DTrace : détection de problèmes en temps réel

1 Introduction

Dtrace est un système de trace conçu par Sun Microsystems pour la détection de problèmes en temps réel au niveau noyau ou au niveau applicatif. Il est disponible depuis Novembre 2003, et a été intégré en tant que partie de Solaris 10 en Janvier 2005. Dtrace est le premier composant du projet OpenSolaris dont le code a été délivré sous la license Common Development and Distribution License (CDDL).

Dtrace est un système de trace conçu pour donner des informations qui permettent aux utilisateurs d'ajuster des applications et le système d'exploitation lui-même. Il est conçu pour être utilisé dans des environements de production. Ainsi, l'effet des sondes est minimal quand l'action de trace est en cours, et il n'y a pas d'impact de performance pour les sondes non actives. C'est important car un système comprend des dizaines de milliers de sondes dont beaucoup peuvent être actives.

On écrit les programmes de trace (souvent appelés script) en utilisant le langage de programmation D (à ne pas confondre avec D). D est un sous-ensemble du langage C avec l'addition de fonctions et variables prédéfinies spécifiques à l'opération de trace. Un programme écrit en D ressemble par sa structure à un programme écrit en AWK.

Sachant qu'il est toujours un peu long de se faire ses propres scripts à chaque fois, j'ai préférer mettre ici tous ceux que j'ai utilié.

2 Print Utilization statistics per process

Brendan Gregg developed prustat to display the top processes sorted by CPU, Memory, Disk or Network utilization:

prustat

$ prustat -c -t 10 5 PID  %CPU  %Mem  %Disk  %Net COMM 7176 0.88 0.70 0.00 0.00 dtrace 7141 0.00 0.43 0.00 0.00 sshd 7144 0.11 0.24 0.00 0.00 sshd 3 0.34 0.00 0.00 0.00 fsflush 7153 0.03 0.19 0.00 0.00 bash 99 0.00 0.22 0.00 0.00 nscd 7146 0.00 0.19 0.00 0.00 bash 52 0.00 0.17 0.00 0.00 vxconfigd 7175 0.07 0.09 0.00 0.00 sh 98 0.00 0.16 0.00 0.00 kcfd

This script is super useful for getting a high level understanding of what is happening to a Solaris server. Golden!

3 File System Flush Activity

On Solaris systems, the pagedaemon is reponsible for scanning the page cache and adjusting the MMU reference bit of each dirty page it finds. When the fsflush daemon runs, it scans the page cache looking for pages with the MMU reference bit set, and schedules these pages to be written to disk. The fsflush.d D script provides a detailed breakdown of pages scanned, and the number of nanoseconds that were required to scan "SCANNED" pages:

fsflush.d

$ fsflush.d SCANNED EXAMINED LOCKED MODIFIED COALESCE RELEASES TIME(ns) 4254 4255 1 1 0 0 2695024 4254 4255 1 0 0 0 1921518 4254 4255 6 0 0 0 1989044 4254 4255 1 0 0 0 2401266 4254 4255 4 1 0 0 2562138 4254 4255 89 4 0 0 2425988 4254 3744 80 25 0 0 2394895 4254 4255 28 8 0 0 1776222 4254 4255 216 8 0 0 2350826 4254 4255 108 7 0 0 2356146

Now you might be wondering why "SCANNED" is less than "EXAMINED?" This is due to a bug in fsflush, and a bug report was filed to address this anomaly. Tight!

4 Seek Sizes

Prior to Solaris 10, determining if an application accessed data in a sequential or random pattern required reviewing mounds of truss(1m) and vxtrace(1m) data. With the introduction of DTrace and Brendan Gregg's seeksize.d D script, this question is trivial to answer:

seeksize.d

$ seeksize.d Sampling... Hit Ctrl-C to end. ^C   PID CMD 7312 dd if=/dev/dsk/c1t1d0s2 of=/dev/null bs=1048576   value ------------- Distribution ------------- count -1 | 0 0 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 1762 1 | 0   0 sched   value ------------- Distribution ------------- count -1048576 | 0 -524288 |@@@@ 1 -262144 | 0 -131072 | 0 -65536 | 0 -32768 | 0 -16384 | 0 -8192 | 0 -4096 | 0 -2048 | 0 -1024 | 0 -512 | 0 -256 | 0 -128 | 0 -64 | 0 -32 | 0 -16 | 0 -8 | 0 -4 | 0 -2 | 0 -1 | 0 0 | 0 1 | 0 2 | 0 4 | 0 8 | 0 16 | 0 32 | 0 64 | 0 128 |@@@@ 1 256 |@@@@ 1 512 |@@@@ 1 1024 |@@@@ 1 2048 | 0 4096 | 0 8192 | 0 16384 |@@@@ 1 32768 |@@@@ 1 65536 |@@@@@@@@ 2 131072 | 0 262144 | 0 524288 |@@@@ 1 1048576 | 0

This script measures the seek distance between consecutive reads and writes, and provides a histogram with the seek distances. For applications that are using sequential access patterns (e.g., dd in this case), the distribution will be small. For applications accessing data in a random nature (e.g, sched in this example), you will see a wide distribution. Shibby!

5 Print Overall Paging Activity

Prior to the introduction of DTrace, it was difficult to extract data on which files and disk devices were active at a specific point in time. With the introduction of fspaging.d, you can get a detailed view of which files are being accessed:

fspaging.d

$ fspaging.d Event Device Path RW Size Offset get-page /lib/sparcv9/libc.so.1 8192 get-page /usr/lib/sparcv9/libdtrace.so.1 8192 get-page /lib/sparcv9/libc.so.1 8192 get-page /lib/libc.so.1 8192 put-page /etc/svc/volatile/system-system-log:default.log 8192 put-page /etc/svc/volatile/svc_nonpersist.db 8192 put-page /etc/svc/volatile/svc_nonpersist.db 8192 put-page /etc/svc/volatile/init-next.state 8192 put-page /etc/svc/volatile/network-ssh:default.log 8192 put-page /etc/svc/volatile/network-pfil:default.log 8192

This is a super useful script! Niiiiiiiiiiice!

6 Getting System Wide errno Information

When system calls have problems executing, they usually return a value to indicate success or failure, and set the global "ERRNO" variable to a value indicating what went wrong. To get a system wide view of which system calls are erroring out, we can use Brendan Gregg's errinfo D script:

errinfo

$ errinfo -c Sampling... Hit Ctrl-C to end. ^C EXEC SYSCALL ERR COUNT DESC ttymon read 11 1 Resource temporarily unavailable utmpd ioctl 25 2 Inappropriate ioctl for device init ioctl 25 4 Inappropriate ioctl for device nscd lwp_kill 3 13 No such process fmd lwp_park 62 48 timer expired nscd lwp_park 62 48 timer expired svc.startd lwp_park 62 48 timer expired vxesd accept 4 49 interrupted system call svc.configd lwp_park 62 49 timer expired inetd lwp_park 62 49 timer expired svc.startd portfs 62 490 timer expired

This will display the process, system call, and errno number and description from /usr/src/sys/errno.h! Jeah!

7 I/O per process

Several Solaris utilities provide a summary of the time spent waiting for I/O (which is a meaningless metric), but fail to provide facilities to easily correlate I/O activity with a process. With the introduction of psio.pl, you can see exactly which processes are responsible for generating I/O:

psio.pl

$ psio.pl UID PID PPID %I/O STIME TTY TIME CMD root 7312 7309 70.6 16:00:59 pts/2 02:36 dd if=/dev/dsk/c1t1d0s2 of=/dev/null bs=1048576 root 0 0 0.0 10:24:18 ? 00:02 sched root 1 0 0.0 10:24:18 ? 00:03 /sbin/init root 2 0 0.0 10:24:18 ? 00:00 pageout root 3 0 0.0 10:24:18 ? 00:51 fsflush root 7 1 0.0 10:24:20 ? 00:06 /lib/svc/bin/svc.startd root 9 1 0.0 10:24:21 ? 00:14 /lib/svc/bin/svc.configd

Once you find I/O intensive processes, you can use fspaging, iosnoop, and rwsnoop to get additional information:

iosnoop

$ iosnoop -n MAJ MIN UID PID D BLOCK SIZE COMM PATHNAME 136 8 0 990 R 341632 8192 dtrace /lib/sparcv9/ld.so.1 136 8 0 990 R 341568 8192 dtrace /lib/sparcv9/ld.so.1 136 8 0 990 R 14218976 8192 dtrace /lib/sparcv9/libc.so.1 [ ... ]

iosnoop

$ iosnoop -e DEVICE UID PID D BLOCK SIZE COMM PATHNAME dad1 0 404 R 481712 8192 vxsvc /lib/librt.so.1 dad1 0 3 W 516320 3072 fsflush /var/adm/utmpx dad1 0 3 W 18035712 8192 fsflush /var/adm/wtmpx [ ... ]

rwsnoop

$ rwsnoop UID PID CMD D BYTES FILE 100 902 sshd R 42 /devices/pseudo/clone@0:ptm 100 902 sshd W 80 100 902 sshd R 65 /devices/pseudo/clone@0:ptm 100 902 sshd W 112 100 902 sshd R 47 /devices/pseudo/clone@0:ptm 100 902 sshd W 96 0 404 vxsvc R 1024 /etc/inet/protocols [ ... ]

Smooooooooooth!

