MySQL 101: Monitor Disk I/O with pt-diskstats - Peter Zaitsev

http://ift.tt/1rtYmFd

Here on the Percona Support team we often ask customers to retrieve disk stats to monitor disk IO and to measure block devices iops and latency. There are a number of tools available to monitor IO on Linux. iostat is one of the popular tools and Percona Toolkit, which is free, contains the pt-diskstats tool for this purpose. The pt-diskstats tool is similar to iostat but it’s more interactive and contains extended information. pt-diskstats reports current disk activity and shows the statistics for the last second (which by default is 1 second) and will continue until interrupted. The pt-diskstats tool collects samples of /proc/diskstats.

In this post, I will share some examples about how to monitor and check to see if the IO subsystem is performing properly or if any disks are a limiting factor – all this by using the pt-diskstats tool.

pt-diskstats output consists on number of columns and in order to interpret pt-diskstats output we need to know what each column represents.

• rd_s tells about number of reads per second while wr_s represents number of writes per second.
• rd_rt and wr_rt shows average response time in milliseconds for reads & writes respectively, which is similar to iostat tool output await column but pt-diskstats shows individual response time for reads and writes at disk level. Just a note, modern iostat splits read and write latency out, but most distros don’t have the latest iostat in their systat (or equivalent) package.
• rd_mrg and wr_mrg are other two important columns in pt-diskstats output.  *_mrg is telling us how many of the original operations the IO elevator (disk scheduler) was able to merge to reduce IOPS, so *_mrg is telling us a quite important thing by letting us know that the IO scheduler was able to consolidate many or few operations. If rd_mrg/wr_mrg is high% then the IO workload is sequential on the other hand, If rd_mrg/wr_mrg is a low% then IO workload is all random. Binary logs, redo logs (aka ib_logfile*), undo log and doublewrite buffer all need sequential writes.
• qtime and stime are last two columns in pt-diskstats output where qtime reflects to time spent in disk scheduler queue i.e. average queue time before sending it to physical device and on the other hand stime is average service time which is time accumulated to process the physical device request. Note, that qtime is not discriminated between reads and writes and you can check if response time is higher for qtime than it signal towards disk scheduler. Also note that service time (stime field and svctm field in in pt-diskstats & iostat output respectively) is not reliable on Linux. If you read the iostat manual you will see it is deprecated.

Along with that, there are many other parameters for pt-diskstats – you can found full documentation here. Below is an example of pt-disktats in action. I used the  –devices-regex option which prints only device information that matches this Perl regex.

  $ pt-diskstats --devices-regex=sd --interval 5  #ts device rd_s rd_avkb rd_mb_s rd_mrg rd_cnc rd_rt wr_s wr_avkb wr_mb_s wr_mrg wr_cnc wr_rt busy in_prg io_s qtime stime  1.1 sda   21.6   22.8    0.5      45%    1.2  29.4 275.5   4.0    1.1      0%   40.0  145.1   65%   158 297.1 155.0   2.1  1.1 sdb   15.0   21.0    0.3      33%    0.1   5.2   0.0   0.0    0.0      0%    0.0    0.0   11%     1  15.0   0.5   4.7  1.1 sdc    5.6   10.0    0.1       0%    0.0   5.2   1.9   6.0    0.0     33%    0.0    2.0    3%     0   7.5   0.4   3.6  1.1 sdd    0.0    0.0    0.0       0%    0.0   0.0   0.0   0.0    0.0      0%    0.0    0.0    0%     0   0.0   0.0   0.0  5.0 sda   17.0   14  

Truncated by Planet PHP, read more at the original (another 12219 bytes)

from Planet MariaDB http://ift.tt/1lvLCrP