Wednesday, May 12, 2021

Dynamic Tracing of Memory Allocations in MySQL With bcc Tools

Last year I started my lame attempts to apply different Linux dynamic tracing tools and approaches to frequent events like memory allocations. In this blog post I already described how to use perf to add user probe to trace malloc() calls with the number of bytes requested. Unfortunately this approach is not practical for production use for more than several seconds.

Recently I plan with bpftrace a lot and so far ended up with an easy way to trace calls and call stacks, and was on my way to trace only outstanding allocations and care only about memory areas not yet freed. If you are interested, the primitive approach may look like this:

[openxs@fc33 ~]$ cat malloc.bt
#!/usr/bin/env bpftrace

BEGIN
{
    printf("Tracing MariaDB's malloc() calls, Ctrl-C to stop\n");
}

interval:s:$1 { exit(); }

uprobe:/lib64/libc.so.6:malloc
/ comm == "mariadbd" /
{
    @size[tid] += arg0;
/*    printf("Allocating %d bytes in thread %u...\n", arg0, tid);  */
}

uretprobe:/lib64/libc.so.6:malloc
/ comm == "mariadbd" && @size[tid] > 0 /
{
    @memory[tid,retval] = @size[tid];
    @stack[ustack(perf)] += @size[tid];

    print(@stack);
    clear(@stack);

    delete(@size[tid]);
}

uprobe:/lib64/libc.so.6:free
/ comm == "mariadbd" /
{
    delete(@memory[tid, arg0]);
/*    printf("Freeing %p...\n", arg0); */
}

END
{
    clear(@size);
    clear(@memory);
    clear(@stack);
}
[openxs@fc33 ~]$

But while it works (both for tracing that is commented out above and for summarizing the non-freed allocations) and produced some outputs as expected:

[openxs@fc33 ~]$ time sudo ./malloc.bt 1 2>/dev/null >/tmp/malloc_raw.txt

real    8m47.963s
user    2m53.513s
sys     5m50.685s

[openxs@fc33 maria10.5]$ ls -l /tmp/malloc_raw.txt
-rw-r--r--. 1 openxs openxs 461675 Apr 22 10:13 /tmp/malloc_raw.txt
[openxs@fc33 ~]$ tail /tmp/malloc_raw.txt
        558c68a27d12 row_purge_step(que_thr_t*)+770 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c689e8256 que_run_threads(que_thr_t*)+2166 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68a47ab7 purge_worker_callback(void*)+375 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68b88989 tpool::task_group::execute(tpool::task*)+137 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68b87bdf tpool::thread_pool_generic::worker_main(tpool::worker_data*)+79 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        7f599ae455f4 0x7f599ae455f4 ([unknown])
        558c69fc26b0 0x558c69fc26b0 ([unknown])
        558c68b87cb0 std::thread::_State_impl<std::thread::_Invoker<std::tuple<void (tpool::thread_pool_generic::*)(tpool::worker_data*), tpool::thread_pool_generic*, tpool::worker_data*> > >::~_State_impl()+0 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        dde907894810c083 0xdde907894810c083 ([unknown])
]: 33

...

@stack[
        558c689bd25c mem_heap_create_block_func(mem_block_info_t*, unsigned long, unsigned long)+108 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68a3f2ae row_vers_old_has_index_entry(bool, unsigned char const*, mtr_t*, dict_index_t*, dtuple_t const*, unsigned long, unsigned long)+126 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68a24c60 row_purge_poss_sec(purge_node_t*, dict_index_t*, dtuple_t const*, btr_pcur_t*, mtr_t*, bool)+512 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68a2627b row_purge_remove_sec_if_poss_leaf(purge_node_t*, dict_index_t*, dtuple_t const*)+971 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68a272c3 row_purge_record_func(purge_node_t*, unsigned char*, que_thr_t const*, bool)+1459 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68a27d12 row_purge_step(que_thr_t*)+770 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c689e8256 que_run_threads(que_thr_t*)+2166 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68a47ab7 purge_worker_callback(void*)+375 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68b88989 tpool::task_group::execute(tpool::task*)+137 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        558c68b87bdf tpool::thread_pool_generic::worker_main(tpool::worker_data*)+79 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        7f599ae455f4 0x7f599ae455f4 ([unknown])
        558c69fc26b0 0x558c69fc26b0 ([unknown])
        558c68b87cb0 std::thread::_State_impl<std::thread::_Invoker<std::tuple<void (tpool::thread_pool_generic::*)(tpool::worker_data*), tpool::thread_pool_generic*, tpool::worker_data*> > >::~_State_impl()+0 (/home/openxs/dbs/maria10.5/bin/mariadbd)
        dde907894810c083 0xdde907894810c083 ([unknown])
]: 1152

