About %30 degradation of performance when enabling pseudo nmi

About %30 degradation of performance when enabling pseudo nmi

[liwei (GF)]

Hi Julien,

	I did a redis benchmark test on kernel-4.19.33, with pseudo nmi enabled.
The result indicated that there is about %30 degradation of performance comparing
with the one pseudo nmi disabled.

    Redis 4.0.8
    Test: LPOP
    Requests Per Second > Higher Is Better
    redis-419-numa-aware-nmi . 277342 |===========================
    redis-4.19-numa-aware .... 365635 |====================================


    Redis 4.0.8
    Test: SADD
    Requests Per Second > Higher Is Better
    redis-419-numa-aware-nmi . 335625 |==========================
    redis-4.19-numa-aware .... 459141 |====================================


    Redis 4.0.8
    Test: LPUSH
    Requests Per Second > Higher Is Better
    redis-419-numa-aware-nmi . 270748 |===========================
    redis-4.19-numa-aware .... 363677 |====================================


    Redis 4.0.8
    Test: GET
    Requests Per Second > Higher Is Better
    redis-419-numa-aware-nmi . 349754 |=========================
    redis-4.19-numa-aware .... 496333 |====================================


    Redis 4.0.8
    Test: SET
    Requests Per Second > Higher Is Better
    redis-419-numa-aware-nmi . 310413 |===========================
    redis-4.19-numa-aware .... 408679 |====================================

	I tried to analyse and found the hotpoint is around "msr    daif" and "dsb"
instructions which i think is appropriate. I have no good idea about how to improve,
could you please look into this and give some suggestion?

Thanks,
Wei


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: About %30 degradation of performance when enabling pseudo nmi

[Hanjun Guo]

On 2019/4/6 18:39, liwei (GF) wrote:
> Hi Julien,
> 
> 	I did a redis benchmark test on kernel-4.19.33, with pseudo nmi enabled.
> The result indicated that there is about %30 degradation of performance comparing
> with the one pseudo nmi disabled.
> 
>     Redis 4.0.8
>     Test: LPOP
>     Requests Per Second > Higher Is Better
>     redis-419-numa-aware-nmi . 277342 |===========================
>     redis-4.19-numa-aware .... 365635 |====================================
> 
> 
>     Redis 4.0.8
>     Test: SADD
>     Requests Per Second > Higher Is Better
>     redis-419-numa-aware-nmi . 335625 |==========================
>     redis-4.19-numa-aware .... 459141 |====================================
> 
> 
>     Redis 4.0.8
>     Test: LPUSH
>     Requests Per Second > Higher Is Better
>     redis-419-numa-aware-nmi . 270748 |===========================
>     redis-4.19-numa-aware .... 363677 |====================================
> 
> 
>     Redis 4.0.8
>     Test: GET
>     Requests Per Second > Higher Is Better
>     redis-419-numa-aware-nmi . 349754 |=========================
>     redis-4.19-numa-aware .... 496333 |====================================
> 
> 
>     Redis 4.0.8
>     Test: SET
>     Requests Per Second > Higher Is Better
>     redis-419-numa-aware-nmi . 310413 |===========================
>     redis-4.19-numa-aware .... 408679 |====================================
> 
> 	I tried to analyse and found the hotpoint is around "msr    daif" and "dsb"
> instructions which i think is appropriate. I have no good idea about how to improve,
> could you please look into this and give some suggestion?

Although it was tested on 4.19 stable kernel, I think mainline kernel will hit the
same performance issue as well.

Thanks
Hanjun


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel