Posts Tagged ‘overhead’

Tracking I/O hard disk server bottlenecks with iostat on GNU / Linux and FreeBSD

Tuesday, March 27th, 2012

Reading Time: 7minutes

Hard disk overhead tracking on Linux and FreeBSD with iostat

I've earlier wrote an article How to find which processes are causing hard disk i/o overhead on Linux there I explained very rawly few tools which can be used to benchmark hard disk read / write operations. My prior article accent was on iotop and dstat and it just mentioned of iostat. Therefore I've wrote this short article in attempt to explain a bit more thoroughfully on how iostat can be used to track problems with excessive server I/O read/writes.

Here is the command man page description;
iostatReport Central Processing Unit (CPU) statistics and input/output statistics for devices, partitions and network filesystems

I will further proceed with few words on how iostat can be installed on various Linux distros, then point at few most common scenarious of use and a short explanation on the meaning of each of the command outputs.

1. Installing iostat on Linux

iostat is a swiss army knife of finding a server hard disk bottlenecks. Though it is a must have tool in the admin outfut, most of Linux distributions will not have iostat installed by default.
To have it on your server, you will need to install sysstat package:

a) On Debian / Ubuntu and other Debian GNU / Linux derivatives to install sysstat:

debian:~# apt-get --yes install sysstat

b) On Fedora, CentOS, RHEL etc. install is with yum:

[root@centos ~]# yum -y install sysstat

c) On Slackware Linux sysstat package which contains iostat is installed by default. 

d) In FreeBSD, there is no need for installation of any external package as iostat is part of the BSD world (bundle commands).
I should mention bsd iostat and Linux's iostat commands are not the same and hence there use to track down hard disk bottlenecks differs a bit, however the general logic of use is very similar as with most tools in BSD and Linux.

2. Checking a server hard disk for i/o disk bottlenecks on G* / Linux

Once having the sysstat installed on G* / Linux systems, the iostat command will be added in /usr/bin/iostat
a) To check what is the hard disk read writes per second (in megabytes) use:

debian:~# /usr/bin/iostat -m
Linux 2.6.32-5-amd64 (debian) 03/27/2012 _x86_64_ (8 CPU)
avg-cpu: %user %nice %system %iowait %steal %idle
15.34 0.36 2.76 2.66 0.00 78.88
Device: tps MB_read/s MB_wrtn/s MB_read MB_wrtn
sda 63.89 0.48 8.20 6730223 115541235
sdb 64.12 0.44 8.23 6244683 116039483
md0 2118.70 0.22 8.19 3041643 115528074

In the above output the server, where I issue the command is using sda and sdb configured in software RAID 1 array visible in the output as (md0)

The output of iostat should already be easily to read, for anyone who didn't used the tool here is a few lines explanation of the columns:

The %user 15.34 meaning is that 15.34 out of 100% possible i/o load is generad by system level read/write operations.
%nice – >Show the percentage of CPU utilization that occurred while executing at the user level with nice priority.
%iowait – just like the top command idle it shows the idle time when the system didn't have an outstanding disk I/O requests.
%steal – show percentage in time spent in time wait of CPU or virtual CPUs to service another virtual processor (high numbers of disk is sure sign for i/o problem).
%idle – almost the same as meaning to %iowait
tps – HDD transactions per second
MB_read/s (column) – shows the actual Disk reads in Mbytes at the time of issuing iostat
MB_wrtn/s – displays the writes p/s at the time of iostat invocation
MB_read – shows the hard disk read operations in megabytes, since the server boot 'till moment of invocation of iostat
MB_wrtn – gives the number of Megabytes written on HDD since the last server boot filesystem mount

The reason why the Read / Write values for sda and sdb are similar in this example output is because my disks are configured in software RAID1 (mirror)

The above iostat output reveals in my specific case the server is experiencing mostly Disk writes (observable in the high MB_wrtn/s 8.19md0 in the above sample output).

It also reveals, the I/O reads experienced on that server hard disk are mostly generated as a system (user level load) – see (%user 15.34 and md0 2118.70).

