[VL] Set s.g.s.c.b.v.coalesceBatchesBeforeShuffle=true by default

[zhztheplayer]

Related to #6009

Set spark.gluten.sql.columnar.backend.velox.coalesceBatchesBeforeShuffle=true to make sure shuffle's performance is always not impacted by the issue.

The default batch size is changed to 0.8 * GLUTEN_MAX_BATCH_SIZE to add some flexibilities to avoid unexpected combinations from buffers that are already large enough.

[github-actions]

Thanks for opening a pull request!

Could you open an issue for this pull request on Github Issues?

https://github.com/apache/incubator-gluten/issues

Then could you also rename commit message and pull request title in the following format?

[GLUTEN-${ISSUES_ID}][COMPONENT]feat/fix: ${detailed message}

See also:

• Other pull requests

[github-actions]

Run Gluten Clickhouse CI

[github-actions]

Run Gluten Clickhouse CI

[github-actions]

Run Gluten Clickhouse CI

[zhztheplayer]

/Benchmark TPCDS

[zhztheplayer]

/Benchmark TPCDS

[github-actions]

Run Gluten Clickhouse CI

[zhztheplayer]

/Benchmark Velox TPCDS

[GlutenPerfBot]

===== Performance report for TPCDS SF2000 with Velox backend, for reference only ====

query	log/native_6056_time.csv	log/native_master_06_16_2024_a08a57c61f_time.csv	difference	percentage
q1	14.66	14.69	0.024	100.17%
q2	16.46	15.18	-1.281	92.22%
q3	4.33	4.91	0.575	113.27%
q4	63.86	66.82	2.963	104.64%
q5	6.62	8.33	1.709	125.80%
q6	3.45	2.16	-1.283	62.77%
q7	6.02	6.30	0.273	104.53%
q8	5.63	5.96	0.329	105.85%
q9	19.22	17.57	-1.654	91.40%
q10	11.33	11.03	-0.298	97.37%
q11	38.91	35.51	-3.394	91.28%
q12	1.50	1.44	-0.058	96.16%
q13	5.46	5.31	-0.145	97.35%
q14a	45.55	43.42	-2.136	95.31%
q14b	41.52	38.81	-2.713	93.47%
q15	2.71	3.85	1.145	142.30%
q16	39.45	42.00	2.559	106.49%
q17	4.86	4.73	-0.126	97.41%
q18	6.33	6.16	-0.165	97.39%
q19	2.20	3.59	1.399	163.72%
q20	1.47	2.71	1.242	184.47%
q21	1.03	1.35	0.320	131.01%
q22	10.91	8.13	-2.778	74.54%
q23a	80.98	82.95	1.969	102.43%
q23b	103.32	100.23	-3.089	97.01%
q24a	75.42	70.78	-4.636	93.85%
q24b	69.77	76.15	6.372	109.13%
q25	4.29	6.22	1.926	144.90%
q26	4.29	4.83	0.538	112.53%
q27	3.40	3.25	-0.149	95.63%
q28	20.92	21.89	0.975	104.66%
q29	8.29	7.91	-0.383	95.38%
q30	8.33	4.35	-3.980	52.22%
q31	6.45	6.50	0.054	100.84%
q32	1.31	1.11	-0.200	84.74%
q33	7.02	7.95	0.929	113.23%
q34	4.42	5.74	1.322	129.89%
q35	7.51	7.07	-0.433	94.23%
q36	3.82	3.56	-0.267	93.01%
q37	5.11	4.21	-0.894	82.51%
q38	11.75	12.83	1.073	109.13%
q39a	3.28	3.18	-0.099	96.98%
q39b	2.75	2.83	0.083	103.02%
q40	8.61	6.35	-2.259	73.75%
q41	0.70	0.61	-0.092	86.97%
q42	2.75	1.10	-1.648	40.03%
q43	3.87	3.90	0.022	100.58%
q44	10.42	13.59	3.170	130.42%
q45	5.63	4.16	-1.467	73.92%
q46	3.31	3.39	0.080	102.43%
q47	14.32	14.38	0.061	100.43%
q48	4.63	4.63	-0.001	99.99%
q49	9.50	9.65	0.152	101.60%
q50	27.59	24.71	-2.881	89.56%
q51	8.73	8.97	0.237	102.72%
q52	0.99	1.11	0.120	112.16%
q53	1.98	2.99	1.014	151.24%
q54	3.28	3.36	0.077	102.34%
q55	1.04	1.17	0.122	111.72%
q56	4.36	4.57	0.209	104.79%
q57	8.68	9.20	0.514	105.92%
q58	2.47	2.66	0.187	107.58%
q59	15.86	14.06	-1.803	88.64%
q60	4.68	4.64	-0.036	99.22%
q61	5.44	5.22	-0.218	95.99%
q62	4.41	6.60	2.195	149.81%
q63	2.23	2.10	-0.122	94.54%
q64	53.49	53.36	-0.129	99.76%
q65	13.83	16.93	3.096	122.38%
q66	3.45	3.16	-0.289	91.61%
q67	377.08	363.54	-13.536	96.41%
q68	3.54	3.72	0.181	105.11%
q69	7.48	9.96	2.480	133.16%
q70	8.95	8.93	-0.029	99.68%
q71	2.63	2.57	-0.059	97.74%
q72	188.52	192.42	3.896	102.07%
q73	2.20	2.26	0.052	102.38%
q74	21.44	21.15	-0.293	98.63%
q75	27.37	24.94	-2.434	91.11%
q76	10.33	12.49	2.165	120.96%
q77	2.22	2.33	0.113	105.07%
q78	43.34	42.10	-1.239	97.14%
q79	3.54	3.64	0.099	102.79%
q80	12.91	15.73	2.819	121.84%
q81	5.38	5.14	-0.235	95.63%
q82	7.71	7.76	0.051	100.67%
q83	1.48	2.36	0.879	159.51%
q84	2.71	2.82	0.110	104.06%
q85	7.91	8.11	0.199	102.52%
q86	3.41	3.26	-0.149	95.64%
q87	12.33	12.39	0.061	100.50%
q88	25.13	28.73	3.599	114.32%
q89	3.44	3.43	-0.006	99.82%
q90	4.29	4.35	0.057	101.33%
q91	2.62	2.62	-0.001	99.97%
q92	1.27	1.25	-0.017	98.66%
q93	32.35	30.88	-1.473	95.45%
q94	22.77	22.55	-0.218	99.04%
q9	89.00	87.33	-1.679	98.11%
q5	3.86	3.78	-0.079	97.94%
q96	12.29	11.99	-0.298	97.57%
q97	2.07	2.02	-0.048	97.69%
q98	8.51	13.03	4.522	153.14%
q99	8.51	13.03	4.522	153.14%
total	1972.25	1969.68	-2.575	99.87%

[GlutenPerfBot]

===== Performance report for TPCH SF2000 with Velox backend, for reference only ====

query	log/native_6056_time.csv	log/native_master_06_16_2024_a08a57c61_time.csv	difference	percentage
q1	37.54	37.13	-0.404	98.92%
q2	23.89	24.00	0.101	100.42%
q3	40.70	37.23	-3.469	91.48%
q4	32.45	32.75	0.298	100.92%
q5	73.14	71.56	-1.577	97.84%
q6	9.81	8.16	-1.647	83.20%
q7	80.72	80.98	0.265	100.33%
q8	85.24	86.78	1.538	101.80%
q9	118.38	121.57	3.190	102.69%
q10	45.07	44.17	-0.906	97.99%
q11	20.44	21.42	0.977	104.78%
q12	26.35	25.32	-1.029	96.09%
q13	38.73	39.84	1.106	102.86%
q14	19.73	19.23	-0.501	97.46%
q15	32.28	33.05	0.771	102.39%
q16	14.37	14.54	0.170	101.18%
q17	103.29	102.23	-1.057	98.98%
q18	149.08	142.93	-6.143	95.88%
q19	13.81	15.44	1.630	111.80%
q20	28.16	27.99	-0.170	99.40%
q21	262.81	261.41	-1.401	99.47%
q22	12.39	12.41	0.021	100.17%
total	1268.39	1260.15	-8.235	99.35%

[github-actions]

Run Gluten Clickhouse CI

[github-actions]

Run Gluten Clickhouse CI

[FelixYBW]

@zhztheplayer @marin-ma Similarly if the batch size is too large, each column exceeds the L2 cache size, the performance should be very bad as well. Can you submit a similar PR to fix this? Split the large batchs into the small one during Split.

[marin-ma]

@zhztheplayer @marin-ma Similarly if the batch size is too large, each column exceeds the L2 cache size, the performance should be very bad as well. Can you submit a similar PR to fix this? Split the large batchs into the small one during Split.

@FelixYBW Do we need a new Operator for it? If it's only for split, we already have it #5536

[zhztheplayer]

Perhaps we can tweak the code of VeloxAppendBatches including renaming it to make it be able to do both appending and slicing by giving it a target batch size range. So the happening of these two operations can be made more consistent.

[zzcclp]

Run Gluten Clickhouse CI

[github-actions]

Run Gluten Clickhouse CI

[marin-ma]

LGTM. Thanks!

[GlutenPerfBot]

===== Performance report for TPCH SF2000 with Velox backend, for reference only ====

query	log/native_6056_time.csv	log/native_master_06_17_2024_5b87efa56_time.csv	difference	percentage
q1	34.11	34.50	0.390	101.14%
q2	23.28	23.69	0.419	101.80%
q3	40.46	39.94	-0.515	98.73%
q4	30.62	32.62	2.004	106.54%
q5	67.97	67.62	-0.346	99.49%
q6	8.05	6.63	-1.416	82.41%
q7	81.43	81.08	-0.347	99.57%
q8	83.95	85.44	1.489	101.77%
q9	121.42	120.83	-0.588	99.52%
q10	48.86	46.34	-2.524	94.83%
q11	19.99	19.96	-0.029	99.85%
q12	29.07	25.57	-3.503	87.95%
q13	38.26	38.36	0.099	100.26%
q14	22.05	19.32	-2.726	87.64%
q15	30.34	32.79	2.446	108.06%
q16	14.20	13.93	-0.269	98.11%
q17	102.44	103.20	0.758	100.74%
q18	147.36	147.38	0.023	100.02%
q19	14.75	14.12	-0.632	95.71%
q20	29.34	28.78	-0.567	98.07%
q21	261.33	258.28	-3.044	98.84%
q22	12.22	12.23	0.010	100.09%
total	1261.49	1252.62	-8.870	99.30%