https://bugs.openldap.org/show_bug.cgi?id=9316

Issue ID: 9316 Summary: performance issue when writing a high number of large objects Product: LMDB Version: 0.9.24 Hardware: x86_64 OS: Linux Status: UNCONFIRMED Severity: normal Priority: --- Component: liblmdb Assignee: bugs@openldap.org Reporter: JGabler@univa.com Target Milestone: ---

Created attachment 755 --> https://bugs.openldap.org/attachment.cgi?id=755&action=edit lmdb performance test reproducing the issue

When writing a high number of big objects we see an extreme variation in performance from very fast to extremely slow.

In the test scenario we write 10 chunks of 10.000 "jobs" (some 10kB) and their corresponding "job script" (some 40kB), 200.000 objects in total. Then delete all objects.

We do 10 iterations of this scenario.

When running this scenario as part of Univa Grid Engine with LMDB as database backend we get the following performance values (rows are the iteration, columns the chunk of jobs):

Iteration 0 1 2 3 4 5 6 7 8 9 0 21.525 21.250 21.574 21.722 22.693 21.992 22.438 22.650 21.972 22.017 1 22.262 21.656 22.339 22.914 21.549 24.906 23.862 1531.189 1695.041 1491.255 2 36.071 21.619 22.074 22.927 23.455 27.239 22.640 22.802 633.956 1882.008 3 52.163 21.651 21.571 22.686 22.727 22.024 40.980 22.156 22.429 595.362 4 64.977 21.511 22.519 22.148 22.354 23.292 57.740 20.835 37.680 250.594 5 54.724 21.074 21.200 23.744 22.109 21.351 62.225 21.447 91.292 375.260 6 49.065 21.573 22.309 26.084 21.226 21.248 68.580 22.531 59.338 249.936 7 44.666 21.830 21.009 28.760 21.533 21.611 72.291 23.144 86.281 118.326 8 35.486 21.720 21.840 24.729 22.045 20.877 76.473 21.193 120.387 136.836 9 41.159 23.365 21.721 23.024 21.835 20.972 77.409 21.784 193.885 306.158

So usually writing of 10.000 "jobs"+"job_script" takes some 22 seconds but after some time performance massively breaks in.

With other database backends we do not see this behaviour, see the following performance data of the same test done with PostgreSQL backend which is slower (as expected going over the network) but provides constant throughput:

Iteration 0 1 2 3 4 5 6 7 8 9 0 36.937 37.110 36.952 37.279 37.580 37.364 37.950 37.390 37.682 37.439 1 37.464 38.110 37.679 38.366 37.576 37.624 37.476 37.412 37.265 37.727 2 36.394 37.635 37.347 37.603 37.402 37.515 37.802 37.898 37.355 37.939 3 37.213 37.539 36.771 37.706 37.055 37.780 37.283 37.488 36.955 37.460 4 36.554 37.557 37.368 37.960 37.070 37.892 37.459 37.857 37.228 37.833 5 37.047 38.164 37.167 37.885 37.268 37.676 37.355 37.572 37.347 37.569 6 37.118 37.735 36.857 37.602 36.717 37.716 37.444 37.685 37.085 38.151 7 36.787 37.647 36.844 37.601 36.934 37.440 37.632 37.291 37.174 37.926 8 36.884 37.560 37.117 37.239 37.034 37.748 37.289 37.635 36.822 37.693 9 37.178 37.496 36.849 37.799 37.289 37.644 37.461 37.622 37.022 37.670

We can reproduce the issue with a small C program (see attachment) which does essentially the same database operations as our database layer in Univa Grid Engine but depends only on liblmdb. It simulates the scenario described above and gives us the following performance data showing the extreme performance variation:

Iteration 0 1 2 3 4 5 6 7 8 9 0 0.686 0.625 0.660 0.637 0.631 0.741 0.757 0.658 0.651 0.614 1 0.705 0.838 0.690 0.772 0.663 3.248 0.605 542.762 1114.374 898.477 2 13.336 1.299 0.659 0.637 0.626 0.712 11.172 0.663 29.833 1161.884 3 26.774 0.647 0.607 0.586 0.583 0.639 24.893 0.629 3.837 423.248 4 32.802 0.629 0.616 0.560 0.550 0.605 31.133 0.625 6.606 195.150 5 34.819 0.623 0.628 0.582 0.564 0.609 32.275 0.607 7.599 134.106 6 26.319 0.622 0.582 0.548 0.551 0.590 28.536 0.611 36.429 160.781 7 21.878 0.814 0.668 0.736 0.614 0.543 24.355 0.626 36.583 148.337 8 4.129 0.654 5.674 0.596 0.566 0.554 7.158 0.633 0.599 48.799 9 30.278 0.608 0.608 0.560 0.549 0.587 29.253 0.606 9.593 128.339

It can be compiled on Linux 64bit with gcc -I <path to lmdb>/include -L <path to lmdb>/lib -o test_lmdb_perf test_lmdb_perf.c -llmdb

To run the given scenario call it with the following parameters: ./test_lmdb_perf <path to database directory> 10 10 10000

We built and ran it on - CentOS Linux release 7.7.1908 (Core) - Linux biber 3.10.0-1062.9.1.el7.x86_64 #1 SMP Fri Dec 6 15:49:49 UTC 2019 x86_64 x86_64 x86_64 GNU/Linux - it was built with gcc (GCC) 7.2.1 20170829 (Red Hat 7.2.1-1) from devtoolset-7