For all those not familiar with system also called user / level load, this is all kind of load which is generated by running programs on the server – (any kind of load not generated by the Linux kernel or loaded kernel modules).

b) To periodically keep an eye on HDD i/o operations with iostat, there are two ways:

– Use watch in conjunction with iostat;

[root@centos ~]# watch "/usr/bin/iostat -m"
Every 2.0s: iostat -m Tue Mar 27 11:00:30 2012
Linux 2.6.32-5-amd64 (centos) 03/27/2012 _x86_64_ (8 CPU)
avg-cpu: %user %nice %system %iowait %steal %idle
15.34 0.36 2.76 2.66 0.00 78.88
Device: tps MB_read/s MB_wrtn/s MB_read MB_wrtn
sda 63.89 0.48 8.20 6730255 115574152
sdb 64.12 0.44 8.23 6244718 116072400
md0 2118.94 0.22 8.20 3041710 115560990
Device: tps MB_read/s MB_wrtn/s MB_read MB_wrtn
sda 55.00 0.01 25.75 0 51
sdb 52.50 0.00 24.75 0 49
md0 34661.00 0.01 135.38 0 270

Even though watch use and -d might appear like identical, they're not watch does refresh the screen, executing instruction similar to the clear command which clears screen on every 2 seconds, so the output looks like the top command refresh, while passing the -d 2 will output the iostat command output on every 2 secs in a row so all the data is visualized on the screen. Hence -d 2 in cases, where more thorough debug is necessery is better. However for a quick routine view watch + iostat is great too.

c) Outputting extra information for HDD input/output operations;

root@debian:~# iostat -x
Linux 2.6.32-5-amd64 (debian) 03/27/2012 _x86_64_ (8 CPU)
avg-cpu: %user %nice %system %iowait %steal %idle
15.34 0.36 2.76 2.66 0.00 78.88
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
sda 4.22 2047.33 12.01 51.88 977.44 16785.96 278.03 0.28 4.35 3.87 24.72
sdb 3.80 2047.61 11.97 52.15 906.93 16858.32 277.05 0.03 5.25 3.87 24.84
md0 0.00 0.00 20.72 2098.28 441.75 16784.05 8.13 0.00 0.00 0.00 0.00

This command will output extended useful Hard Disk info like;
r/s – number of read requests issued per second
w/s – number of write requests issued per second
rsec/s – numbers of sector reads per second
b>wsec/s – number of sectors wrote per second
etc. etc.

Most of ppl will never need to use this, but it is good to know it exists.

3. Tracking read / write (i/o) hard disk bottlenecks on FreeBSD

BSD's iostat is a bit different in terms of output and arguments.

a) Here is most basic use:

freebsd# /usr/sbin/iostat
tty ad0 cpu
tin tout KB/t tps MB/s us ni sy in id
1 561 45.18 44 1.95 14 0 5 0 82

b) Periodic watch of hdd i/o operations;

freebsd# iostat -c 10
tty ad0 cpu
tin tout KB/t tps MB/s us ni sy in id
1 562 45.19 44 1.95 14 0 5 0 82
0 307 51.96 113 5.73 44 0 24 0 32
0 234 58.12 98 5.56 16 0 7 0 77
0 43 0.00 0 0.00 1 0 0 0 99
0 485 0.00 0 0.00 2 0 0 0 98
0 43 0.00 0 0.00 0 0 1 0 99
0 43 0.00 0 0.00 0 0 0 0 100
...

As you see in the output, there is information like in the columns tty, tin, tout which is a bit hard to comprehend.
Thanksfully the tool has an option to print out only more essential i/o information:

freebsd# iostat -d -c 10
ad0
KB/t tps MB/s
45.19 44 1.95
58.12 97 5.52
54.81 108 5.78
0.00 0 0.00
0.00 0 0.00
0.00 0 0.00
20.48 25 0.50

The output info is quite self-explanatory.

Displaying a number of iostat values for hard disk reads can be also achieved by omitting -c option with:

freebsd# iostat -d 1 10
...

Tracking a specific hard disk partiotion with iostat is done with:

freebsd# iostat -n /dev/ad0s1a
tty cpu
tin tout us ni sy in id
1 577 14 0 5 0 81
c) Getting Hard disk read/write information with gstat

gstat is a FreeBSD tool to print statistics for GEOM disks. Its default behaviour is to refresh the screen in a similar fashion like top command, so its great for people who would like to periodically check all attached system hard disk and storage devices:

freebsd# gstat
dT: 1.002s w: 1.000s
L(q) ops/s r/s kBps ms/r w/s kBps ms/w %busy Name
0 10 0 0 0.0 10 260 2.6 15.6| ad0
0 10 0 0 0.0 10 260 2.6 11.4| ad0s1
0 10 0 0 0.0 10 260 2.8 12.5| ad0s1a
0 0 0 0 0.0 0 0 0.0 20.0| ad0s1b
0 0 0 0 0.0 0 0 0.0 0.0| ad0s1c
0 0 0 0 0.0 0 0 0.0 0.0| ad0s1d
0 0 0 0 0.0 0 0 0.0 0.0| ad0s1e
0 0 0 0 0.0 0 0 0.0 0.0| acd0

