iotop | A top utility for IO

When I use sysbench to test mysql, I use iotop to monitor io and I find only have DiSH WRITE speed, the DISK READ speed is always 0. Then I use free -h and I find that buffer/cache increase, does it mean that sysbench's test data is not write in disk but in buffer and no auto update into disk?

Thank you so much!

The problem is that I am running no performance intensive program but the system load continues to be high. Actually the normal load is below 0.05, but from yesterday it has been always been higher than 1.5. After some time digging in the reason, I think that it is the jbd2/sda2-8's io usage that caused the problem.

Later I went to room where the PC locates and found out that the HDD LED light keeps flashing, maybe lots of times in a second. It means that io usage is really a problem.

Here, https://www.webhostingtalk.com/showthread.php?t=1148545, it tells me that jbd2 is not the root cause and I must found out which program is really writing or reading the disk. So I found out that the real cause is snapd.

I tried temporarily stopping the snapd service, and the load went down immediately.

I am working on a python script which needs to read the output of iotop. But stdout appears to be blank. The goal is to have the output of iotop stored as a string which I can then later search and use. This is running on Ubuntu 16.04 w/ python 3.6. I am calling the script with "sudo python3.6 script.py" as iotop requires admin rights.

Lets say I have an iotop view like this:

How to find the SELECT which is hitting the hard disk so heavy here?

5GB database takes 15 minutes to restore. CPU is utilized 10% on one thread, iotop shows 20MB/s or less. Can't understand why it is so slow and does not use all HW resources?

I have a table with around 270,000,000 rows and this is how I created it.

I have directories containing 100K - 1 million images. I'm going to create a hash for each image so that I can, in the future, find an exact match based on these hashes. My current approach is:

I am maintaing the code for a Go project that reads and writes a lot of data and that has done so successfully for some time. Recently, I made a change: a CSV file with about 2 million records is loaded in a map with struct values at the beginning of the program. This map is only used in part B, but first part A is executed. And this first part already runs noticeably slower than before (processing time is quadrupled). That is very strange since that part of the logic did not change. I have spent a week trying to explain how this can happen. Here are the steps I have taken (when I mention performance, I always refer to part A, which does not include the time to load the data in memory and actually has nothing to do with it):

• The program was running on a server inside a Docker container. But I have been able to reproduce it on my laptop without container: the performance indeed decreases compared to when I run it without the data from the file loaded in memory.
• The server had a huge amount of RAM. Although obviously more memory is used when the file is loaded, no limits are hit. I also did not see spikes or other strange patterns in memory usage and disk I/O. For these checks, I have used pprof, htop and iotop.
• When the data is loaded but then the map set to nil, performance is OK again.
• Loading the data in a slice instead of a map reduces the performance decrease from x4 to x2 (but the memory usage is more or less the same as with the map).
• This made me wonder whether the map/slice is accessed somewhere in part A, even though it shouldn’t. The map is stored in a field of a struct type. I checked and this struct is always passed by pointer (including all goroutines). Making it a global variable instead of a pointer field did not solve the issue.
• There is one dependency outside of the standard library. Is the problem caused by the library? It forces some garbage collects. Disabling this does not make a difference. I found another similar library that is unrelated and using this one as a replacement improves performance, but it still takes longer when the data of the file is loaded.

Here I have plotted the metrics with and without the data in memory:

What could cause this effect or how do I find it out?

I finded how to get summary value grouping column by PID: