0

I run simulations on a public computation server. Due to the nature of the simulation its CPU usage oscillates in a wide range (0-60%). Does this affect the speed of other people's work. The overall CPU usage doesn't exceeds 90%.

The question is how oscillation in CPU usage by one task can affect speed of other tasks where the overall CPU usage always remains less than 100% i.e. CPU is not overloaded.

The picture attached shows the oscillatory use of CPU I am talking about

enter image description here

4
  • That is interesting of you to tell us. However you seem to have no question...
    – Hennes
    May 6, 2015 at 16:36
  • I think this has something to do with architecture of CPU and the way Windows server manages the tasks. Unfortunately I don't have much information on this area. The question is how oscillation in CPU usage by one task can affect speed of other tasks where the overall CPU usage always remains less than 100% i.e. CPU is not overloaded.
    – Kevin Bell
    May 6, 2015 at 17:01
  • In that case rewrite the question using the eit link. The answer is likely contain things like 'fewer cache hits', delays due to context switches, less spare thermal budget, ...
    – Hennes
    May 6, 2015 at 17:02
  • @KevinBell: What program are you using for the simulations? It probably has an option to set how many threads to use, in this case you could limit it to say, 4 threads which would mean that you would have a reduced impact on the overall server load.
    – James P
    May 11, 2015 at 10:54

1 Answer 1

1

Since nobody seems to be answering this: Yes, it will slow down the other processes.

From a CPU/computing standpoint:

The CPU's might not have more work per second than they can do per second (aka load), but they do get less efficient when they have to do different tasks with different data sets. Keywords to research in this topic are context switches and cache hits.

To use something equivalent: Think of the CPU as a secretary which you give work. (s)he will work more efficient if she has single job and works on that for an hour compared to trying to do 60 jobs each lasting a minutes. Part of this is switching from one job (context) to another job, which takes time.

Another part is the cache on the CPU. It keeps a local copy of the data it works with. This is done because memory access is relative slow. As soon as you switch tasks you start working on a new set of data. This means fetching new info. And since you have limited space in the cache that means throwing old data out. And once you switch back this happens again. And again...

Then on modern CPUs there is a thermal budget. A CPU can run at regular max. speed all the time. It will get hot doing so, but heat produced and heat dissipated should stay in balance. If the CPU has less work it can cool down. This effectively give it a small heat buffer. This buffer is used with what Intel and AMD now call turbo. When The CPU is relative cold and it has a lot of work the CPU increases its clockspeed and works faster. It can not sustain that for long, but a short but intense task on a cold CPU (with spare thermal budget) will briefly run faster than on a CPU which already spent its thermal budget.

Memory:

From a memery standpoint: You application will use at least some memory. That is less memory available to other tasks (such as IO buffers). This will slow the system down.

IO:

If your application maxes out the I/O (e..g disk access) then it does not even matter if it slows down the CPU. If every other program has to wait in a queue for disk access then you can slow the system down even without exceeding 100% CPU load.


Summary: Yes, it is very likely that a continuous busy application will slow down the system. Just how much can vary from barely perceptiable up to a significant slowdown.

1
  • Thanks for the explanation, I have also attached a picture showing the oscillatory usage of CPU I am talking about.
    – Kevin Bell
    May 11, 2015 at 10:46

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .