Re: Question about LLC-load-misses event

[Chulmin Kim <cmkim@core.kaist.ac.kr>]

2012-10-23 오후 2:53, Chulmin Kim 쓴 글:
> 2012-10-23 오후 2:39, Namhyung Kim 쓴 글:
>> Hi Chulmin,
>>
>> On Mon, 15 Oct 2012 22:31:34 +0900, Chulmin Kim wrote:
>>> 2012-10-15 오후 10:07, Chulmin Kim 쓴 글:
>>>> (perf command : perf stat -a -A -e LLC-loads -e LLC-load-misses -e
>>>> instructions sleep 3)
>>>>
>>>> The problem is,, the bandwidth from STREAM benchmark does not match 
>>>> with
>>>> the monitored value.
>>>>
>>>> e.g.
>>>> I got 9395MB/s from Stream.
>>>>
>>>> "perf" shows 134,642,063 LLC-load-misses for 3 seconds.
>>>> -> BW = ((# of events)/(3 seconds)) * 64 bytes / (1024*1024) = 
>>>> 2739MB/s
>>>> In this equation, the term (64bytes) is for cache line size, and the
>>>> term(1024*1024) is for (MB/s).
>>>>
>>>> Why does this mismatch occur?
>>> In case of Oprofile, the value for a certain event represents the 
>>> number
>>> of the overflows which occur when the number of the event exceeds the
>>> predefined value.
>>> Is it a similar case with that?
>> I guess not.  And what's the result of the LLC-loads?  AFAIK it counts
>> all cache accesses including hits and misses.
>
> Sorry for the lack of information.
>
> I used STREAM benchmark which generates 100% cache miss.
> Of course, the value of LLC-loads shows bit larger number than that of 
> LLC-load-misses (but, they are almost same).
>
>
>>    Did you calculate the
>> bandwidth using the result of LLC-loads?
>
> Bandwidth results:
> 9395MB/s from Stream
> 2739MB/s from LLC-load (including both hit and miss)
> I also want to add the BW from the mem write (about 3000MB/s from 
> LLC-store (including both hit and miss) )
>
> I'm wondering why this difference happens?  (9395MB/s vs about 5739MB/s)
>
>
>
>> I suspect the h/w *might*
>> prefetches a couple of lines when cache-miss occurred, but I'm not
>> sure. :)
>
> I also suspected PREFETCH.
>
> After i uploaded this question, i checked prefetch events using "perf".
> I got 100% prefetch miss also. (but i don't know the meaning of this 
> result thoroughly.)
>
> Do you know what this means?
> Are Prefetch events and LLC-load (or store) events exclusive? or 
> correlated?
>
> Perf tool is too hard for me.. No remarkable document or site to help 
> the user.  (If you know, recommend it please! :) )
>
>
> Thanks for your attention!
>
>

In the end, I changed BIOS setting of my own machine to turn off 
prefetch features.
(Hardware Prefetch & Adjacent Cache Line Prefetch)

Finally, the bandwidth results of STREAM and PMU events are consistent!

I guess it was an issue related with the prefetching though I couldn't 
anlalyze it thouroughly.



Thanks!

>
>
>> Thanks,
>> Namhyung
>>
> -- 
> To unsubscribe from this list: send the line "unsubscribe 
> linux-perf-users" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>