It even has colors if your tty supports colors 🙂

Another useful tool in debugging the culprit of excessive hdd I/O operations is procstat command:

Here is a sample procstat run to track (httpd) one of my processes imposing i/o hdd load:

freebsd# procstat -f 50404
PID COMM FD T V FLAGS REF OFFSET PRO NAME
50404 httpd cwd v d -------- - - - /
50404 httpd root v d -------- - - - /
50404 httpd 0 v c r------- 56 0 - -
50404 httpd 1 v c -w------ 56 0 - -
50404 httpd 2 v r -wa----- 56 75581 - /var/log/httpd-error.log
50404 httpd 3 s - rw------ 105 0 TCP ::.80 ::.0
50404 httpd 4 p - rw---n-- 56 0 - -
50404 httpd 5 p - rw------ 56 0 - -
50404 httpd 6 v r -wa----- 56 25161132 - /var/log/httpd-access.log
50404 httpd 7 v r rw------ 56 0 - /tmp/apr8QUOUW
50404 httpd 8 v r -w------ 56 0 - /var/run/accept.lock.49588
50404 httpd 9 v r -w------ 1 0 - /var/run/accept.lock.49588
50404 httpd 10 v r -w------ 1 0 - /tmp/apr8QUOUW
50404 httpd 11 ? - -------- 2 0 - -

Btw fstat is sometimes helpful in identifying the number of open files and trying to estimate which ones are putting the hdd load.
Hope this info helps someone. If you know better ways to track hdd excessive loads on Linux / BSD pls share 'em pls.
 

diskinfo Linux hdparm FreeBSD equivalent command for disk info and benchmarking

Thursday, March 8th, 2012

Reading Time: 3minutes

FreeBSD Linux hdparm equivalent is diskinfo artistic logo

On Linux there is the hdparm tool for various hard disk benchmarking and extraction of hard disk operations info.
As the Linux manual states hdparmget/set SATA/IDE device parameters

Most Linux users should already know it and might wonder if there is hdparm port or equivalent for FreeBSD, the aim of this short post is to shed some light on that.

The typical use of hdparm is like this:

linux:~# hdparm -t /dev/sda8

/dev/sda8:
Timing buffered disk reads: 76 MB in 3.03 seconds = 25.12 MB/sec
linux:~# hdparm -T /dev/sda8
/dev/sda8:
Timing cached reads: 1618 MB in 2.00 seconds = 809.49 MB/sec

The above output here is from my notebook Lenovo R61i.
If you're looking for alternative command to hdparm you should know in FreeBSD / OpenBSD / NetBSD, there is no exact hdparm equivalent command.
The somehow similar hdparm equivallent command for BSDs (FreeBSD etc.) is:
diskinfo

diskinfo is not so feature rich as linux's hdparm. It is just a simple command to show basic information for hard disk operations without no possibility to tune any hdd I/O and seek operations.
All diskinfo does is to show statistics for a hard drive seek times I/O overheads. The command takes only 3 arguments.

The most basic and classical use of the command is:

freebsd# diskinfo -t /dev/ad0s1a
/dev/ad0s1a
512 # sectorsize
20971520000 # mediasize in bytes (20G)
40960000 # mediasize in sectors
40634 # Cylinders according to firmware.
16 # Heads according to firmware.
63 # Sectors according to firmware.
ad:4JV48BXJs0s0 # Disk ident.

