reserved vector capacity upfront when the size is known, eliminating multiple resize operations and significantly enhancing performance.

Benchmark	Before Change (ns)	After Change (ns)	Change in Time	Change in Speed
BM_Summary_Observe/0/iterations:262144	8074	1129	-6945	7.14x Faster
BM_Summary_Observe/1/iterations:262144	8032	1244	-6788	6.46x Faster
BM_Summary_Observe/8/iterations:262144	10122	8410	-1712	1.20x Faster
BM_Summary_Observe/64/iterations:262144	15786	17444	+1658	0.90x Slower
BM_Summary_Collect/0/0	41.8	42.7	+0.9	1.02x Slower
BM_Summary_Collect/1/0	96.2	110	+13.8	1.14x Slower
BM_Summary_Collect/8/0	338	429	+91	1.27x Slower
BM_Summary_Collect/64/0	2355	2788	+433	1.18x Slower
BM_Summary_Collect/0/1	44.1	42.4	-1.7	1.04x Faster
BM_Summary_Collect/1/1	101	111	+10	1.10x Slower
BM_Summary_Collect/8/1	496	521	+25	1.05x Slower
BM_Summary_Collect/64/1	8301	9060	+759	1.09x Slower
BM_Summary_Collect/0/8	44.9	35.5	-9.4	1.26x Faster
BM_Summary_Collect/1/8	149	137	-12	1.09x Faster
BM_Summary_Collect/8/8	1384	1787	+403	1.29x Slower
BM_Summary_Collect/64/8	45587	50081	+4494	1.10x Slower
BM_Summary_Collect/0/64	31.4	35.7	+4.3	1.14x Slower
BM_Summary_Collect/1/64	283	132	-151	2.14x Faster
BM_Summary_Collect/8/64	7899	9547	+648	1.21x Slower
BM_Summary_Collect/64/64	405506	420493	+14987	1.04x Slower
BM_Summary_Collect/0/512	34.5	33.6	-0.9	1.03x Faster
BM_Summary_Collect/1/512	1534	157	-1377	9.77x Faster
BM_Summary_Collect/8/512	39361	9249	-30112	4.25x Faster
BM_Summary_Collect/64/512	2029824	432225	-1597600	4.70x Faster
BM_Summary_Collect/0/4096	35.3	34.5	-0.8	1.02x Faster
BM_Summary_Collect/1/4096	3086	273	-2813	11.32x Faster
BM_Summary_Collect/8/4096	117444	116648	-796	1.01x Faster
BM_Summary_Collect/64/4096	5433854	5666187	+232333	1.04x Slower
BM_Summary_Collect/0/32768	33.2	39.3	+6.1	1.18x Slower
BM_Summary_Collect/1/32768	4012	2411	-1601	1.66x Faster
BM_Summary_Collect/8/32768	297063	511257	+214194	1.72x Slower
BM_Summary_Collect/64/32768	24117380	21697663	-2429702	1.11x Faster
BM_Summary_Collect/0/262144	40.7	34.9	-5.8	1.17x Faster
BM_Summary_Collect/1/262144	13560	11292	-2268	1.20x Faster
BM_Summary_Collect/8/262144	869739	1540760	+670021	1.77x Slower
BM_Summary_Collect/64/262144	36734543	100230906	+63596363	2.73x Slower
BM_Summary_Observe_Common/iterations:262144	4580	2480	-2100	1.85x Faster
BM_Summary_Collect_Common/0	196	310	+114	1.58x Slower
BM_Summary_Collect_Common/1	250	352	+102	1.41x Slower
BM_Summary_Collect_Common/8	579	899	+320	1.55x Slower
BM_Summary_Collect_Common/64	1810	816	-994	2.22x Faster
BM_Summary_Collect_Common/512	5108	1828	-3280	2.80x Faster
BM_Summary_Collect_Common/4096	32956	5892	-27064	5.60x Faster
BM_Summary_Collect_Common/32768	183060	28329	-154731	6.46x Faster
BM_Summary_Collect_Common/262144	159511	166397	+6868	1.04x Slower

The observations clearly indicate that for a limited number of quantiles (typically around 8), which represent the most common use case, there are substantial performance gains. However, for a larger number of quantiles (around 64), the performance improvements are either negligible or slightly diminished.