Analysis for IO bottleneck in Unix

by

in

# sar -d

Linux 2.4.21-27.ELsmp (pw101) 12/04/2005

12:00:00 AM DEV tps rd_sec/s wr_sec/s

….

Average: dev8-128 7.16 5.37 75.07

Average: dev8-129 7.16 5.37 75.07

Average: dev8-130 0.00 0.00 0.00

….

The above command finds the busiest device. To determine what that device is, do:

# more /proc/devices

Character devices:

1 mem

2 pty

3 ttyp

4 ttyS

5 cua

7 vcs

Block devices:

1 ramdisk

2 fd

3 ide0

7 loop

8 sd

71 sd

129 sd

The above example indicates the ‘Block device’ 8 is ‘sd’ or scsi disk, and #129 is one of

scsi disks in the system. So ‘sar -d’ tells you that one of the scsi disks, or device

‘dev8-129’ is busy.

Now the following command is a bit busy, but you can always strip away piped commands to see

what output is like for each command. But the shown example give you which program (along with

its pid) has the most open files on a given busy device. In our example, the busy device we

are interested is ‘dev8-129’, but is represented as ‘8,129’ in the ‘lsof’ output.

# lsof | grep “8,129” | awk ‘{print $1″ “$2}’ | uniq -c | sort -n -r

[count p pid]

307 java 31916

307 java 31915

307 java 31914

307 java 31913

307 java 31912

307 java 31911

307 java 31910

307 java 31909

307 java 31908

307 java 31907

307 java 31906

307 java 31905

307 java 31904

307 java 31903

307 java 27645

55 db2sysc 32011

51 db2sysc 32012

46 httpd 32009

46 httpd 32008

46 httpd 32006

46 httpd 31901

46 httpd 31900

46 httpd 31899

46 httpd 31898

46 httpd 31897

46 httpd 31874

45 db2sysc 32019

40 db2fmp 31996

31 sshd 26713

29 db2sysc 32025

29 db2sysc 32024

The above output gives a clue as to which process(es) to look into for disk I/O problems

(i.e., ‘java’). The ‘java’ processes are having the largest number of open files, and

chances are doing the most disk I/O. Of course, some files maybe memory-cached, so this

method may not always work. I’d like to hear from you if you have a better solution. Please

contact me (see footer for contact info).


Comments

Leave a Reply

Your email address will not be published. Required fields are marked *