We need elaborate profiling for ABACUS version 2.2.0, please collect as many case as you can!

[dyzheng]

Describe Current Status and Possible Solution

We can seperate the time consumed by the program into several parts collected in the end of screem or OUT.suffix/running_scf.txt outputs.

For example:

|CLASS_NAME---------|NAME---------------|TIME(Sec)-----|CALLS----|AVG------|PER%-------
Gint_k vlocal 46.585 10 4.7 33 %

We can focus on PER% > 1% and mark NAME, CALLS, AVG;
another significant feature is cores in this calculation.
Better to note what type of machine and note which compiler.

[dyzheng]

units: second

<style> </style>

oneapi+8269CY	scalapack对角化	vlocal格点积分矩阵元	nonlocal矩阵元	S矩阵元	sum_bands	cal_dk_k	cal_rho_k	mix_rho
128Si	2.4	4.7	7.6	0.67	2.9	0.36	2.3	1.1
256Si	22	9.5	17	1.5	7.9	1.9	4.9	1.2
512Si	2.00E+02	20	39	3.2	28	11	10	1.2
GNU+5600x	elpa对角化	vlocal格点积分矩阵元	nonlocal矩阵元	S矩阵元	sum_bands	cal_dk_k	cal_rho_k	mix_rho
128Si	1.6	22	3.4	0.37	11	0.25	10	0.79
256Si	14	46	7.7	23	23	1.3	21	0.7
512Si	1.10E+02	1.00E+02	17	1.9	59	6.9	47	0.74

[caic99]

To optimize bottleneck, I suggest we don't need so much samples. The bottlenecks should be consistent across samples, while the algorithm is static. A few profiling cases are sufficient to guide our optimize direction.

[pxlxingliang]

The format of time in running_scf.log is not scientific and only 2 significant digits are reserved, it is better to make it same with that in OUTPUT.

[dyzheng]

Some timer of function has error, we should keep repairing them.

[dyzheng]

To optimize bottleneck, I suggest we don't need so much samples. The bottlenecks should be consistent across samples, while the algorithm is static. A few profiling cases are sufficient to guide our optimize direction.

The new target is to elaborate profiling for several cases.

[caic99]

To optimize bottleneck, I suggest we don't need so much samples. The bottlenecks should be consistent across samples, while the algorithm is static. A few profiling cases are sufficient to guide our optimize direction.

The new target is to elaborate profiling for several cases.

That's good. Will you provide such cases? @dyzheng

[mohanchen]

The bottlenecks should be consistent across samples

Of course not, please learn some DFT knowledge when you have time.

[mohanchen]

Some timer of function has error, we should keep repairing them.

This should be another issue.

[mohanchen]

The format of time in running_scf.log is not scientific and only 2 significant digits are reserved, it is better to make it same with that in OUTPUT.

This is another issue.