DATAS tuning changes for small HCs

[Maoni0]

changes included -

• currently we have a very simplistic formula for actually adapting to the size and this basically just makes all the asp.net benchmarks with low surv rate adjust to the min 2.5 mb gen0 budget, while those run ok with such a small budget on a 28 core machine, it doesn't work if we limit the heap count to a small number, eg, 4. what happens is the % time in GC is very high, some benchmarks run with 20% to 40% time in GC. this is obviously not desirable. I reworked this to make it actually adapting to the size. here's a chart for different gen2 sizes -

and we'll take the min of this and what we calculated without DATAS.

• the formula I had previously did not handle small HCs well so I also adjust that.

• got rid of the adjusting to cache size in gc1 for DATAS, this just makes things unpredictable especially for small workloads.

I will add data for the final build soon. there's more work to do that I will address in future PRs -

• do some refactoring as I've been adding a few tuning things that are of the same nature so it'd be nice to have a utility class to do that;
• make setting the min/max gen0 budget for DATAS part of the static data calculation.
• the gen0 budget computation should take conserve_mem_setting into consideration.
• need to get rid of more adjustments we do for gen0 budget that make things volatile.
• we do have this time tuning thing for WKS GC but not for SVR, we should bring that to DATAS to make sure we are not going for a very long time without doing a higher gen GC (especially a gen2 GC)).
• there are also a couple of things related to small HC tuning that I need to fix but I think this PR is at a good place.

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/gc
See info in area-owners.md if you want to be subscribed.

[Maoni0]

this obviously increases the memory footprint but increases RPS. this is a run with max HC set to 4 -

bench	b_max_mem	max_mem_%	b_p50_mem	p50_mem_%	b_rps	rps_%
Stage1GrpcServerGC	70.75	73.50	59.00	48.73	518.28	79.17
Stage1Grpc	69.00	60.51	59.50	48.32	524.22	77.56
FortunesPlatformEF	87.00	84.48	76.75	37.46	191.91	59.39
FortunesEf	58.00	93.97	55.50	40.99	137.16	40.50
FortunesDapper	47.50	52.63	43.50	50.00	139.56	38.75
PlaintextMvc	43.50	9.77	41.00	11.59	1671.11	25.77
Stage1TrimR2RSingleFile	29.50	51.69	28.25	49.56	599.50	25.19
Stage1	34.25	50.36	31.75	32.28	601.09	25.08
CachingPlatform	45.75	54.64	44.00	30.68	342.20	19.99
Stage2	62.00	38.31	59.00	24.58	168.18	18.03
ConnectionCloseHttpSys	25.25	76.24	23.50	48.94	81.11	15.66
SingleQueryPlatform	86.25	33.62	58.00	36.21	375.61	15.12
JsonMvc	44.25	32.20	42.75	25.73	457.95	12.69
PlaintextWithParametersNoFilter	44.00	21.59	43.00	17.44	2910.62	10.86
UpdatesPlatform	69.50	100.00	59.50	100.00	25.29	10.76
MultipleQueriesPlatform	63.75	100.78	61.25	93.88	40.96	7.65
PlaintextNoParametersEmptyFilter	35.25	28.37	33.75	15.56	3806.39	6.46
JsonHttps	51.25	43.41	47.00	15.43	653.70	-1.73

for example -

run	benchmark	gen0	pause	gen1	pause	ngc2	pause	bgc	pause	allocMB	alloc/gc	pct in GC
baseline_0	CachingPlatform	3282	1.7	8	3.07	1	5.49	2	1.87	9371.31	2.85	14.83
baseline_1	CachingPlatform	3539	1.9	5	3.73	2	4.6	1	1.61	8838.93	2.49	18.13
baseline_2	CachingPlatform	3572	1.84	6	3.58	1	5.51	2	1.82	8943.07	2.5	17.7
baseline_3	CachingPlatform	3629	1.76	6	3.49	1	5.44	2	1.89	9083.47	2.5	15.94
fix_0	CachingPlatform	1348	1.29	5	4.28	1	5.6	2	1.32	10911.95	8.05	4.66
fix_1	CachingPlatform	1283	1.23	5	4.21	1	5.83	2	1.39	10984.42	8.51	4.3
fix_2	CachingPlatform	1346	1.27	6	3.93	1	5.43	2	1.4	10842.49	8	4.38
fix_3	CachingPlatform	1397	1.22	6	3.92	1	5.34	2	1.93	10797	7.68	4.55

run	benchmark	gen0	pause	gen1	pause	ngc2	pause	bgc	pause	allocMB	alloc/gc	pct in GC
baseline_0	ConnectionCloseHttpSys	2463	1.55	1228	1.15	1	3.82	0	0	9289.57	2.52	17.33
baseline_1	ConnectionCloseHttpSys	2471	1.51	1230	1.11	1	4.13	0	0	9316.42	2.52	16.9
baseline_2	ConnectionCloseHttpSys	2481	1.52	1238	1.09	1	4.33	0	0	9364.03	2.52	15.92
baseline_3	ConnectionCloseHttpSys	2459	1.55	1228	1.15	1	4.43	0	0	9281.21	2.52	17.29
fix_0	ConnectionCloseHttpSys	580	1.6	309	1.21	1	3.08	0	0	10759.68	12.09	4.31
fix_1	ConnectionCloseHttpSys	557	1.58	297	1.21	0	0	1	0.49	10813.69	12.65	4.11
fix_2	ConnectionCloseHttpSys	562	1.59	287	1.2	0	0	1	0.45	10814.27	12.72	4.1
fix_3	ConnectionCloseHttpSys	556	1.64	286	1.21	0	0	1	0.68	10810.13	12.82	4.17

comparing with default SVR GC with max HC set to 4 -

without setting a small max HC, this does increase memory footprint a bit while the rps is mostly the same (+/- 5% from baseline). still much lower than default SVR GC -