linux-kernel - Re: [BUG REPORT] perf tools: x86_64: weight column displays odd memory latency data

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <MEYPR01MB66152548BEC4D0113BBBD7A5DBDF9@MEYPR01MB6615.ausprd01.prod.outlook.com>
Date:   Wed, 1 Jun 2022 05:03:01 +0000
From:   Tyson Thomas <tyson.thomas@...ney.edu.au>
To:     "Liang, Kan" <kan.liang@...ux.intel.com>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "linux-perf-users@...r.kernel.org" <linux-perf-users@...r.kernel.org>
Subject: Re: [BUG REPORT] perf tools: x86_64: weight column displays odd
 memory latency data

Hi Kan,

Thanks for the reply.

The exact command used was: perf mem report

The columns are in order: Overhead, Samples, Local Weight, Memory Access.
The results I sent before are: Overhead, Local Weight and Memory.

So, to update the table from before, each tuple has a sample of 1, so it is:

   0.02%,1 ,62515 ,L1 or L1 hit
   0.02%,1 ,54048 ,L1 or L1 hit
   0.02%,1 ,52206 ,L1 or L1 hit
   0.02%,1 ,49831 ,L1 or L1 hit
   0.02%,1 ,49056 ,Local RAM or RAM hit
   0.01%,1 ,40666 ,LFB or LFB hit
   0.01%,1 ,38080 ,L1 or L1 hit
   0.01%,1 ,36772 ,L1 or L1 hit

So, the weights are absurdly high here. 

I'll give a reply on perf report -D shortly.

Regards,
Tyson



From: Liang, Kan <kan.liang@...ux.intel.com>
Sent: Wednesday, 25 May 2022 12:58 AM
To: Tyson Thomas <tyson.thomas@...ney.edu.au>; linux-kernel@...r.kernel.org <linux-kernel@...r.kernel.org>; linux-perf-users@...r.kernel.org <linux-perf-users@...r.kernel.org>
Subject: Re: [BUG REPORT] perf tools: x86_64: weight column displays odd memory latency data 
 


On 5/22/2022 6:14 PM, Tyson Thomas wrote:
> Hi Kan, Linux-Perf Team
> 
> I have observed some odd behaviour within perf when using perf-mem. Specifically the reported latency under the weight column appears to be unreasonably high.
> 
> Here is a given sample from a recent test, I find that some of the  latencies are close to an unsigned short and I cannot seem to be understand why that would be outside of it being an issue with the perf events.
> 
> This can be replicated using a NAS benchmark, specifically cg.D.
> 
> I observe the following results in perf mem report (just getting the top 10 results)
>


Could you please show me the exact perf command used?

With my perf mem report,
The first column is the Overhead.
The second column is the number of samples.
The third column is the weight.
The fourth is the Memory access.

Seems like the weight is missed?

Could you please check the perf report -D?
It will dump the weight for each Sample.
Does it look correct?

Thanks,
Kan

>   0.02% ,62515 ,L1 or L1 hit
>   0.02% ,54048 ,L1 or L1 hit
>   0.02% ,52206 ,L1 or L1 hit
>   0.02% ,49831 ,L1 or L1 hit
>   0.02% ,49056 ,Local RAM or RAM hit
>   0.01% ,40666 ,LFB or LFB hit
>   0.01% ,38080 ,L1 or L1 hit
>   0.01% ,36772 ,L1 or L1 hit
>   0.01% ,36729 ,LFB or LFB hit
>   0.01% ,27101 ,LFB or LFB hit
> 
> Is it possible for someone to shed some light on this or am I misunderstanding how the weight column is used here?
> This appears to have been an issue on 5.4, 5.10 and 5.15. I am looking into seeing if it is still present in 5.17 and 5.18.
> 
> I've also tried this on different Intel CPUs such as Intel Xeon 6230, i5-1135G7, Intel Xeon 6330
> 
> Any insight or help would be appreciated,
> Tyson