benchmark: Compare performance on i8g and i7ie and i4i

[VoVAllen]

On laion 5m dataset
test pure cold read

[cutecutecat]

On Branch: 9b70fb7

data = laion-5m
nlist = 8192
shared buffer = 28G
jit = false
effective_io_concurrency = 200
top 10 with prewarm 
top 100 run twice

Build time

i8g > i7ie > i4i

i4i.xlarge: 1292.06s
i7ie.xlarge: 835.91s
i8g.xlarge: 691.53s

Benchmark compare

i7ie > i8g > i4i

nprob / epsilon / top	i7ie QPS	i4i QPS	i8g QPS	Recall
100 / 1.5 / 10	206.49	156.17	179.09	0.9803
100 / 1.0 / 10	227.53	167.73	197.68	0.9672
100 / 0.8 / 10	233.07	174.16	203.97	0.9543
100 / 0.6 / 10	238.77	174.97	209.97	0.9339
100 / 0.8 / 100	192.74	148.75	167.27	0.9513
100 / 0.6 / 100	200.05	149.97	173.69	0.9332
200 / 1.5 / 100	104.10	79.22	91.10	0.9847

[VoVAllen]

5m experiment on #157, i4i

[VoVAllen]

5m experiment on vchord-fscan, i4i @usamoi

[cutecutecat]

laion-5m

on #157

nprob / epsilon / top	QPS + Mem	QPS + Disk	Recall
100 / 1.5 / 10	128.89	41.40	0.9803
100 / 1.0 / 10	147.65	55.57	0.9672
100 / 0.8 / 10	160.18	63.95	0.9544
100 / 0.8 / 100	105.05	31.76	0.9513
100 / 0.6 / 100	111.26	33.91	0.9332
200 / 1.5 / 100	51.01	22.47	0.9847

on main

nprob / epsilon / top	QPS + Mem	QPS + Disk	Recall
100 / 1.5 / 10	159.13	42.71	0.9804
100 / 1.0 / 10	169.15	55.34	0.9672
100 / 0.8 / 10	180.41	61.86	0.9543
100 / 0.8 / 100	132.84	25.76	0.9513
100 / 0.6 / 100	142.71	26.69	0.9332
200 / 1.5 / 100	75.32	20.20	0.9847

on v0.1.0

nprob / epsilon / top	QPS + Mem	QPS + Disk	Recall
100 / 1.5 / 10	240.87	46.37	0.9803
100 / 1.0 / 10	258.12	61.22	0.9672
100 / 0.8 / 10	274.05	69.89	0.9543
100 / 0.8 / 100	193.82	26.94	0.9513
100 / 0.6 / 100	202.70	28.26	0.9332
200 / 1.5 / 100	102.47	21.80	0.9847

[cutecutecat]

Prewarm pages usage

# v0.1.0 branch

postgres=# SELECT vchordrq_prewarm('laion_embedding_idx'::regclass);
     vchordrq_prewarm      
---------------------------
 height of root: 2        +
 number of tuples: 1      +
 number of pages: 1       +
 number of tuples: 8192   +
 number of pages: 128     +
 number of tuples: 5000000+
 number of pages: 160277  +
 
# Main branch

postgres=# SELECT vchordrq_prewarm('laion_embedding_idx'::regclass);
     vchordrq_prewarm      
---------------------------
 height of root: 2        +
 number of tuples: 1      +
 number of pages: 1       +
 number of tuples: 8192   +
 number of pages: 142     +
 number of tuples: 5000000+
 number of pages: 90278   +

# PR 157 branch

postgres=# SELECT vchordrq_prewarm('laion_embedding_idx'::regclass);
     vchordrq_prewarm      
---------------------------
 height of root: 2        +
 number of tuples: 1      +
 number of pages: 1       +
 number of tuples: 8192   +
 number of pages: 142     +
 number of tuples: 5000000+
 number of pages: 86055   +

[usamoi]

i4i.xlarge, e26e5e3, vchordrqfscan, lists = 8192, shared_buffer = 28G, jit = off, effective_io_concurrency = 200

probes / epsilon / top	QPS	Recall
100 / 1.5 / 10	174.96	0.9717
100 / 1.0 / 10	187.15	0.9597
100 / 0.8 / 10	191.62	0.9476
100 / 0.8 / 100	94.72	0.9492
100 / 0.6 / 100	151.84	0.9318
200 / 1.5 / 100	76.07	0.9805

   vchordrqfscan_prewarm   
---------------------------
 height of root: 2        +
 number of tuples: 1      +
 number of pages: 1       +
 number of tuples: 8192   +
 number of pages: 128     +
 number of tuples: 5000000+
 number of pages: 160277  +

i4i.xlarge, e26e5e3, vchordrq, lists = 8192, shared_buffer = 28G, jit = off, effective_io_concurrency = 200

probes / epsilon / top	QPS	Recall
100 / 1.5 / 10	139.10	0.9717
100 / 1.0 / 10	148.04	0.9596
100 / 0.8 / 10	150.17	0.9476
100 / 0.8 / 100	120.18	0.9493
100 / 0.6 / 100	125.12	0.9319
200 / 1.5 / 100	63.29	0.9806

     vchordrq_prewarm      
---------------------------
 height of root: 2        +
 number of tuples: 1      +
 number of pages: 1       +
 number of tuples: 8192   +
 number of pages: 142     +
 number of tuples: 5000000+
 number of pages: 90278   +

i4i.xlarge, db1c2ed, vchordrq, lists = 8192, shared_buffer = 26G, jit = off, effective_io_concurrency = 200

probes / epsilon / top	QPS	Recall
100 / 1.5 / 10	186.43	0.9716
100 / 1.0 / 10	197.50	0.9595
100 / 0.8 / 10	205.65	0.9475
100 / 0.8 / 100	157.91	0.9492
100 / 0.6 / 100	163.79	0.9319
200 / 1.5 / 100	63.92	0.9806

     vchordrq_prewarm      
---------------------------
 height of root: 2        +
 number of tuples: 1      +
 number of pages: 1       +
 number of tuples: 8192   +
 number of pages: 128     +
 number of tuples: 5000000+
 number of pages: 160277  +