Textscan and cpu usage

I used textscan to read a csv file of 0.5 GB. Sometimes it took less than a minute to complete but in another time it took almost 10 minutes! When I compare the cpu usages during those instances, I noticed that the cpu usage is high for the former (25% in a quadcores machine, so a full core) and low for the latter (less than 5%). Anybody has this experience?

[per isakson]

• How much RAM do you have? What did the Task Manager (Windows) free ram show during execution.
• Did anything I/O-intensive run simultaneously?

[Ari]

I have 8GB RAM, and no I didn't run anything simultaneously.

When it reads fast then the memory usage (in the Task Manager - process - Matlab) increases rapidly and so is the CPU. When it reads slow the memory usage stays constant and so is the CPU (at low percentage). What is strange is that if I read the file first using fileread, although just as a dummy, then do the textscan (on the file and not on the string), it reads fast all the time. I came across this 'trick' by looking at what importdata does. Importdata uses also textscan, but instead of textscanning the file direclty, it reads the file into a string first (using fileread), then do the textscan on the string.

[per isakson]

• This large difference in speed makes me think about swapping, but that isn't likely with 8GB RAM. Did try to use the Resource Manager to see what's going on?
• I once tested speed of I/O with some large files. I had problems to reproduce the results. End of story: I drew the conclusion that my test "messed up" the system cache, which in turn increased the execution times, but certainly not an order of magnitude.
• I often use fileread in combination with textscan, when the string needs some fixing before parsing. I was initially surprised it is nearly as fast as reading with textscan.
• Which versions of Matlab and Windows do you use?

Answer 1

[Kian Azami]

I just heard that the computation process by the cpu is a very nonlinear process and for this reason every time you see a different behavior. There are some publication about this issue, to study the behavior of the cpu computations.

I put a youtube link which one of the prominent scientists talks about this issue. Worth to listen! https://www.youtube.com/watch?time_continue=17&v=iW2QJRDEBMw