This is not a performance tuning article. If your system is always slow, this article may not be what you are looking for. I'll be covering some general performance related issues here, but the main focus is for the system that was running fine yesterday but is sucking mud today. The typical response to such problems is "Reboot it", and while that may indeed fix the problem, it does not address the root cause, so you are likely to have the situation again.
You may also want to review General Troubleshooting
You need to figure out WHY it is slow. It would be nice if you had prepared ahead of time by having such tools as "sar" enabled for historical data or at least were familiar with what "top" and "vmstat" look like when your system is "normal"- that is, doing whatever tasks it is supposed to do, with the usual number of users, and not seeming "slow". A baseline of "normal" performance is very useful when trying to figure out what is causing abnormal performance. For instance, here's a "sar -b" from a Linux system:
Linux 2.4.9-12 (apl) 12/03/2001 07:01:00 AM tps rtps wtps bread/s bwrtn/s 07:11:00 AM 1.28 0.00 1.28 0.03 11.93 07:21:00 AM 1.34 0.10 1.25 2.53 12.37 08:01:00 AM 1.00 1.00 1.00 1.00 1.00 08:11:00 AM 32.51 28.41 4.10 235.71 64.94 08:21:00 AM 111.83 101.62 10.21 7127.43 181.95 08:31:00 AM 75.79 52.29 23.50 2348.18 625.73 Average: 400.01 335.11 64.91 20863.12 1691.67
Does the activity beginning just after 8:00 AM represent an abnormal condition for this machine? Will the system seem slow with this amount of disk activity? If the system is slow right now, is this why- is it because of disk access? Suppose I now take a "sar -b 5 10" sample and see this:
Linux 2.4.9-12 (apl) 12/03/2001 08:32:20 AM tps rtps wtps bread/s bwrtn/s 08:32:25 AM 0.00 0.00 0.00 0.00 0.00 08:32:30 AM 0.00 0.00 0.00 0.00 0.00 08:32:35 AM 0.00 0.00 0.00 0.00 0.00 08:32:40 AM 158.80 0.40 158.40 3.20 5360.40 08:32:45 AM 143.60 3.60 140.00 67.20 2299.60 08:32:50 AM 45.20 44.60 0.60 356.80 6.40 08:32:55 AM 0.60 0.00 0.60 0.00 6.40 08:33:00 AM 2.80 0.00 2.80 0.00 32.00 08:33:05 AM 32.00 20.60 11.40 164.80 1225.20 08:33:10 AM 43.20 43.00 0.20 344.00 102.40 Average: 42.62 11.22 31.40 93.60 903.24
What does that indicate about the disk activity?
Well, in fact, this wouldn't be a particularly heavy load on the hardware I ran it on, and wouldn't noticeably affect performance on the single user desktop that it is. The "sar -b 5 10" shows that whatever the disk activity is, it isn't constant. But is it "normal"? The only way you'd know that is if your were familiar with the system or could look at historical data- which is why it is important to collect that data from the very first day you put a machine into service.
Your beginning tool for that is "sar"- enable it on SCO Unix if it isn't already enabled with
Linux folks will find their stuff just a bit below this - hang tight.
Ignore the warning about rebooting- it is not necessary.
If you already know about sar, you may want to skip ahead to What's Wrong? now.
All that does is uncomment entries for sa1 and sa2 in the sys and root crontabs. By default, SCO runs sar every twenty minutes during "working hours" and ever hour otherwise; you should adjust the "sys" crontab to meet your specific needs. The daily summary (the sa2 script run from root's crontab) creates summary files in /var/adm/sa. These will have names like "sar01", "sar02", etc.; the daily data files are named "sa01" through "sa31". The daily data files are binary data. If you wanted to examine memory statistics from the 15th of the month, you'd run
sar -f /var/adm/sa15 -r
The "sar" summary files are text. The "sar15" is the output of running
sar -f /var/adm/sa15 -A > /var/adm/sar15
and therefore can be viewed directly with "more", printed directly or whatever.
On Red Hat Linux systems, sar data collection runs from /etc/cron.hourly/sysstat:
#!/bin/sh # snapshot system usage every 10 minutes six times. umask 0022 /usr/lib/sa/sa1 600 6 &
The daily summary is done by:
#!/bin/sh # generate a daily summary of process accounting. umask 0022 /usr/lib/sa/sa2 -A &
The sar files are found in /var/log/sa, and use the same naming as on SCO. If you have been running sar for a month or more, you'll always have one months worth of historical data. As daily files are overwritten by daily files from the following month, you don't have to be concerned with using up disk space. Having this historical data lets you quickly decide if the current sar statistics represent an unusual condition.
The flags for sar vary on different OS's, so read the man page. On all systems, sar without any arguments gives you cpu usage, but even there the output will vary. Linux systems have a "nice" column that SCO Unix lacks, and SCO includes a useful "wio" column not found on Linux:
Linux 2.4.9-12 (apl) 12/04/2001 05:32:15 AM CPU %user %nice %system %idle 05:32:20 AM all 0.00 0.00 0.20 99.80 05:32:25 AM all 0.00 0.00 0.00 100.00 05:32:30 AM all 4.20 0.00 1.00 94.80 05:32:35 AM all 4.40 0.00 0.60 95.00 05:32:40 AM all 0.00 0.00 0.00 100.00 Average: all 1.72 0.00 0.36 97.92 SCO_SV scosysv 3.2v5.0.6 PentIII 12/04/2001 05:48:33 %usr %sys %wio %idle (-u) 05:48:38 0 0 0 100 05:48:43 0 0 0 100 05:48:48 0 0 0 100 05:48:53 0 0 0 100 05:48:58 0 0 0 100 Average 0 0 0 100
If you run sar without any numerical arguments, it will look for today's historical data (and complain if it can't find it). If you run it with numerical arguments, it samples what is happening now. The first argument is the time between samples (5 seconds is a good choice), the second is the number of samples. So "sar 5 2" gives two samples, 5 seconds apart.
So, the system is slow. Let's try to find out why.
A quick ps might spot a problem:
ps -A -o pcpu -o pid -o comm | sort -r | head
That sorts by "percent cpu" ("top" would do the same thing)
If it is the cpu that is pegged busy, it *may* be a run away process that is eating cpu cycles. Do this:
for x in 1 2 3 4 5 do ps -e | sort -r +2 | head -5 echo "===" sleep 5 done
Look for a process who's time column has gone up by 3 to 5 seconds each time- if you have something like that, that's your problem- you need to kill it. The TIME column is time on the cpu- normally a process doesn't spend a great deal of time actually running- it's waiting for the disk, waiting for you to type something, etc. Most processes spend most of their time sleeping, waiting for something else to happen, so something that gains 3 seconds or more in 5 seconds of wall time is usually suspect.
Also see processes hogging CPU?.
If you watch it over a few minutes, the time it gains here divided by the elapsed wall clock time is the percentage of your cpu this process is taking for itself. A shortlived process can take a lot of the cpu to print, or to redraw an X screen etc., so you have to use some good judgement here. But 3 seconds out of 5 is very likely a real problem.
Of course you need to understand what you are killing: you probably wouldn't want to kill the main Oracle database, for example.
If you kill the errant process and another copy of it pops right back to the top of the list, then you need to track down its parent:
# for example, if process 15246 is the problem ps -p 15246 -o ppid
Of course, it may go further up the chain. Here's a script that traces back to init:
# This works on SCO or Linux, just pass a process ID as an argument. MYPROC=$1 NEXTPROC=$MYPROC while [ $NEXTPROC != 0 ] do ps -lp $NEXTPROC MYPROC=$NEXTPROC NEXTPROC=`ps -p $MYPROC -o "ppid=" ` done
Sometimes you'll have a badly written network program that starts sucking resources when its client dies. If you can't get the supplier to fix it, you may want to write a script to track down and kill these things. One clue that might help: the difference between a good "xyz" process and a bad one might just be whether or not it has an attached tty. So, if you see this:
5821 ? 00:00:42 xyz 6689 ttyp0 00:00:08 xyz 7654 ttyp1 00:00:12 xyz
It's probably the one with a "?" that will start accumulating time. So a script that watched for and killed those might look like this:
set -f # turn off shell expansion because of "?" ps -e | grep "xyz$" | while read line do set $line [ "$2" = "?" ] && kill -9 $1 done
If you can't do it that way, you have to get more clever, and watch for changing time:
set -f mkdir /tmp/mystuff ps -e | grep "xyz$" | while read line do set $line ps -p $1 > /tmp/mystuff/first sleep 5 #adjust sleep as necessary ps -p $1 > /tmp/mystuff/second diff /tmp/mystuff/first /tmp/mystuff/second || kill -9 $1 done
And even that may not be clever enough for your particular situation, so test and tread carefully. You may even need to do math on the time field to see what has really happened. If all else fails, you might be able to set ulimit on cpu time or some other limit that wouldn't affect a "normal" running of this process.
Another thing you may see is a process that has used a lot of time but isn't gaining time right now. I've seen that many times where the process is "deliver"- MMDF's mail delivery agent on SCO systems that aren't running sendmail. What happens is that for whatever reason (a root.lock file from a crash in /usr/spool/mail or a missing "sys" home directory), there are thousands of undelivered messages in the subdirectories of /usr/spool/mmdf/lock/home
The fix for that is simple if you don't care about the messages: rm -r all those directories and recreate them empty with the same ownership and permissions
cd /usr/spool/mmdf/lock/home /etc/rc2.d/P86mmdf stop rm -r * chown mmdf:mmdf * chmod 777 * cd /usr/spool/mail rm *.lock /etc/rc2.d/P86mmdf start
You'd then want to verify that mail is working normally and that whatever caused the problem isn't still happening- for example, if /usr/sys is missing this problem will come right back again very quickly.
Another possibility is a program that is rapidly spawning off other programs. You should be able to see that in "ps -e". First, are the number of processes growing?:
ps -e | wc -l sleep 5 ps -e | wc -l
Or, are there new processes briefly showing up at the end of the listing?:
ps -e | tail sleep 5 ps -e | tail
In either case, you need to track down the parent and kill it.
If sar -r shows low memory or (worse) swapping, go buy more memory. That's going to be easy to spot on SCO's sar, but Linux is a bit harder. Let's look at SCO first:
SCO_SV scosysv 3.2v5.0.6 PentIII 12/06/2001 11:37:06 freemem freeswp availrmem availsmem (-r) 11:37:06 unix restarts 11:40:00 52972 786000 56222 150408 12:00:00 52903 786000 56234 150643 12:20:00 52996 786000 56240 150723 12:40:00 53018 786000 56240 150723 13:00:00 53018 786000 56240 150723 13:20:00 53018 786000 56240 150723 13:40:00 53018 786000 56240 150723 14:00:00 52885 786000 56231 150606 14:20:00 52999 786000 56240 150723 14:40:00 53016 786000 56240 150723 15:00:00 53018 786000 56240 150723
This machine consistently has over 200 MB of memory available- unused ( freemem pages are 4K each). Obviously no problem there, and in fact, if this is always the case (which you'd know from sar historical data), you may want to use some of that memory for disk buffers- see Adding RAM to SCO Unix.
Linux looks much different:
Linux 2.4.9-12 (apl) 12/06/2001 03:19:25 PM kbmemfree kbmemused %memused kbmemshrd kbbuffers kbcached kbswpfree kbswpused %swpused 03:19:30 PM 4864 250080 98.09 700 3920 77136 442192 87912 16.58 03:19:35 PM 4864 250080 98.09 700 3920 77140 442192 87912 16.58 03:19:40 PM 4864 250080 98.09 700 3920 77140 442192 87912 16.58 03:19:45 PM 4864 250080 98.09 700 3920 77140 442192 87912 16.58 03:19:50 PM 4864 250080 98.09 700 3920 77140 442192 87912 16.58 Average: 4864 250080 98.09 700 3920 77139 442192 87912 16.58
As is immediately obvious, all memory is in use. The reason is that Linux always uses "unused" memory for the file system buffer cache. So, roughly 75 MB has been put to work for that (kbcached column).
