imx7 vs imx8 unexpected difference in performance

[mikyago]

Good morning,

we're performing a performance test on our 2 platforms, imx7d and imx8 mmini, to investigate an issue we're facing.

The test is based in a basic communication between 2 threads, implemented with the pthread library, and it works as follow:

1) One thread is in waiting with a conditional wait

2) the other thread wakes it up

3) elapsed time computation

4) threads exchange roles

5) cycle restarts

Fact is: we see very different performance with this test between imx7 and imx8 platforms.

To make a deeper analysis, we make these measurements in 3 different conditions:

• threads on the same core
• threads on different cores
• OS decides on its own

The most critical condition, the one in which the 2 platforms perform very differently is when threads are in different cores. We even tried to put 2 cores offline in the imx8 platform, but as it is possible to see below, nothing changes. Can someone help us with this situation? Are we missing something?

Further details:

We tried to use this system settings to get a better system stability for all the tests:

• scaling_governor: performance
• dynamic frequency scaling driver: disabled

(i.e. on imx8: echo 0 > /sys/bus/platform/drivers/imx_busfreq/busfreq/enable)

Data

imx7d

uname -r:  4.9.11+gf1a31cc

Forcing threads on different cores

[T1] Average is 10 us;

[T2] Average is 14 us;

Forcing threads on same core

[T1] Average is 10 us;

[T2] Average is 14 us;

OS scheduler decides thread-core affinity

[T1] Average is 11 us;

[T2] Average is 15 us;

imx8mmini (all cores online)

uname -r: 4.14.78-imx_4.14.78_1.0.0_ga_dev+g991fec2

Forcing threads on different cores

[T1] Average is 493 us;

[T2] Average is 458 us;

Forcing threads on same core

[T1] Average is 10 us;

[T2] Average is 10 us;

OS scheduler decides thread-core affinity

[T1] Average is 507 us;

[T2] Average is 448 us;

imx8mmini (only 2 cores online)

uname -r: 4.14.78-imx_4.14.78_1.0.0_ga_dev+g991fec2

Forcing threads on different cores

[T1] Average is 474 us;

[T2] Average is 379 us;

Forcing threads on same core

[T1] Average is 10 us;

[T2] Average is 10 us;

OS scheduler decides thread-core affinity

[T1] Average is 480 us;

[T2] Average is 384 us;

As it is possible to see the difference is sensible. We're available for any further useful test, just let us know.

Any help would be appreciated, thanks!

[igorpadykov]

Hi Miky

nxp provides linux bsps described on below link, which are optimized for

demonstration all processor (as much as possible) capabilities

i.MX Software | NXP 

For specific customer needs may be recommended to proceed with help

of extended support of NXP Professional Services | NXP 

so necessary adjustments were made for improving performance

for specific customer use cases.

Best regards
igor
-----------------------------------------------------------------------------------------------------------------------
Note: If this post answers your question, please click the Correct Answer button. Thank you!
-----------------------------------------------------------------------------------------------------------------------