Seek times:
Full stroke: 250 iter in 3.272735 sec = 13.091 msec
Half stroke: 250 iter in 3.507849 sec = 14.031 msec
Quarter stroke: 500 iter in 9.705555 sec = 19.411 msec
Short forward: 400 iter in 2.605652 sec = 6.514 msec
Short backward: 400 iter in 4.333490 sec = 10.834 msec
Seq outer: 2048 iter in 1.150611 sec = 0.562 msec
Seq inner: 2048 iter in 0.215104 sec = 0.105 msec

Transfer rates:
outside: 102400 kbytes in 3.056943 sec = 33498 kbytes/sec
middle: 102400 kbytes in 2.696326 sec = 37978 kbytes/sec
inside: 102400 kbytes in 3.178711 sec = 32214 kbytes/sec

Another common use of diskinfo is to measure hdd I/O command overheads with -c argument:

freebsd# diskinfo -c /dev/ad0s1e
/dev/ad0s1e
512 # sectorsize
39112312320 # mediasize in bytes (36G)
76391235 # mediasize in sectors
75784 # Cylinders according to firmware.
16 # Heads according to firmware.
63 # Sectors according to firmware.
ad:4JV48BXJs0s4 # Disk ident.

I/O command overhead:
time to read 10MB block 1.828021 sec = 0.089 msec/sector
time to read 20480 sectors 4.435214 sec = 0.217 msec/sector
calculated command overhead = 0.127 msec/sector

Above diskinfo output is from my FreeBSD home router.

As you can see, the time to read 10MB block on my hard drive is 1.828021 (which is very high number),
this is a sign the hard disk experience too many read/writes and therefore needs to be shortly replaced with newer faster one.
diskinfo is part of the basis bsd install (bsd world). So it can be used without installing any bsd ports or binary packages.

For the purpose of stress testing hdd, or just some more detailed benchmarking on FreeBSD there are plenty of other tools as well.
Just to name a few:
 

  • rawio – obsolete in FreeBSD 7.x version branch (not available in BSD 7.2 and higher)
  • iozone, iozone21 – Tools to test the speed of sequential I/O to files
  • bonnie++ – benchmark tool capable of performing number of simple fs tests
  • bonnie – predecessor filesystem benchmark tool to bonnie++
  • raidtest – test performance of storage devices
  • mdtest – Software to test metadata performance on filsystems
  • filebench – tool for micro-benchmarking storage subsystems

Linux hdparm allows also changing / setting various hdd ATA and SATA settings. Similarly, to set and change ATA / SATA settings on FreeBSD there is the:

  • ataidle

tool.

As of time of writting ataidle is in port path /usr/ports/sysutils/ataidle/

To check it out install it as usual from the port location:

FreeBSD also has also the spindown port – a small program for handling automated spinning down ofSCSI harddrive
spindown is useful in setting values to SATA drives which has problems with properly controlling HDD power management.

To keep constant track on hard disk operations and preliminary warning in case of failing hard disks on FreeBSD there is also smartd service, just like in Linux.
smartd enables you to to control and monitor storage systems using the Self-Monitoring, Analysisand Reporting Technology System (S.M.A.R.T.) built into most modern ATA and SCSI hard disks.
smartd and smartctl are installable via the port /usr/ports/sysutils/smartmontools.

To install and use smartd, ataidle and spindown run:

freebsd# cd /usr/ports/sysutils/smartmontools
freebsd# make && make install clean
freebsd# cd /usr/ports/sysutils/ataidle/
freebsd# make && make install clean
freebsd# cd /usr/ports/sysutils/spindown/
freebsd# make && make install clean

Check each one's manual for more info.

How to find and kill Abusers on OpenVZ Linux hosted Virtual Machines (Few bash scripts to protect OpenVZ CentOS server from script kiddies and easify the daily admin job)

Friday, July 22nd, 2011

Reading Time: 2minutes
OpenVZ Logo - Anti Denial Of Service shell scripts

These days, I’m managing a number of OpenVZ Virtual Machine host servers. Therefore constantly I’m facing a lot of problems with users who run shit scripts inside their Linux Virtual Machines.

Commonly user Virtual Servers are used as a launchpad to attack hosts do illegal hacking activities or simply DDoS a host..
The virtual machines users (which by the way run on top of the CentOS OpenVZ Linux) are used to launch a Denial service scripts like kaiten.pl, trinoo, shaft, tfn etc.

As a consequence of their malicious activities, oftenly the Data Centers which colocates the servers are either null routing our server IPs until we suspend the Abusive users, or the servers go simply down because of a server overload or a kernel bug hit as a result of the heavy TCP/IP network traffic or CPU/mem overhead.

Therefore to mitigate this abusive attacks, I’ve written few bash shell scripts which, saves us a lot of manual check ups and prevents in most cases abusers to run the common DoS and “hacking” script shits which are now in the wild.

