Crashing: gmetad of Ganglia crashing because of Buffer Overflow

Question

I use Ganglia to monitor Hadoop Flume Agents' performance. For almost 1 year now, it had been working very well. Last week gmetad started crashing with buffer overflow. Only thing that has changed in last few days is we started monitoring more instances of flume agent.

> gmond -V
gmond 3.7.2
> gmetad -V
gmetad 3.7.2

Below is the output I get when I run gmetad at command prompt with debug=100. Please suggest how to overcome the buffer overflow problem.

Writing Summary data for source atl-ganglia, metric flume.SINK.hdfsSink.StartTime
Writing Summary data for source atl-ganglia, metric flume.CHANNEL.fileChannel.ChannelFillPercentage
*** buffer overflow detected ***: /usr/sbin/gmetad terminated
======= Backtrace: =========
/lib64/libc.so.6(__fortify_fail+0x37)[0x7f9899751597]
/lib64/libc.so.6(+0x100480)[0x7f989974f480]
/lib64/libc.so.6(+0xff8d9)[0x7f989974e8d9]
/lib64/libc.so.6(_IO_default_xsputn+0xc9)[0x7f98996c3639]
/lib64/libc.so.6(_IO_vfprintf+0x41c0)[0x7f9899697190]
/lib64/libc.so.6(__vsprintf_chk+0x9d)[0x7f989974e97d]
/lib64/libc.so.6(__sprintf_chk+0x7f)[0x7f989974e8bf]
/usr/sbin/gmetad[0x40a714]
/usr/sbin/gmetad[0x40841f]
/usr/lib64/libganglia.so.0(hash_foreach+0x59)[0x7f989af731f9]
/usr/sbin/gmetad[0x408151]
/lib64/libexpat.so.1(+0xa836)[0x7f9899e16836]
/lib64/libexpat.so.1(+0xbbce)[0x7f9899e17bce]
/lib64/libexpat.so.1(+0xd4fa)[0x7f9899e194fa]
/lib64/libexpat.so.1(+0xde3b)[0x7f9899e19e3b]
/lib64/libexpat.so.1(XML_ParseBuffer+0x6d)[0x7f9899e1288d]
/usr/sbin/gmetad[0x409ddc]
/usr/sbin/gmetad[0x405116]
/lib64/libpthread.so.0(+0x7a51)[0x7f98999eaa51]
/lib64/libc.so.6(clone+0x6d)[0x7f989973796d]
======= Memory map: ========
00400000-00417000 r-xp 00000000 08:02 1077406                            /usr/sbin/gmetad
00617000-00619000 rw-p 00017000 08:02 1077406                            /usr/sbin/gmetad
00619000-0061a000 rw-p 00000000 00:00 0
0157e000-0159f000 rw-p 00000000 00:00 0                                  [heap]
7f988c000000-7f988c116000 rw-p 00000000 00:00 0
7f988c116000-7f9890000000 ---p 00000000 00:00 0
7f9890f0f000-7f9890f25000 r-xp 00000000 08:02 1569794                    /lib64/libgcc_s-4.4.7-20120601.so.1
7f9890f25000-7f9891124000 ---p 00016000 08:02 1569794                    /lib64/libgcc_s-4.4.7-20120601.so.1
.
.
.
7f989817f000-7f989837e000 ---p 00003000 08:02 1569886                    /lib64/libgmodule-2.0.so.0.2800.8
7f989837e000-7f989837f000 rw-p 00002000 08:02 1569886                    /lib64/libgmodule-2.0.so.0.2800.8/bin/bash: line 1: 19584 Aborted                 /usr/sbin/gmetad
                                                           [FAILED]

Answer 1

Looks like the format of performance numbers sent in XML formatted string by Flume to Ganglia is different when the Flume agent run on Linux vs Windows. From Linux based Flume agent, the numbers are formatted as Float, whereas from Windows, the float comes further formatted in scientific notation.

This is from CentOs based Flume agent:

<METRIC NAME="flume.CHANNEL.memoryChannel.ChannelFillPercentage" VAL=**"0.018000000000000002"** TYPE="float" UNITS="" TN="39" TMAX="60" DMAX="0" SLOPE="both">

This is from Windows based Flume agent:

<METRIC NAME="flume.CHANNEL.fileChannel.ChannelFillPercentage" VAL=**"1.7976931348623157E308"** TYPE="float" UNITS="" TN="0" TMAX="60" DMAX="0" SLOPE="both">  

Seems that the Ganglia daemon gmetad parses the XML string but does try to convert the string to number before storing in rrds. That when the buffer overflow occurs.

I stopped sending data from Windows based flume agents and now Ganglia monitoring is stable. Now need to find how to monitor windows based agents.