it took 8 minutes(!) to deal with data collected for over 1 second of tracing under high load, and caused notable load:

top - 09:59:04 up  1:16,  3 users,  load average: 1.60, 7.14, 5.31
Tasks: 228 total,   2 running, 226 sleeping,   0 stopped,   0 zombie
%Cpu(s): 10.8 us, 17.5 sy,  0.0 ni, 71.6 id,  0.0 wa,  0.2 hi,  0.0 si,  0.0 st
MiB Mem :   7916.5 total,   1759.6 free,   2527.0 used,   3629.9 buff/cache
MiB Swap:   3958.0 total,   3958.0 free,      0.0 used.   4910.8 avail Mem

    PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND
  14813 root      20   0  307872 111852  75516 R  99.7   1.4   7:57.17 bpftrace
...

and drop in performance for the system in the meantime:

[openxs@fc33 maria10.5]$ sysbench oltp_read_write --db-driver=mysql --tables=5 --table-size=100000 --mysql-user=openxs --mysql-socket=/tmp/mariadb.sock --mysql-db=sbtest --threads=32 --report-interval=10 --time=300 run
sysbench 1.1.0-174f3aa (using bundled LuaJIT 2.1.0-beta2)

Running the test with following options:
Number of threads: 32
Report intermediate results every 10 second(s)
Initializing random number generator from current time


Initializing worker threads...

Threads started!

[ 10s ] thds: 32 tps: 759.62 qps: 15246.37 (r/w/o: 10676.06/3047.88/1522.44) lat (ms,95%): 80.03 err/s: 0.00 reconn/s: 0.00
...
[ 70s ] thds: 32 tps: 708.20 qps: 14174.96 (r/w/o: 9920.64/2837.91/1416.41) lat (ms,95%): 74.46 err/s: 0.00 reconn/s: 0.00
[ 80s ] thds: 32 tps: 354.60 qps: 7080.28 (r/w/o: 4964.46/1406.62/709.21) lat (ms,95%): 134.90 err/s: 0.00 reconn/s: 0.00
[ 90s ] thds: 32 tps: 332.91 qps: 6661.34 (r/w/o: 4657.50/1338.03/665.81) lat (ms,95%): 132.49 err/s: 0.00 reconn/s: 0.00

...

So I mention that my lame tracing approach here for no one to try to do the same - monitoring that caused 2 times drop in TPS for minutes is hardly acceptable. I obviously made some mistake that I have yet to identify. Probably resolving stack traces and summarizing them in kernel context was too much for the system, and I can not do better in bpftrace itself, if only use it to just produce raw traces.

The approach above is too primitive also, as I traced only malloc(), while theoretically realloc() and callock() calls may be used. So in a hope to see how this task is approached by really experienced people, I checked what bcc tools provide for tracing memory allocations.

The memleak.py tool there is quite advanced. It allows both to trace individual calls and output periodic summaries of outstanding allocations:

openxs@ao756:~/git/BPF-tools$ /usr/share/bcc/tools/memleak -h
usage: memleak [-h] [-p PID] [-t] [-a] [-o OLDER] [-c COMMAND]
               [--combined-only] [--wa-missing-free] [-s SAMPLE_RATE] [-T TOP]
               [-z MIN_SIZE] [-Z MAX_SIZE] [-O OBJ] [--percpu]
               [interval] [count]

Trace outstanding memory allocations that weren't freed.
Supports both user-mode allocations made with libc functions and kernel-mode
allocations made with kmalloc/kmem_cache_alloc/get_free_pages and corresponding
memory release functions.

positional arguments:
  interval              interval in seconds to print outstanding allocations
  count                 number of times to print the report before exiting

optional arguments:
  -h, --help            show this help message and exit
  -p PID, --pid PID     the PID to trace; if not specified, trace kernel
                        allocs
  -t, --trace           print trace messages for each alloc/free call
  -a, --show-allocs     show allocation addresses and sizes as well as call
                        stacks
  -o OLDER, --older OLDER
                        prune allocations younger than this age in
                        milliseconds
  -c COMMAND, --command COMMAND
                        execute and trace the specified command
  --combined-only       show combined allocation statistics only
  --wa-missing-free     Workaround to alleviate misjudgments when free is
                        missing
  -s SAMPLE_RATE, --sample-rate SAMPLE_RATE
                        sample every N-th allocation to decrease the overhead
  -T TOP, --top TOP     display only this many top allocating stacks (by size)
  -z MIN_SIZE, --min-size MIN_SIZE
                        capture only allocations larger than this size
  -Z MAX_SIZE, --max-size MAX_SIZE
                        capture only allocations smaller than this size
  -O OBJ, --obj OBJ     attach to allocator functions in the specified object
  --percpu              trace percpu allocations

EXAMPLES:

./memleak -p $(pidof allocs)
        Trace allocations and display a summary of "leaked" (outstanding)
        allocations every 5 seconds
./memleak -p $(pidof allocs) -t
        Trace allocations and display each individual allocator function call
./memleak -ap $(pidof allocs) 10
        Trace allocations and display allocated addresses, sizes, and stacks
        every 10 seconds for outstanding allocations
./memleak -c "./allocs"
        Run the specified command and trace its allocations
./memleak
        Trace allocations in kernel mode and display a summary of outstanding
        allocations every 5 seconds
./memleak -o 60000
        Trace allocations in kernel mode and display a summary of outstanding
        allocations that are at least one minute (60 seconds) old
./memleak -s 5
        Trace roughly every 5th allocation, to reduce overhead
openxs@ao756:~/git/BPF-tools$

I've applied it to MySQL 8.0.25 recently built on my Ubuntu 20.04 and running sysbench oltp_read_write load test:

openxs@ao756:~/dbs/8.0$ sysbench oltp_read_write --db-driver=mysql --tables=5 --table-size=100000 --mysql-user=root --mysql-socket=/tmp/mysql8.sock --mysql-db=sbtest --time=300 --report-interval=10 --threads=4 run
sysbench 1.0.18 (using system LuaJIT 2.1.0-beta3)

Running the test with following options:
Number of threads: 4
Report intermediate results every 10 second(s)
Initializing random number generator from current time


Initializing worker threads...

Threads started!

[ 10s ] thds: 4 tps: 18.49 qps: 377.44 (r/w/o: 264.49/75.57/37.38) lat (ms,95%): 344.08 err/s: 0.00 reconn/s: 0.00
[ 20s ] thds: 4 tps: 14.80 qps: 296.01 (r/w/o: 207.21/59.20/29.60) lat (ms,95%): 530.08 err/s: 0.00 reconn/s: 0.00
[ 30s ] thds: 4 tps: 25.99 qps: 519.89 (r/w/o: 363.92/103.98/51.99) lat (ms,95%): 314.45 err/s: 0.00 reconn/s: 0.00
[ 40s ] thds: 4 tps: 25.30 qps: 506.04 (r/w/o: 354.23/101.21/50.60) lat (ms,95%): 292.60 err/s: 0.00 reconn/s: 0.00
[ 50s ] thds: 4 tps: 21.90 qps: 437.92 (r/w/o: 306.54/87.58/43.79) lat (ms,95%): 356.70 err/s: 0.00 reconn/s: 0.00
[ 60s ] thds: 4 tps: 23.51 qps: 470.05 (r/w/o: 329.10/93.93/47.01) lat (ms,95%): 308.84 err/s: 0.00 reconn/s: 0.00
[ 70s ] thds: 4 tps: 20.29 qps: 405.99 (r/w/o: 284.12/81.28/40.59) lat (ms,95%): 450.77 err/s: 0.00 reconn/s: 0.00
[ 80s ] thds: 4 tps: 20.51 qps: 408.20 (r/w/o: 286.17/81.02/41.01) lat (ms,95%): 390.30 err/s: 0.00 reconn/s: 0.00

[ 90s ] thds: 4 tps: 22.80 qps: 457.80 (r/w/o: 320.03/92.18/45.59) lat (ms,95%): 383.33 err/s: 0.00 reconn/s: 0.00
^C

Ignore ther absolute numbers, but note that (unlike with my bpftrace program) there was no very significant drop in QPS over that 20+ seconds I was collecting stacks for outstanding allocations in another shell:

openxs@ao756:~/git/BPF-tools$ time sudo /usr/share/bcc/tools/memleak -p $(pidof mysqld) --top 100 >/tmp/memleak.out
[sudo] password for openxs:
^C
real    0m21,142s
user    0m0,998s
sys     0m0,466s

Now, what was collected? Let' check top 40 rows:

openxs@ao756:~/git/BPF-tools$ head -40 /tmp/memleak.out
Attaching to pid 3416, Ctrl+C to quit.
[13:51:26] Top 100 stacks with outstanding allocations:
        1536 bytes in 2 allocations from stack
                mem_heap_create_block_func(mem_block_info_t*, unsigned long, unsigned long)+0xc8 [mysqld]
                mem_heap_add_block(mem_block_info_t*, unsigned long)+0x53 [mysqld]
                row_vers_build_for_consistent_read(unsigned char const*, mtr_t*, dict_index_t*, unsigned long**, ReadView*, mem_block_info_t**, mem_block_info_t*, unsigned char**, dtuple_t const**, lob::undo_vers_t*)+0x78c [mysqld]
                row_search_mvcc(unsigned char*, page_cur_mode_t, row_prebuilt_t*, unsigned long, unsigned long)+0x2b78 [mysqld]
                ha_innobase::index_read(unsigned char*, unsigned char const*, unsigned int, ha_rkey_function)+0x32e [mysqld]
                handler::ha_index_read_map(unsigned char*, unsigned char const*, unsigned long, ha_rkey_function)+0x389 [mysqld]
                handler::read_range_first(key_range const*, key_range const*, bool, bool)+0x6e [mysqld]
                ha_innobase::read_range_first(key_range const*, key_range const*, bool, bool)+0x27 [mysqld]
                handler::multi_range_read_next(char**)+0x135 [mysqld]
                handler::ha_multi_range_read_next(char**)+0x2c [mysqld]
                QUICK_RANGE_SELECT::get_next()+0x5a [mysqld]
                IndexRangeScanIterator::Read()+0x3f [mysqld]
                FilterIterator::Read()+0x18 [mysqld]
                MaterializeIterator::MaterializeQueryBlock(MaterializeIterator::QueryBlock const&, unsigned long long*)+0x133 [mysqld]
                MaterializeIterator::Init()+0x319 [mysqld]
                filesort(THD*, Filesort*, RowIterator*, unsigned long, unsigned long long, Filesort_info*, Sort_result*, unsigned long long*)+0x39d [mysqld]
                SortingIterator::DoSort()+0x72 [mysqld]
                SortingIterator::Init()+0x34 [mysqld]
                Query_expression::ExecuteIteratorQuery(THD*)+0x2ea [mysqld]
                Query_expression::execute(THD*)+0x33 [mysqld]
                Sql_cmd_dml::execute_inner(THD*)+0x30b [mysqld]
                Sql_cmd_dml::execute(THD*)+0x545 [mysqld]
                mysql_execute_command(THD*, bool)+0x9f0 [mysqld]
                Prepared_statement::execute(String*, bool)+0x8b0 [mysqld]
                Prepared_statement::execute_loop(String*, bool)+0x117 [mysqld]
                mysqld_stmt_execute(THD*, Prepared_statement*, bool, unsigned long, PS_PARAM*)+0x1b1 [mysqld]
                dispatch_command(THD*, COM_DATA const*, enum_server_command)+0x175d [mysqld]
                do_command(THD*)+0x1a4 [mysqld]
                handle_connection+0x258 [mysqld]
                pfs_spawn_thread+0x162 [mysqld]
                start_thread+0xd9 [libpthread-2.31.so]
        2304 bytes in 3 allocations from stack
                mem_heap_create_block_func(mem_block_info_t*, unsigned long, unsigned long)+0xc8 [mysqld]
                mem_heap_add_block(mem_block_info_t*, unsigned long)+0x53 [mysqld]
                row_vers_build_for_consistent_read(unsigned char const*, mtr_t*, dict_index_t*, unsigned long**, ReadView*, mem_block_info_t**, mem_block_info_t*, unsigned char**, dtuple_t const**, lob::undo_vers_t*)+0x78c [mysqld]
                row_search_mvcc(unsigned char*, page_cur_mode_t, row_prebuilt_t*, unsigned long, unsigned long)+0x2b78 [mysqld]
                ha_innobase::index_read(unsigned char*, unsigned char const*, unsigned int, ha_rkey_function)+0x32e [mysqld]
openxs@ao756:~/git/BPF-tools$

This information may be really useful for further analysis. But it is not easy to collapse/summarize for further memory flame graphs creation, while I'd really use them to quickly see where most of the memory is allocated.

One day I'll get back to memleak and create some script to fold its output. But for this blog post I was looking for qucik and dirty ways and, according to the page linked above, for that I had either to use general purpose stackcount.py (that just counts the number of occurences per unique stack), or Brendan Gregg's unsupported mallocstacks, which is similar to stackcount but sums the size_t argument to malloc() as the metric. I've used the latter and had to make a small fix to make it run on my Ubuntu 20.04 netbook:

openxs@ao756:~/git/BPF-tools$ git diff old/2017-12-23/mallocstacks.py
diff --git a/old/2017-12-23/mallocstacks.py b/old/2017-12-23/mallocstacks.py
index 8891e82..92271ed 100755
--- a/old/2017-12-23/mallocstacks.py
+++ b/old/2017-12-23/mallocstacks.py
@@ -96,7 +96,7 @@ struct key_t {
     char name[TASK_COMM_LEN];
 };
 BPF_HASH(bytes, struct key_t);
-BPF_STACK_TRACE(stack_traces, STACK_STORAGE_SIZE)
+BPF_STACK_TRACE(stack_traces, STACK_STORAGE_SIZE);

 int trace_malloc(struct pt_regs *ctx, size_t size) {
     u32 pid = bpf_get_current_pid_tgid();
openxs@ao756:~/git/BPF-tools$

Then I've used it as follows, run for some 10+ seconds against the same MySQL 8.0.25 under the same load (and with similar acceptable impact as with memleak):

openxs@ao756:~/git/BPF-tools/old/2017-12-23$ sudo ~/git/BPF-tools/old/2017-12-23/mallocstacks.py -p $(pidof mysqld) -f >/tmp/alloc.out
^C
openxs@ao756:~/git/BPF-tools/old/2017-12-23$ ls -l /tmp/alloc.out
-rw-rw-r-- 1 openxs openxs 859059 тра 12 10:28 /tmp/alloc.out
openxs@ao756:~/git/BPF-tools$ head -2 /tmp/alloc.out
mysqld;[unknown];std::thread::_State_impl<std::thread::_Invoker<std::tuple<Runnable, void (*)()> > >::~_State_impl();dict_stats_thread();[unknown];std::thread::_State_impl<std::thread::_Invoker<std::tuple<Runnable, void (*)()> > >::_M_run();dict_stats_thread();dict_stats_update(dict_table_t*, dict_stats_upd_option_t);dict_stats_save(dict_table_t*, index_id_t const*);dict_stats_exec_sql(pars_info_t*, char const*, trx_t*);que_eval_sql(pars_info_t*, char const*, unsigned long, trx_t*);que_run_threads(que_thr_t*);row_sel_step(que_thr_t*);row_sel(sel_node_t*, que_thr_t*);eval_cmp(func_node_t*);eval_node_alloc_val_buf(void*, unsigned long);__libc_malloc 33
mysqld;[unknown];std::thread::_State_impl<std::thread::_Invoker<std::tuple<Runnable, void (*)()> > >::~_State_impl();dict_stats_thread();[unknown];std::thread::_State_impl<std::thread::_Invoker<std::tuple<Runnable, void (*)()> > >::_M_run();dict_stats_thread();dict_stats_update(dict_table_t*, dict_stats_upd_option_t);dict_stats_save(dict_table_t*, index_id_t const*);dict_stats_exec_sql(pars_info_t*, char const*, trx_t*);trx_commit_for_mysql(trx_t*);trx_commit(trx_t*);trx_commit_low(trx_t*, mtr_t*);__libc_malloc 40
openxs@ao756:~/git/BPF-tools/old/2017-12-23$ cd
openxs@ao756:~$ cat /tmp/alloc.out | ~/git/FlameGraph/flamegraph.pl --color=mem --title="malloc() Flame Graph" --countname="bytes" >/tmp/mysql8_malloc.svg

So, we've got nice long folded stacks with the amount of bytes allocated (no matter if freed or not already), pure (as low impact as possible) malloc() tracing in a way immediately usable by flamegraph.pl, that produced the following output:

Flame graph showing where most of the memory was allocated from in MySQL 8.0.25 over 10+ seconds, while running standard sysbench oltp_read_write test.

Primitive, but it worked and allowed to see that most of allocations were related to filesort.

* * *

To summarize:
  • Relatively low impact tracing of memory allocations (ongoing and outstanding) is possible, using bcc tools.
  • One may use memleak tool to get quick insights on outstanding memory allocations, no matter where and how they were made, with period sampling to reduce the performance impact, or rely on some custom or general tracing tools to collect some metric per stack trace and represent the result as memory flame graphs.
  • Looks like it may make sense to do just primitive tracing with bpftrace and not try to overload the tool wiuth collecting per stack data in the maps, as stacks resolution seem to take too much CPU resources.

Sunday, May 9, 2021

Off-CPU Analysis Attempt to Find the Reason of Performance Regression in MariaDB 10.4

I did not write new blog posts for more than 2 months already. Busy days... But now I am on vacation and my Percona Live Online 2021 talk on flame graphs is coming soon, so I decided to renew my experience with bcc tools and try to get some insights for one of the bugs I've reported for MariaDB using the off-CPU flame graphs.

The idea was to check why sysbench oltp_read_write test started to work notably slower in a newer version of MariaDB 10.4.x after 10.4.15. On my good, old and slow Acer netbook with recently updated Ubuntu: 

openxs@ao756:~/git/server$ cat /etc/lsb-release
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=20.04
DISTRIB_CODENAME=focal
DISTRIB_DESCRIPTION="Ubuntu 20.04.2 LTS"
openxs@ao756:~/git/server$ uname -a
Linux ao756 5.8.0-50-generic #56~20.04.1-Ubuntu SMP Mon Apr 12 21:46:35 UTC 2021 x86_64 x86_64 x86_64 GNU/Linux

I've compiled current MariaDB 10.4 from GitHub source following my usual way:

openxs@ao756:~/git/server$ git checkout 10.4
...
openxs@ao756:~/git/server$ git pull
...
openxs@ao756:~/git/server$ git submodule update --init --recursive
Submodule path 'extra/wolfssl/wolfssl': checked out '9c87f979a7f1d3a6d786b260653d566c1d31a1c4'
Submodule path 'libmariadb': checked out '180c543704d627a50a52aaf60e24ca14e0ec4686'
Submodule path 'wsrep-lib': checked out 'f271ad0c6e3c647df83c1d5ec9cd26d77cef2337'
Submodule path 'wsrep-lib/wsrep-API/v26': checked out '76cf223c690845bbf561cb820a46e06a18ad80d1'
openxs@ao756:~/git/server$ git branch
  10.3
* 10.4
  10.5
  10.6
openxs@ao756:~/git/server$ git log -1
commit 583b72ad0ddbc46a7aaeda1c1373b89d4bded9ea (HEAD -> 10.4, origin/bb-10.4-merge, origin/10.4)
Merge: 473e85e9316 a4139f8d68b
Author: Oleksandr Byelkin <sanja@mariadb.com>
Date:   Fri May 7 11:50:24 2021 +0200

    Merge branch 'bb-10.4-release' into 10.4
openxs@ao756:~/git/server$ rm CMakeCache.txt
openxs@ao756:~/git/server$ cmake . -DCMAKE_INSTALL_PREFIX=/home/openxs/dbs/maria10.4 -DCMAKE_BUILD_TYPE=RelWithDebInfo -DBUILD_CONFIG=mysql_release -DFEATURE_SET=community -DWITH_EMBEDDED_SERVER=OFF -DPLUGIN_TOKUDB=NO -DWITH_SSL=system
...
-- Generating done
-- Build files have been written to: /home/openxs/git/server
openxs@ao756:~/git/server$ time make -j 3
...
[100%] Building C object extra/mariabackup/CMakeFiles/mariabackup.dir/__/__/libmysqld/libmysql.c.o
[100%] Linking CXX executable mariabackup
[100%] Built target mariabackup

real    74m9,550s
user    134m10,837s
sys     7m0,387s

openxs@ao756:~/git/server$ make install && make clean
...
openxs@ao756:~/git/server$ cd ~/dbs/maria10.4
openxs@ao756:~/dbs/maria10.4$ bin/mysqld_safe --no-defaults --port=3309 --socket=/tmp/mariadb.sock --innodb_buffer_pool_size=1G --innodb_flush_log_at_trx_commit=2 &
[1] 27483
openxs@ao756:~/dbs/maria10.4$ 210507 19:15:37 mysqld_safe Logging to '/home/openxs/dbs/maria10.4/data/ao756.err'.
210507 19:15:37 mysqld_safe Starting mysqld daemon with databases from /home/openxs/dbs/maria10.4/data

openxs@ao756:~/dbs/maria10.4$ bin/mysql --socket=/tmp/mariadb.sock
Welcome to the MariaDB monitor.  Commands end with ; or \g.
Your MariaDB connection id is 8
Server version: 10.4.19-MariaDB MariaDB Server

Copyright (c) 2000, 2018, Oracle, MariaDB Corporation Ab and others.

Type 'help;' or '\h' for help. Type '\c' to clear the current input statement.

MariaDB [(none)]> drop database if exists sbtest;
Query OK, 5 rows affected (1.297 sec)

MariaDB [(none)]> create database sbtest;
Query OK, 1 row affected (0.001 sec)

MariaDB [(none)]> exit
Bye

and compared to 10.4.15 with the following test:

openxs@ao756:~/dbs/maria10.4$ sysbench oltp_read_write --db-driver=mysql --tables=5 --table-size=100000 --mysql-user=openxs --mysql-socket=/tmp/mariadb.sock --mysql-db=sbtest --threads=4 prepare
sysbench 1.0.18 (using system LuaJIT 2.1.0-beta3)

Initializing worker threads...