The first script I’ve written is kill_abusers.sh , what the script does is to automatically look up for a number of listed processes and kills them while logging in /var/log/abusers.log about the abusive VM user procs names killed.

I’ve set this script to run 4 times an hour and it currently saves us a lot of nerves and useless ticket communication with Data Centers (DCs), not to mention that reboot requests (about hanged up servers) has reduced significantly.
Therefore though the scripts simplicity it in general makes the servers run a way more stable than before.

Here is OpenVZ kill/suspend Abusers procs scriptkill_abusers.sh ready for download

Another script which later on, I’ve written is doing something similar and still different, it does scan the server hard disk using locate and find commands and tries to identify users which has script kiddies programs in their Virtual machines and therefore are most probably crackers.
The scripts looks up for abusive network scanners, DoS scripts, metasploit framework, ircds etc.

After it registers through scanning the server hdd, it lists only files which are preliminary set in the script to be dangerous, and therefore there execution inside the user VM should not be.

search_for_abusers.sh then logs in a files it’s activity as well as the OpenVZ virtual machines user IDs who owns hack related files. Right after it uses nail mailing command to send email to a specified admin email and reports the possible abusers whose VM accounts might need to either be deleted or suspended.

search_for_abusers can be download here

Honestly I truly liked my search_for_abusers.sh script as it became quite nice and I coded it quite quickly.
I’m intending now to put the Search for abusers script on a cronjob on the servers to check periodically and report the IDs of OpenVZ VM Users which are trying illegal activities on the servers.

I guess now our beloved Virtual Machine user script kiddies are in a real trouble ;P
 

How to find out which processes are causing a hard disk I/O overhead in GNU/Linux

Wednesday, September 28th, 2011

Reading Time: 3minutes
iotop monitor hard disk io bottlenecks linux
To find out which programs are causing the most read/write overhead on a Linux server one can use iotop

Here is the description of iotop – simple top-like I/O monitor, taken from its manpage.

iotop does precisely the same as the classic linux top but for hard disk IN/OUT operations.

To check the overhead caused by some daemon on the system or some random processes launching iotop without any arguments is enough;

debian:~# iotop

The main overview of iostat statistics, are the:

Total DISK READ: xx.xx MB/s | Total DISK WRITE: xx.xx K/s
If launching iotop, shows a huge numbers and the server is facing performance drop downs, its a symptom for hdd i/o overheads.
iotop is available for Debian and Ubuntu as a standard package part of the distros repositories. On RHEL based Linuxes unfortunately, its not available as RPM.

While talking about keeping an eye on hard disk utilization and disk i/o’s as bottleneck and a possible pitfall to cause a server performance down, it’s worthy to mention about another really great tool, which I use on every single server I administrate. For all those unfamiliar I’m talking about dstat

dstat is a – versatile tool for generating system resource statistics as the description on top of the manual states. dstat is great for people who want to have iostat, vmstat and ifstat in one single program.
dstat is nowdays available on most Linux distributions ready to be installed from the respective distro package manager. I’ve used it and I can confirm tt is installable via a deb/rpm package on Fedora, CentOS, Debian and Ubuntu linuces.

Here is how the tool in action looks like:

dstat Linux hdd load stats screenshot

The most interesting things from all the dstat cmd output are read, writ and recv, send , they give a good general overview on hard drive performance and if tracked can reveal if the hdd disk/writes are a bottleneck to create server performance issues.
Another handy tool in tracking hdd i/o problems is iostat its a tool however more suitable for the hard core admins as the tool statistics output is not easily readable.

In case if you need to periodically grasp data about disks read/write operations you will definitely want to look at collectl i/o benchmarking tool .Unfortunately collect is not included as a packaget for most linux distributions except in Fedora. Besides its capabilities to report on servers disk usage, collect is also capable to show brief stats on cpu, network.

Collectl looks really promosing and even seems to be in active development the latest tool release is from May 2011. It even supports NVidia’s GPU monitoring 😉 In short what collectl does is very similar to sysstat which by the way also has some possibilities to track disk reads in time.  collectl’s website praises the tool, much and says that in most machines the extra load the tool would add to a system to generate reports on cpu, disk and disk io is < 0.1%.  I couldn’t find any data online on how much sysstat (sar) extra loads a system. It will be interesting if some of someone concluded some testing and can tell which of the two puts less load on a system.