Re: [Valgrind-users] tool=cachegrind always reports LL misses, even if the cache is bigger than the used memory

[Eliot M. <mo...@cs...>]

On 4/25/2023 3:56 PM, Volker Dirr wrote:
> Hallo,
> 
> maybe I misunderstood, but it look like I don't understand tool=cachegrind correct (or there is a bug).
> 
> I have a software. If it runs, then the task manager tells me that is use only 38 MB memory.
> 
> Now i used cachegrind like this:
> valgrind --tool=cachegrind --LL=2097152,16,64 ./fet-cl --inputfile=German-100_and_0.fet  
> --randomseeds10=10 --randomseeds11=11 --randomseeds12=12 --randomseeds20=20 --randomseeds21=21 
> --randomseeds22=22
> 
> i got this report:
> [...]
> ==43032== LL misses:           8,283,261  (      7,840,142 rd   + 443,119 wr)
> 
> 
> i doubled the LL cache by this:
> valgrind --tool=cachegrind --LL=4194304,16,64 ./fet-cl --inputfile=German-100_and_0.fet  
> --randomseeds10=10 --randomseeds11=11 --randomseeds12=12 --randomseeds20=20 --randomseeds21=21 
> --randomseeds22=22
> 
> and got this:
> [...]
> ==48663== LL misses:             230,426  (         89,082 rd   + 141,344 wr)
> 
> 
> That sound ok (larger cache should reduce the misses).
> 
> But I continued to double the cache again and again. Up to:
> valgrind --tool=cachegrind --LL=268435456,16,64 ./fet-cl --inputfile=German-100_and_0.fet  
> --randomseeds10=10 --randomseeds11=11 --randomseeds12=12 --randomseeds20=20 --randomseeds21=21 
> --randomseeds22=22
> 
> So much more memory then my program use at all.
> But the misses never drop down to 0. They stay at:
> [...]
> ==6637== LL misses:             180,120  (         41,252 rd   + 138,868 wr)
> 
> 
> I don't understand that. Shouldn't the misses drop down to 0 as soon as LL is >64MB (since my 
> software use only 38 MB)? (But i tried up to 256MB and it doesn't drop).
> Is that a bug in valgrind or is there a bug in my logic in understanding the LL misses?
> 
> Please let me know.

I would think that, regardless of cache size, the *first* access to a line
causes a miss.  These are sometimes called mandatory misses.  A large cache
*will* eliminate *capacity* misses.

HTH -- Eliot Moss