Creating table 'sbtest2'...
Creating table 'sbtest3'...
Creating table 'sbtest4'...
Creating table 'sbtest1'...
Inserting 100000 records into 'sbtest1'
Inserting 100000 records into 'sbtest4'
Inserting 100000 records into 'sbtest2'
Inserting 100000 records into 'sbtest3'
Creating a secondary index on 'sbtest1'...
Creating a secondary index on 'sbtest4'...
Creating table 'sbtest5'...
Inserting 100000 records into 'sbtest5'
Creating a secondary index on 'sbtest2'...
Creating a secondary index on 'sbtest3'...
Creating a secondary index on 'sbtest5'...

openxs@ao756:~/dbs/maria10.4$ sysbench oltp_read_write --db-driver=mysql --tables=5 --table-size=100000 --mysql-user=openxs --mysql-socket=/tmp/mariadb.sock --mysql-db=sbtest --time=300 --report-interval=10 --threads=4 run
...
    transactions:                        35630  (118.54 per sec.)
    queries:                             712600 (2370.87 per sec.)
    ignored errors:                      0      (0.00 per sec.)
    reconnects:                          0      (0.00 per sec.)

General statistics:
    total time:                          300.5612s
    total number of events:              35630

Latency (ms):
         min:                                    3.30
         avg:                                   33.70
         max:                                 2200.17
         95th percentile:                      164.45
...

openxs@ao756:~/dbs/maria10.4.15$ sysbench oltp_read_write --db-driver=mysql --tables=5 --table-size=100000 --mysql-user=openxs --mysql-socket=/tmp/mariadb.sock --mysql-db=sbtest --time=300 --report-interval=10 --threads=4 run
...
    transactions:                        56785  (189.25 per sec.)
    queries:                             1135700 (3784.99 per sec.)
    ignored errors:                      0      (0.00 per sec.)
    reconnects:                          0      (0.00 per sec.)

General statistics:
    total time:                          300.0501s
    total number of events:              56785

Latency (ms):
         min:                                    3.15
         avg:                                   21.13
         max:                                  704.36
         95th percentile:                      108.68
...

So, basically with the same test with all tables fitting into the buffer pool (1G) and all but few default settings current MariaDB 10.4.19 demonstrate up to 60% drop in throughput and increase of 95th latency on this netbook (even more than 15% or so reported on faster quad core Fedora desktop previously).

If you read MDEV-24272 carefully, the regression was tracked up to a specific commit, but I tried to apply various tools to actually see where more time is spent now, specifically. Profiling with perf and creating on-CPU flame graphs had not given me any clear insight that would explain that increase in latency, so my next idea was to trace off-CPU time spent, that is, try to find out how long MariaDB server waits and where in the code that's mostly happen.

For this I've used the offcputime tool:

openxs@ao756:~$ /usr/share/bcc/tools/offcputime -h
usage: offcputime [-h] [-p PID | -t TID | -u | -k] [-U | -K] [-d] [-f]
                  [--stack-storage-size STACK_STORAGE_SIZE]
                  [-m MIN_BLOCK_TIME] [-M MAX_BLOCK_TIME] [--state STATE]
                  [duration]

Summarize off-CPU time by stack trace

positional arguments:
  duration              duration of trace, in seconds

optional arguments:
  -h, --help            show this help message and exit
  -p PID, --pid PID     trace this PID only
  -t TID, --tid TID     trace this TID only
  -u, --user-threads-only
                        user threads only (no kernel threads)
  -k, --kernel-threads-only
                        kernel threads only (no user threads)
  -U, --user-stacks-only
                        show stacks from user space only (no kernel space
                        stacks)
  -K, --kernel-stacks-only
                        show stacks from kernel space only (no user space
                        stacks)
  -d, --delimited       insert delimiter between kernel/user stacks
  -f, --folded          output folded format
  --stack-storage-size STACK_STORAGE_SIZE
                        the number of unique stack traces that can be stored
                        and displayed (default 1024)
  -m MIN_BLOCK_TIME, --min-block-time MIN_BLOCK_TIME
                        the amount of time in microseconds over which we store
                        traces (default 1)
  -M MAX_BLOCK_TIME, --max-block-time MAX_BLOCK_TIME
                        the amount of time in microseconds under which we
                        store traces (default U64_MAX)
  --state STATE         filter on this thread state bitmask (eg, 2 ==
                        TASK_UNINTERRUPTIBLE) see include/linux/sched.h