8 I/O Sizes Per Process

As a Solaris administrator, we are often asked to identify application I/O sizes. This information can be acquired for a single process with truss(1m), or system wide with Brendan Gregg's bitesize.d D script:

bitesize.d

$ bitesize.d Sampling... Hit Ctrl-C to end.   3 fsflush   value ------------- Distribution ------------- count 512 | 0 1024 |@ 1 2048 | 0 4096 |@@ 2 8192 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 39 16384 | 0   7312 dd if=/dev/dsk/c1t1d0s2 of=/dev/null bs=1048576   value ------------- Distribution ------------- count 16 | 0 32 | 2 64 | 0 128 | 0 256 | 0 512 | 2 1024 | 0 2048 | 0 4096 | 0 8192 | 0 16384 | 0 32768 | 0 65536 | 0 131072 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 76947 262144 | 0

If only Dorothy could see this!

9 TCP Top

Snoop(1m) and ethereal are amazing utilities, and provide a slew of options to filter data. When you don't have time to wade through snoop data or download and install ethereal, you can use tcptop to get an overview of TCP activity on a system:

tcptop

$ tcptop 5 2005 Jul 19 14:09:06, load: 0.01, TCPin: 2679 Kb, TCPout: 12 Kb UID PID LADDR LPORT RADDR RPORT SIZE NAME 0 7138 192.168.1.3 44084 192.18.108.40 21 544 ftp 0 352 192.168.1.3 22 192.168.1.8 49805 1308 sshd 100 7134 192.168.1.3 44077 192.168.1.1 22 1618 ssh 0 7138 192.168.1.3 44089 24.98.83.96 51731 2877524 ftp

Now this is some serious bling!

10 Who's paging and DTrace enhanced vmstat

With Solaris 9, the "-p" option was added to vmstat to break paging activity up into "executable," "anonymous" and "filesystem" page types:

vmstat

$ vmstat -p 5 memory page executable anonymous filesystem swap free re mf fr de sr epi epo epf api apo apf fpi fpo fpf 1738152 832320 5 9 0 0 0 0 0 0 0 0 0 1 0 0 1683280 818800 0 2 0 0 0 0 0 0 0 0 0 0 0 0 1683280 818800 0 0 0 0 0 0 0 0 0 0 0 0 0 0

This was super useful information, but unfortunatley doesn't provide the executable responsible for the paging activity. With the introduction of whospaging.d, you can get paging activity per process:

whospaging.d

$ whospaging.d   Who's waiting for pagein (milliseconds): Who's on cpu (milliseconds): svc.configd 0 sendmail 0 svc.startd 0 sshd 0 nscd 1 dtrace 3 fsflush 14 dd 1581 sched 3284

Once we get the process name that is reponsible for the paging activity, we can use dvmstat to break down the types of pages the application is paging (similar to vmstat -p, but per process!):

dvmstat

$ dvmstat -p 0 re maj mf fr epi epo api apo fpi fpo sy 0 0 0 13280 0 0 0 0 0 13280 0 0 0 0 13504 0 0 0 0 0 13504 0 0 0 0 13472 0 0 0 0 0 13472 0 0 0 0 13472 0 0 0 0 0 13472 0 0 0 0 13248 0 0 0 0 0 13248 0 0 0 0 13376 0 0 0 0 0 13376 0 0 0 0 13664 0 0 0 0 0 13664 0

Once we have an idea of which pages are being paged in or out, we can use iosnoop, rwsnoop and fspaging.d to find out which files or devices the application is writing to! Since these rockin' scripts go hand in hand, I am placing them together. Shizam!

And without further ado, number 1 goes to ... (*drum roll*)

11 I/O Top

After careful thought, I decided to make iotop and rwtop #1 on my top ten list. I have long dreamed of a utility that could tell me which applications were actively generating I/O to a given file, device or file system. With the introduction of iotop and rwtop, my wish came true:

iotop

$ iotop 5   2005 Jul 19 13:33:15, load: 0.24, disk_r: 95389 Kb, disk_w: 0 Kb   UID PID PPID CMD DEVICE MAJ MIN D BYTES 0 99 1 nscd dad1 136 8 R 16384 0 7037 7033 find dad1 136 8 R 2266112 0 7036 7033 dd sd7 32 58 R 15794176 0 7036 7033 dd sd6 32 50 R 15826944 0 7036 7033 dd sd5 32 42 R 15826944 0 7036 7033 dd vxio21000 100 21000 R 47448064 <syntaxhighlight lang=text>}} {{command|rwtop|<syntaxhighlight lang=text> $ rwtop 5 2005 Jul 24 10:47:26, load: 0.18, app_r: 9 Kb, app_w: 8 Kb   UID PID PPID CMD D BYTES 100 922 920 bash R 3 100 922 920 bash W 15 100 902 899 sshd R 1223 100 926 922 ls R 1267 100 902 899 sshd W 1344 100 926 922 ls W 2742 100 920 917 sshd R 2946 100 920 917 sshd W 4819 0 404 1 vxsvc R 5120

12 References

http://brendangregg.com/
DTrace User Guide
Observing I/O Behavior with the DTraceToolkit
DTrace Toolkit
DTrace Topics

12.1 Dtrace GUI

http://www.netbeans.org/kb/docs/ide/NetBeans_DTrace_GUI_Plugin_0_4.html