• State
  • Locked Locked
  • Replies 2 replies
  • Subscribers 98 subscribers
  • Views 675 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

PROCESSOR-SDK-AM57X: Context switch latency

hungwei yang
Expert 1090 points
Part Number: PROCESSOR-SDK-AM57X


Hello,

Software: PROCESSOR-SDK-LINUX-RT-AM57X V04.03

Hardware: AM574X-IDK

I m measuring the overhead of context switch by lat_ctx which is mentioned on this page. The result show that it spends around 10~15us on context switch each time which is more than I expected. Is the number reasonable? If it is, do we have any ways to optimize the overhead? Like increasing CPU clock, or upgrade Linux version? Thanks for any comment.

root@am57xx-evm:~# uname -a
Linux am57xx-evm 4.9.65-rt23-g7069a470d5 #2 SMP PREEMPT RT Mon Mar 26 12:04:24 EDT 2018 armv7l GNU/Linux

root@am57xx-evm:~# ./lat_ctx -s 0 processes 2 4 8 16 20

"size=0k ovr=6.05
2 10.13
4 13.85
8 14.26
16 14.16
20 13.68
root@am57xx-evm:~# ./lat_ctx -s 0 processes 2 4 8 16 20

"size=0k ovr=6.14
2 10.25
4 13.12
8 14.34
16 14.65
20 14.32
root@am57xx-evm:~# ./lat_ctx -s 0 processes 2 4 8 16 20

"size=0k ovr=6.24
2 10.08
4 12.83
8 13.90
16 14.02
20 13.94

  • 0
    TI__Mastermind 30646 points

    Hello,

    The SDK you are using is beyond our support window. It would be best to move to a later SDK. We publish results for lat_ctx in our performance guide which can be found here for the RT kernel:

    http://software-dl.ti.com/processor-sdk-linux/esd/docs/06_00_00_07/linux/Release_Specific_Performance_Guide.html#rt-kernel-performance-guide

    What are you basing your expectations on? Or, why do you feel that this is not the expected performance?

    Thank you.

  • 0 in reply to RonB
    Expert 1090 points

    Hello,

    It is nice that you have run the benchmark for the latest kernel version. But why the time of lat_ctx-4-256k is 0?

    lat_ctx-2-128k (us) 4.79
    lat_ctx-2-256k (us) 4.00
    lat_ctx-4-128k (us) 8.07
    lat_ctx-4-256k (us) 0.00

    >What are you basing your expectations on? Or, why do you feel that this is not the expected performance?

    I expect that it spends 1 or 2 us on context switch for simple applications like notepad.

    Lets say we have one program A which is not so smart and sleep per ms by calling usleep(1000). This program A causes context-switch two times (others -> A and A -> others) per ms. Assume the overhead of context-switch is around 5 us, then the theoretical value of overhead will be (5 * 2 / 1000) = 1%. 

    So....usleep(1000) wastes the 1% of CPU resource??