How to get the average/max process scheduling latency
Hi, I would like to retrieve the average latency between 2 schedulings of an application. The purpose is to model the service time of the application (ie. How fast it can empty the TCP buffer). I could not find an answer yet. I've found that on latest kernel latency is on average around 3us with a max at 64us (I consider a moderate load). Is that correct ? Regards Matt
On Mon, 16 Jun 2014 18:23:56 +0200, Teto said:
I could not find an answer yet. I've found that on latest kernel latency is on average around 3us with a max at 64us (I consider a moderate load). Is that correct ?
Depends on your hardware and system load. I'd hardly expect the same numbers from a busy Beaglebone or Raspberry Pi as from a idle high-end Intel i7 system.... (And things like processors that drop into sleep states when idle, with a latency hit to wake up, just add to the confusion....)
Thanks for the answer. Do you know any tool/software to retrieve the average latency for a specific program on a specific platform. This is to modelize the application rate for data processing. 2014-06-19 22:16 GMT+02:00 <Valdis.Kletnieks@vt.edu>:
On Mon, 16 Jun 2014 18:23:56 +0200, Teto said:
I could not find an answer yet. I've found that on latest kernel latency is on average around 3us with a max at 64us (I consider a moderate load). Is that correct ?
Depends on your hardware and system load. I'd hardly expect the same numbers from a busy Beaglebone or Raspberry Pi as from a idle high-end Intel i7 system....
(And things like processors that drop into sleep states when idle, with a latency hit to wake up, just add to the confusion....)
On Fri, 20 Jun 2014 15:36:47 +0200, Teto said:
Do you know any tool/software to retrieve the average latency for a specific program on a specific platform. This is to modelize the application rate for data processing.
You need to add "while running under a specific application load", because latencies can be different if you're under high enough load to keep the CPU running at high clock rate, versus if it's dropping to a low clock rate and then into a sleep state.... The 'latencytop' utility may help here.
participants (2)
-
Teto -
Valdis.Kletnieks＠vt.edu