Debug Tool

In the process of using and developing Doris, we often encounter scenarios that need to debug Doris. Here are some common debugging tools.

The name of the BE binary that appears in this doc is doris_be, which was palo_be in previous versions.

FE debugging

Fe is a java process. Here are just a few simple and commonly used java debugging commands.

1. Statistics of current memory usage details

   jmap -histo:live pid > 1. jmp

   This command can enumerate and sort the memory occupation of living objects. (replace PID with Fe process ID)

   num     #instances         #bytes  class name
   ----------------------------------------------
   1:         33528       10822024  [B
   2:         80106        8662200  [C
   3:           143        4688112  [Ljava.util.concurrent.ForkJoinTask;
   4:         80563        1933512  java. lang.String
   5:         15295        1714968  java. lang.Class
   6:         45546        1457472  java. util. concurrent. ConcurrentHashMap$Node
   7:         15483        1057416  [Ljava.lang.Object;

   You can use this method to view the total memory occupied by the currently living objects (at the end of the file) and analyze which objects occupy more memory.

   Note that this method will trigger fullgc because `: live 'is specified.

2. Check JVM memory usage

   jstat -gcutil pid 1000 1000

   This command can scroll through the memory usage of each region of the current JVM. (replace PID with Fe process ID)

   S0     S1     E      O      M     CCS    YGC     YGCT    FGC    FGCT     GCT
   0.00   0.00  22.61   3.03  95.74  92.77     68    1.249     5    0.794    2.043
   0.00   0.00  22.61   3.03  95.74  92.77     68    1.249     5    0.794    2.043
   0.00   0.00  22.61   3.03  95.74  92.77     68    1.249     5    0.794    2.043
   0.00   0.00  22.92   3.03  95.74  92.77     68    1.249     5    0.794    2.043
   0.00   0.00  22.92   3.03  95.74  92.77     68    1.249     5    0.794    2.043

   The main focus is on the percentage of old area (o) (3% in the example). If the occupancy is too high, oom or fullgc may occur.

3. Print Fe thread stack

   jstack -l pid > 1. js

   This command can print the thread stack of the current Fe. (replace PID with Fe process ID). -L the parameter will detect whether there is deadlock at the same time. This method can check the operation of Fe thread, whether there is deadlock, where it is stuck, etc.

BE debugging

Preparing

pprof: from gperftools, it is used to transform the content generated by gperftools into a format that is easy for people to read, such as PDF, SVG, text, etc.

graphviz: in the absence of this library, pprof can only be converted to text format, but this way is not easy to view. After the library is installed, pprof can be converted to SVG, PDF and other formats, and the call relationship is clearer.

perf: Linux kernel comes with performance analysis tool. here there are some examples of perf usage.

flamegraph: a visualization tool used to show the output of perf in the form of flame graph.

Memory

Debugging memory is generally divided into two aspects. One is whether the total amount of memory use is reasonable. On the one hand, the excessive amount of memory use may be due to memory leak in the system, on the other hand, it may be due to improper use of program memory. The second is whether there is a problem of memory overrun and illegal access, such as program access to memory with an illegal address, use of uninitialized memory, etc. For the debugging of memory, we usually use the following ways to track the problems.

Doris 1.2.1 and previous versions use TCMalloc. Doris 1.2.2 starts to use Jemalloc by default. Select the memory debugging method according to the Doris version used. If you need to switch TCMalloc, you can compile USE_JEMALLOC=OFF sh build.sh --be.

When we find that the memory usage is too large, we can first check the BE log to see if there is a large memory application.

TCMalloc

When using TCMalloc, when a large memory application is encountered, the application stack will be printed to the be.out file, and the general expression is as follows:

tcmalloc: large alloc 1396277248 bytes == 0x3f3488000 @  0x2af6f63 0x2c4095b 0x134d278 0x134bdcb 0x133d105 0x133d1d0 0x19930ed

This indicates that Doris be is trying to apply memory of '1396277248 bytes' on this stack. We can use the 'addr2line' command to restore the stack to a letter that we can understand. The specific example is shown below.

$ addr2line -e lib/doris_be  0x2af6f63 0x2c4095b 0x134d278 0x134bdcb 0x133d105 0x133d1d0 0x19930ed

/home/ssd0/zc/palo/doris/core/thirdparty/src/gperftools-gperftools-2.7/src/tcmalloc.cc:1335
/home/ssd0/zc/palo/doris/core/thirdparty/src/gperftools-gperftools-2.7/src/tcmalloc.cc:1357
/home/disk0/baidu-doris/baidu/bdg/doris-baidu/core/be/src/exec/hash_table.cpp:267
/home/disk0/baidu-doris/baidu/bdg/doris-baidu/core/be/src/exec/hash_table.hpp:86
/home/disk0/baidu-doris/baidu/bdg/doris-baidu/core/be/src/exec/hash_join_node.cpp:239
/home/disk0/baidu-doris/baidu/bdg/doris-baidu/core/be/src/exec/hash_join_node.cpp:213
thread.cpp:?

JEMALLOC

Most of Doris's large memory applications use Allocator, such as HashTable and data serialization. This part of the memory application is expected and will be effectively managed. Other large memory applications are not expected and will be applied The stack is printed to the be.INFO file, which is usually used for debugging, and the general expression is as follows:

MemHook alloc large memory: 8.2GB, stacktrace:
Alloc Stacktrace:
    @     0x55a6a5cf6b4d  doris::ThreadMemTrackerMgr::consume()
    @     0x55a6a5cf99bf  malloc
    @     0x55a6ae0caf98  operator new()
    @     0x55a6a57cb013  doris::segment_v2::PageIO::read_and_decompress_page()
    @     0x55a6a57719c0  doris::segment_v2::ColumnReader::read_page()
    ……

HEAP PROFILE

TCMalloc

Sometimes the application of memory is not caused by the application of large memory, but by the continuous accumulation of small memory. Then there is no way to locate the specific application information by viewing the log, so you need to get the information through other ways.

At this time, we can take advantage of TCMalloc's heapprofile. If the heapprofile function is set, we can get the overall memory application usage of the process. The usage is to set the 'heapprofile' environment variable before starting Doris be. For example:

export HEAPPROFILE=/tmp/doris_be.hprof
./bin/start_be.sh --daemon

In this way, when the dump condition of the heapprofile is met, the overall memory usage will be written to the file in the specified path. Later, we can use the 'pprof' tool to analyze the output content.

$ pprof --text lib/doris_be /tmp/doris_be.hprof.0012.heap | head -30

Using local file lib/doris_be.
Using local file /tmp/doris_be.hprof.0012.heap.
Total: 668.6 MB
   610.6  91.3%  91.3%    610.6  91.3% doris::SystemAllocator::allocate_via_malloc (inline)
    18.1   2.7%  94.0%     18.1   2.7% _objalloc_alloc
     5.6   0.8%  94.9%     63.4   9.5% doris::RowBatch::RowBatch
     5.1   0.8%  95.6%      7.1   1.1% butil::ResourcePool::add_block (inline)
     3.7   0.5%  96.2%      3.7   0.5% butil::iobuf::create_block (inline)
     3.4   0.5%  96.7%      3.4   0.5% butil::FlatMap::init
     3.2   0.5%  97.2%      5.2   0.8% butil::ObjectPool::add_block (inline)
     2.6   0.4%  97.6%      2.6   0.4% __gnu_cxx::new_allocator::allocate (inline)
     2.0   0.3%  97.9%      2.0   0.3% butil::ObjectPool::add_block_group (inline)
     2.0   0.3%  98.2%      2.0   0.3% butil::ResourcePool::add_block_group (inline)
     1.7   0.3%  98.4%      1.7   0.3% doris::SegmentReader::_load_index

Contents of each column of the above documents:

• Column 1: the memory size directly applied by the function, in MB
• Column 4: the total memory size of the function and all the functions it calls.
• The second column and the fifth column are the proportion values of the first column and the fourth column respectively.
• The third column is the cumulative value of the second column.

Of course, it can also generate call relation pictures, which is more convenient for analysis. For example, the following command can generate a call graph in SVG format.

pprof --svg lib/doris_be /tmp/doris_be.hprof.0012.heap > heap.svg 

NOTE: turning on this option will affect the execution performance of the program. Please be careful to turn on the online instance.

pprof remote server

Although heapprofile can get all the memory usage information, it has some limitations. 1. Restart be. 2. You need to enable this command all the time, which will affect the performance of the whole process.

For Doris be, you can also use the way of opening and closing the heap profile dynamically to analyze the memory application of the process. Doris supports the remote server debugging of gperftools. Then you can use 'pprof' to directly perform dynamic head profile on the remote running Doris be. For example, we can check the memory usage increment of Doris through the following command

$ pprof --text --seconds=60 http://be_host:be_webport/pprof/heap 

Total: 1296.4 MB
   484.9  37.4%  37.4%    484.9  37.4% doris::StorageByteBuffer::create
   272.2  21.0%  58.4%    273.3  21.1% doris::RowBlock::init
   157.5  12.1%  70.5%    157.5  12.1% doris::RowBatch::RowBatch
    90.7   7.0%  77.5%     90.7   7.0% doris::SystemAllocator::allocate_via_malloc
    66.6   5.1%  82.7%     66.6   5.1% doris::IntegerColumnReader::init
    47.9   3.7%  86.4%     47.9   3.7% __gnu_cxx::new_allocator::allocate
    20.8   1.6%  88.0%     35.4   2.7% doris::SegmentReader::_load_index
    12.7   1.0%  89.0%     12.7   1.0% doris::DecimalColumnReader::init
    12.7   1.0%  89.9%     12.7   1.0% doris::LargeIntColumnReader::init
    12.7   1.0%  90.9%     12.7   1.0% doris::StringColumnDirectReader::init
    12.3   0.9%  91.9%     12.3   0.9% std::__cxx11::basic_string::_M_mutate
    10.4   0.8%  92.7%     10.4   0.8% doris::VectorizedRowBatch::VectorizedRowBatch
    10.0   0.8%  93.4%     10.0   0.8% doris::PlainTextLineReader::PlainTextLineReader

The output of this command is the same as the output and view mode of heap profile, which will not be described in detail here. Statistics will be enabled only during execution of this command, which has a limited impact on process performance compared with heap profile.

JEMALLOC

For the analysis of the principle of Heap Profile, please refer to Analysis of the Principle of Heap Profiling. It should be noted that Heap Profile records virtual memory

It supports real-time and regular dumping of Heap Profile, and then uses jeprof to analyze Heap Profile.

1. realtime heap dump

Change prof:false of JEMALLOC_CONF in be.conf to prof:true and restart BE, then use the jemalloc heap dump http interface to generate a heap dump file on the corresponding BE machine.

curl http://be_host:be_webport/jeheap/dump

The directory where the heap dump file is located can be configured through the jeprofile_dir variable in be.conf, and the default is ${DORIS_HOME}/log

The default sampling interval is 512K, usually only 10% of memory is recorded by heap dump, and the impact on performance is usually less than 10%. You can modify lg_prof_sample of JEMALLOC_CONF in be.conf, and the default is 19 (2^19 B = 512K), reducing lg_prof_sample can sample more frequently to make the heap profile close to the real memory, but this will bring greater performance loss.

If you are doing profiling, keep prof:false to avoid the performance penalty of heap dump.

2. regular heap dump

First, change prof:false of JEMALLOC_CONF in be.conf to prof:true. The default directory of the heap dump file is ${DORIS_HOME}/log. The file name prefix is ​​JEMALLOC_PROF_PRFIX in be.conf, which is jemalloc_heap_profile_ by default.

Before Doris 2.1.6, JEMALLOC_PROF_PRFIX is empty and needs to be changed to any value as the profile file name

1. Dump when the cumulative memory application reaches a certain value:

Change lg_prof_interval of JEMALLOC_CONF in be.conf to 34. At this time, the profile is dumped once when the cumulative memory application reaches 16GB (2^35 B = 16GB). You can change it to any value to adjust the dump interval.

Before Doris 2.1.6, lg_prof_interval defaults to 32.

2. Dump every time the memory reaches a new high:

Change prof_gdump in JEMALLOC_CONF in be.conf to true and restart BE.

3. Dump when the program exits, and detect memory leaks:

Change prof_leak and prof_final in JEMALLOC_CONF in be.conf to true and restart BE.

4. Dump the cumulative value (growth) of memory instead of the real-time value:

Change prof_accum in JEMALLOC_CONF in be.conf to true and restart BE. Use jeprof --alloc_space to display the cumulative value of heap dump.

3. jeprof parses Heap Profile

Use jeprof to parse the Heap Profile of the above dump. If the process memory is too large, the parsing process may take several minutes, so please wait patiently. If the system does not have the jeprof command, you can package the jeprof binary in the doris/tools directory and upload it to the Heap Dump server.

Addr2line version 2.35.2 or above is required, see QA-1 below
Try to have Heap Dump and `jeprof` to parse Heap Profile on the same server, see QA-2 below

1. Analyze a single heap dump file

jeprof --dot lib/doris_be heap_dump_file_1

After executing the above command, a dot syntax graph will be output in the terminal. Paste it to the online dot drawing website to generate a memory allocation graph, and then analyze it.

If the server is convenient for file transfer, you can also use the following command to directly generate a call relationship graph result.pdf file and transfer it to the local computer for viewing. You need to install the dependencies required for drawing.

yum install ghostscript graphviz
jeprof --pdf lib/doris_be heap_dump_file_1 > result.pdf

2. Analyze the diff of two heap dump files

jeprof --dot lib/doris_be --base=heap_dump_file_1 heap_dump_file_2

Multiple heap files can be generated by running the above command multiple times over a period of time. You can select an earlier heap file as a baseline and compare it with a later heap file to analyze their diff. The method for generating a call relationship graph is the same as above.

4. QA

1. Many errors occurred after running jeprof: addr2line: Dwarf Error: found dwarf version xxx, this reader only handles version xxx

After GCC 11, DWARF-v5 is used by default, which requires Binutils 2.35.2 and above. Doris Ldb_toolchain uses GCC 11. See: https://gcc.gnu.org/gcc-11/changes.html.

Replace addr2line to 2.35.2, refer to:

// Download addr2line source code
wget https://ftp.gnu.org/gnu/binutils/binutils-2.35.tar.bz2

//Install dependencies, if needed
yum install make gcc gcc-c++ binutils

// Compile & install addr2line
tar -xvf binutils-2.35.tar.bz2
cd binutils-2.35
./configure --prefix=/usr/local
make
make install

// verify
addr2line -h

// Replace addr2line
chmod +x addr2line
mv /usr/bin/addr2line /usr/bin/addr2line.bak
mv /bin/addr2line /bin/addr2line.bak
cp addr2line /bin/addr2line
cp addr2line /usr/bin/addr2line
hash -r

Note that you cannot use addr2line 2.3.9, which may be incompatible and cause memory to keep growing.

2. After running jeprof, many errors appear: addr2line: DWARF error: invalid or unhandled FORM value: 0x25. The parsed Heap stack is the memory address of the code, not the function name

This is because the Heap Dump and the execution of jeprof to parse the Heap Profile are not on the same server, which causes jeprof to fail to parse the function name using the symbol table. Try to complete the Dump Heap and jeprof parsing operations on the same machine.

LSAN

LSAN is an address checking tool, GCC has been integrated. When we compile the code, we can enable this function by turning on the corresponding compilation options. When the program has a determinable memory leak, it prints the leak stack. Doris be has integrated this tool, only need to compile with the following command to generate be binary with memory leak detection version.

BUILD_TYPE=LSAN ./build.sh

When the system detects a memory leak, it will output the corresponding information in be. Out. For the following demonstration, we intentionally insert a memory leak code into the code. We insert the following code into the open function of StorageEngine.

    char* leak_buf = new char[1024];
    strcpy(leak_buf, "hello world");
    LOG(INFO) << leak_buf;

We get the following output in be.out

=================================================================
==24732==ERROR: LeakSanitizer: detected memory leaks

Direct leak of 1024 byte(s) in 1 object(s) allocated from:
    #0 0xd10586 in operator new[](unsigned long) ../../../../gcc-7.3.0/libsanitizer/lsan/lsan_interceptors.cc:164
    #1 0xe333a2 in doris::StorageEngine::open(doris::EngineOptions const&, doris::StorageEngine**) /home/ssd0/zc/palo/doris/core/be/src/olap/storage_engine.cpp:104
    #2 0xd3cc96 in main /home/ssd0/zc/palo/doris/core/be/src/service/doris_main.cpp:159
    #3 0x7f573b5eebd4 in __libc_start_main (/opt/compiler/gcc-4.8.2/lib64/libc.so.6+0x21bd4)

SUMMARY: LeakSanitizer: 1024 byte(s) leaked in 1 allocation(s).

From the above output, we can see that 1024 bytes have been leaked, and the stack information of memory application has been printed out.

NOTE: turning on this option will affect the execution performance of the program. Please be careful to turn on the online instance.

NOTE: if the LSAN switch is turned on, the TCMalloc will be automatically turned off

JEMALLOC HEAP PROFILE

1. runtime heap dump by http

Add ,prof:true,lg_prof_sample:10 to JEMALLOC_CONF in start_be.sh and restart BE, then use the jemalloc heap dump http interface to generate a heap dump file on the corresponding BE machine.

The directory where the heap dump file is located can be configured through the jeprofile_dir variable in be.conf, and the default is ${DORIS_HOME}/log

curl http://be_host:be_webport/jeheap/dump

prof: After opening, jemalloc will generate a heap dump file according to the current memory usage. There is a small amount of performance loss in heap profile sampling, which can be turned off during performance testing. lg_prof_sample: heap profile sampling interval, the default value is 19, that is, the default sampling interval is 512K (2^19 B), which will result in only 10% of the memory recorded by the heap profile, lg_prof_sample:10 can reduce the sampling interval to 1K (2^10 B), more frequent sampling will make the heap profile close to real memory, but this will bring greater performance loss.

For detailed parameter description, refer to https://linux.die.net/man/3/jemalloc.

2. jemalloc heap dump profiling

3.1 Generating plain text analysis results

jeprof lib/doris_be --base=heap_dump_file_1 heap_dump_file_2

3.2 Generate call relationship picture

Install dependencies required for plotting

yum install ghostscript graphviz

Multiple dump files can be generated by running the above command multiple times in a short period of time, and the first dump file can be selected as the baseline for diff comparison analysis

jeprof --dot lib/doris_be --base=heap_dump_file_1 heap_dump_file_2

After executing the above command, the terminal will output a diagram of dot syntax, and paste it to online dot drawing website, generate a memory allocation diagram, and then analyze it. This method can Drawing directly through the terminal output results is more suitable for servers where file transfer is not very convenient.

You can also use the following command to directly generate the call relationship result.pdf file and transfer it to the local for viewing

jeprof --pdf lib/doris_be --base=heap_dump_file_1 heap_dump_file_2 > result.pdf

In the above jeprof related commands, remove the --base option to analyze only a single heap dump file

3. heap dump by JEMALLOC_CONF

Perform heap dump by restarting BE after changing the JEMALLOC_CONF variable in start_be.sh

1. Dump every 1MB:

   Two new variable settings prof:true,lg_prof_interval:20 have been added to the JEMALLOC_CONF variable, where prof:true is to enable profiling, and lg_prof_interval:20 means that a dump is generated every 1MB (2^20)

2. Dump each time a new high is reached:

   Added two variable settings prof:true,prof_gdump:true in the JEMALLOC_CONF variable, where prof:true is to enable profiling, and prof_gdump:true means to generate a dump when the memory usage reaches a new high

3. Memory leak dump when the program exits:

   Added three new variable settings prof_leak: true, lg_prof_sample: 0, prof_final: true in the JEMALLOC_CONF variable

ASAN

Except for the unreasonable use and leakage of memory. Sometimes there will be memory access illegal address and other errors. At this time, we can use ASAN to help us find the cause of the problem. Like LSAN, ASAN is integrated into GCC. Doris can open this function by compiling as follows

BUILD_TYPE=ASAN ./build.sh

Execute the binary generated by compilation. When the detection tool finds any abnormal access, it will immediately exit and output the stack illegally accessed in be.out. The output of ASAN is the same as that of LSAN. Here we also actively inject an address access error to show the specific content output. We still inject an illegal memory access into the 'open' function of 'storageengine'. The specific error code is as follows

    char* invalid_buf = new char[1024];
    for (int i = 0; i < 1025; ++i) {
        invalid_buf[i] = i;
    }
    LOG(INFO) << invalid_buf;

We get the following output in be.out

=================================================================
==23284==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x61900008bf80 at pc 0x00000129f56a bp 0x7fff546eed90 sp 0x7fff546eed88
WRITE of size 1 at 0x61900008bf80 thread T0
    #0 0x129f569 in doris::StorageEngine::open(doris::EngineOptions const&, doris::StorageEngine**) /home/ssd0/zc/palo/doris/core/be/src/olap/storage_engine.cpp:106
    #1 0xe2c1e3 in main /home/ssd0/zc/palo/doris/core/be/src/service/doris_main.cpp:159
    #2 0x7fa5580fbbd4 in __libc_start_main (/opt/compiler/gcc-4.8.2/lib64/libc.so.6+0x21bd4)
    #3 0xd30794  (/home/ssd0/zc/palo/doris/core/output3/be/lib/doris_be+0xd30794)

0x61900008bf80 is located 0 bytes to the right of 1024-byte region [0x61900008bb80,0x61900008bf80)
allocated by thread T0 here:
    #0 0xdeb040 in operator new[](unsigned long) ../../../../gcc-7.3.0/libsanitizer/asan/asan_new_delete.cc:82
    #1 0x129f50d in doris::StorageEngine::open(doris::EngineOptions const&, doris::StorageEngine**) /home/ssd0/zc/palo/doris/core/be/src/olap/storage_engine.cpp:104
    #2 0xe2c1e3 in main /home/ssd0/zc/palo/doris/core/be/src/service/doris_main.cpp:159
    #3 0x7fa5580fbbd4 in __libc_start_main (/opt/compiler/gcc-4.8.2/lib64/libc.so.6+0x21bd4)

SUMMARY: AddressSanitizer: heap-buffer-overflow /home/ssd0/zc/palo/doris/core/be/src/olap/storage_engine.cpp:106 in doris::StorageEngine::open(doris::EngineOptions const&, doris::StorageEngine**)

From this message, we can see that at the address of 0x61900008bf80, we tried to write a byte, but this address is illegal. We can also see the application stack of the address [0x61900008bb80, 0x61900008bf80].

NOTE: turning on this option will affect the execution performance of the program. Please be careful to turn on the online instance.

NOTE: if the ASAN switch is turned on, the TCMalloc will be automatically turned off

In addition, if stack information is output in be.out, but there is no function symbol, then we need to handle it manually to get readable stack information. The specific processing method needs a script to parse the output of ASAN. At this time, we need to use asan_symbolize to help with parsing. The specific usage is as follows:

cat be.out | python asan_symbolize.py | c++filt

With the above command, we can get readable stack information.

CPU

When the CPU idle of the system is very low, it means that the CPU of the system has become the main bottleneck. At this time, it is necessary to analyze the current CPU usage. For the be of Doris, there are two ways to analyze the CPU bottleneck of Doris.

pprof

Because Doris has integrated and compatible with GPERF rest interface, users can analyze remote Doris be through the 'pprof' tool. The specific usage is as follows:

pprof --svg --seconds=60 http://be_host:be_webport/pprof/profile > be.svg 

In this way, a CPU consumption graph of be execution can be generated.

perf + flamegragh

This is a quite common CPU analysis method. Compared with pprof, this method must be able to log in to the physical machine of the analysis object. However, compared with pprof, which can only collect points on time, perf can collect stack information through different events. The specific usage is as follows:

perf record -g -p be_pid -- sleep 60

This command counts the CPU operation of be for 60 seconds and generates perf.data. For the analysis of perf.data, the command of perf can be used for analysis.

perf report

The analysis results in the following pictures

To analyze the generated content. Of course, you can also use flash graph to complete the visual display.

perf script | ./FlameGraph/stackcollapse-perf.pl | ./FlameGraph/flamegraph.pl > be.svg

This will also generate a graph of CPU consumption at that time.