examples:
    ./offcputime             # trace off-CPU stack time until Ctrl-C
    ./offcputime 5           # trace for 5 seconds only
    ./offcputime -f 5        # 5 seconds, and output in folded format
    ./offcputime -m 1000     # trace only events that last more than 1000 usec
    ./offcputime -M 10000    # trace only events that last less than 10000 usec
    ./offcputime -p 185      # only trace threads for PID 185
    ./offcputime -t 188      # only trace thread 188
    ./offcputime -u          # only trace user threads (no kernel)
    ./offcputime -k          # only trace kernel threads (no user)
    ./offcputime -U          # only show user space stacks (no kernel)
    ./offcputime -K          # only show kernel space stacks (no user)

I've stored the outputs in /dev/shm to have less impact on disk I/O that I suspected as one of the reasons:

openxs@ao756:~$ ls /dev/shm
openxs@ao756:~$ mkdir /dev/shm/out
openxs@ao756:~$ ls /dev/shm
out

Basically the following commands were used to create folded (ready to use for building flame graphs) user-space only stacks and time spent off-CPU in them over 60 seconds of tracing while sysbench tests were running for clean setup on MariaDB 10.4.15 and then on current 10.4.19, and create flame graphs based on those stacks where start_thread frame is present (to clean up irrelevant details):

openxs@ao756:~$ sudo /usr/share/bcc/tools/offcputime -f -p `pidof mysqld` -U 60 > /dev/shm/out/offcpu_10415.out
WARNING: 27 stack traces lost and could not be displayed.
openxs@ao756:~$ cat /dev/shm/out/offcpu_10415.out | grep start_thread | ~/git/FlameGraph/flamegraph.pl --color=io --title="Off-CPU Time Flame Graph" --countname=us > /tmp/offcpu_10415.svg

openxs@ao756:~$ sudo /usr/share/bcc/tools/offcputime -f -p `pidof mysqld` -U 60 > /dev/shm/out/offcpu_10419.out
WARNING: 24 stack traces lost and could not be displayed.
openxs@ao756:~$ cat /dev/shm/out/offcpu_10419.out | grep start_thread | ~/git/FlameGraph/flamegraph.pl --color=io --title="Off-CPU Time Flame Graph" --countname=us > /tmp/offcpu_10419.svg

As a result I've got the following graphs (.png screenshots from real .svg files below). On 10.4.15 we spent around 43 seconds (out of 60 we monitored) off-CPU, mostly in do_command() and waiting for network I/O, and the graph was the following:

Off-CPU time for MariaDB 10.4.15

In case of 10.4.19 the graph is very different and we seem to have spent 79 seconds off-CPU, mostly in background io_handler_thread():

Off-CPU time for MariaDB 10.4.19

I was surprised to see more than 60 seconds spent off-CPU in this case. Maybe this is possible because I have 2 cores and MariaDB threads were waiting on both most of the time.

I've then tried to use differential flame graph to highlight the call stacks that the main difference is related to. I've crated it from existing folded stack traces with the following command:

openxs@ao756:~$ ~/git/FlameGraph/difffolded.pl /dev/shm/out/offcpu_10415.out /dev/shm/out/offcpu_10419.out | grep start_thread | ~/git/FlameGraph/flamegraph.pl --color=io --title="Off-CPU Time Diff Flame Graph" --countname=us > /tmp/offcpu_diff.svg

The resulting graph is presented below:

The source of increase is highlighted in red

Here we clearly see that main increase in time spent waiting in 10.4.19 is related to io_handler_thread(), but increase happened in almost all background threads.

* * *

To summarize:

  • When some performance regression happens you should check not only those code paths in the application where software started to spend more time working, but also where it started to wait more.
  • In case of https://jira.mariadb.org/browse/MDEV-24272 we clearly started to flush more to disk from the very beginning of sysbench oltp_read_write test in newer versions of MariaDB 10.4.x, and on my slow encrypted HDD this matters a lot. The load that was supposed to be CPU-bound (as we have large enough buffer pool) becomes disk-bound.
  • Flame graphs are cool for highlighting the difference and in this post I've shown both a classical smart way to produce them without too much impact, and a way to highlight the difference in them with a differential flam graph produced by the difffolded.pl tool created by Brendan Gregg.
  • Other cases when Flame Graphs may help MySQL or MariaDB DBAs are discussed during my upcoming Percona Live 2021 talk on May 12. See you there!
  • I'll get back to this nice regression bug to study the test case in more details with other tools, maybe more than once. Stay tuned!