Use std::sort instead of QSortInt

[fuji8]

I profiled the T1050 with the parameters run by AlphaFold. (Only -cpu is changed). I used Score-P as the profiling tool and got the following results.

From this image, we can see that QSortInt in mergeHitsToQuery is taking a long time. With fast enough storage, my hhblits for this condition is about 2100sec, and QSortInt accounts for about 40%.

Instead of this QsortInt, use std::sort.

I ran hhblits installed by conda and using std::sort under the same conditions as before.
In order to avoid I/O effects, I analyze the difference in execution time between the logs that contain this change, instead of the overall execution time. (From

hh-suite/src/hhblits.cpp

Lines 1028 to 1030 in ac76598

	HH_LOG(INFO)
	<< "Realigning " << nhits
	<< " HMM-HMM alignments using Maximum Accuracy algorithm" << std::endl;

to

hh-suite/src/hhhmm.cpp

Line 2337 in b100bb0

HH_LOG(INFO) << "Neutralized His-tag between positions " << imax(i0 - 8, 1) << " and " << i-1 << std::endl;

)

	conda	use std::sort
iteration 1	1232(sec)	477(sec)
iteration 2	631(sec)	374(sec)
iteration 3	306(sec)	232(sec)

This reduced the execution time. I also ran it using the parallelization policy, but the results were not significantly different from std::sort.

This change is due to the different stability of sort, so the execution results may not truly match.

[milot-mirdita]

Cool, thank you!

We have implemented a similar fix in MMseqs2's version of the same code, but haven’t backported it:
https://github.com/soedinglab/MMseqs2/blob/d89fcecf9911a99c45ed81c1c0e5054743debc64/src/alignment/MsaFilter.cpp#L212

Could you repeat the benchmark with a stable sort?

[fuji8]

Thank you for the reply.

I changed the sort to stable_sort and ran it 3 times on hpc in the following environment.

• cpu: 28 cores
• RAM: 235GB

	1	2	3
iteration 1	488(sec)	484(sec)	484(sec)
iteration 2	374(sec)	371(sec)	376(sec)
iteration 3	223(sec)	225(sec)	218(sec)

Because of the large memory, the computational complexity is probably Nlog(N).

[martin-steinegger]

@fuji8 This looks great! Thank you for the PR. Would it be possible to avoid the lambda expression in the sort?

[fuji8]

I apologize for the delay in response.

I rewrote the code to be almost equivalent without using the lambda expression.
I ran it only once, just to be sure.

	no lambda
iteration 1	500(sec)
iteration 2	385(sec)
iteration 3	232(sec)