Another way to look at Linux memory right this moment is to
But what's using it? Well, ps -el will show you in the SZ column how much memory each process is using. So
ps -el | sort -r +9 | head
can tell you a lot, particularly if you see that a process is gaining memory over time.
Remember, this article is not about performance tuning- it's about specific performance degradation. However, it's almost always true that the disk drives are the biggest performance bottleneck. See Raid for a more general discussion of improving disk performance.
On SCO systems, you can get a good overview of disk performance from "sar -d":
# sar -d 10 5 SCO_SV scosysv 3.2v5.0.6 PentIII 12/08/2001 07:57:52 device %busy avque r+w/s blks/s avwait avserv (-d) 07:58:02 wd-0 77.42 1.00 483.92 967.83 0.00 1.60 07:58:12 wd-0 45.45 1.00 242.36 486.11 0.01 1.88 07:58:22 wd-0 6.99 83.30 44.26 529.87 130.05 1.58 #
The first thing you are interested in is %busy. The "avwait" column is the average amount of time that processes wait for the disk to give them their data, so that and "avque" (how many processes are trying to use the disk) can give you a clear picture of load. The "avserve" is a measure of the disks ability to deliver that data, and isn't going to change much for the same hardware.
Note that if %busy is small and avque and avwait are high, you are probably seeing buffer cache flushes. Those can affect performance, and there are tunables that affect how often and how much is flushed, but those issues aren't the focus of this article.
Linux sar doesn't have "-d" (or at least it doesn't on Red Hat 7.2), so the next best thing is iostat:
iostat -d 5 2 Linux 2.4.9-12 (apl) 12/08/2001 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn dev8-0 8.18 77.13 25.77 625890 209094 dev8-1 0.31 2.26 0.38 18376 3064 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn dev8-0 10.80 84.80 1.60 424 8 dev8-1 0.00 0.00 0.00 0 0
The "tps" is the number of transfers per second the disk was asked for. The block read and write columns give you both the number of blocks and the blocks per second.
For a sudden degradation of performance, you are interested only to see if the disk is more busy than is normal. You might also be looking at the blocks per second to compare how much data is being moved around. Of course, if you decide that the disk activity is unusual, your next problem is to find out why: you still have to track down the process that is doing this.
The first place I'd look in such a situation is log files: /var/adm on SCO, or /var/log on Linux. If there's nothing there, and you see large amounts of writes, accompanied by decreasing disk space (df -v), then you are looking for a growing file and the process that is writing it. Finding that can be fairly easy:
cd / du -s * > /tmp/a sleep 30 du -s * > /tmp/b diff /tmp/a /tmp/b
The directory that is growing will pop out of the diff. Change into that directory and repeat the procedure until you find the file that is growing, and finally use "fuser" to identify the process that is writing the file.
fuser -k /tmp/growingfile
A sudden problem with network performance is almost always going to be hardware, but there are other possibilities. For example, are your routes what they should be? If "netstat -rn" shows different routing than what you expect, do you have routed running and some router is giving you bad information? Kill routed and reset your routes.
"netstat -in" will give you an idea of collisions; a bad card somewhere in your network can cause this- if you run switches rather than hubs you won't have collisions at all, but that bad card could still be affecting performance. Know what "normal" network traffic looks like, know how long "ping" response times should be on your WAN, etc.
Mis-negotiation of network speeds and cheap nic cards are another source of network problems:
An interesting mail related problem I had recently is described at /SME/largenewsletter.html
Disk CPU and Network are the most common areas that will cause a sudden performance drop. However, there are other things that can happen. The general rule is that if you have historical data, you'll be able to spot the problem much more quickly. Some of the other things I'd look at if everything above came up blank include "sar -y" (tty activity) and "sar -q" (run queue size)- these may not be available on Linux.
Some of these may be of interest:
Got something to add? Send me email.
More Articles by Tony Lawrence © 2011-05-16 Tony Lawrence