[Qemu-devel] How to precisely monitor all the memory references in QEMU to feed the cache model

[Qemu-devel] How to precisely monitor all the memory references in QEMU to feed the cache model

[shocklink99]

[-- Attachment #1: Type: text/plain, Size: 981 bytes --]

Hi,
I'm adding a cache model  into QEMU 0.12
I have encountered a problem that the cache miss error rate was high
compared to real platform Creator(arm926) .
I used the QEMU integrator board to run the experiment.
I've modified
softmmu_header.h
softmmu_template.h
target-arm/translate.c

e.g. in softmmu_header.h
glue(glue(ld, USUFFIX), MEMSUFFIX)(target_ulong ptr)
I monitored the ptr, I know that ptr is the access address

e.g. in softmmu_template.h
glue(glue(__ld, SUFFIX), MMUSUFFIX)(target_ulong addr


e.g. target-arm/translate.c
static inline TCGv gen_ld8s(TCGv addr, int index)
{
    TCGv tmp = new_tmp();

    gen_helper_cache_access(addr , tcg_const_i32(1) );

    tcg_gen_qemu_ld8s(tmp, addr, index);
    return tmp;
}

I had taken care of all the related function.
Is there  anything I ignored when running with the model?

I have also reduce the timer interrupt to make it closed enough with the
real platform.
SO the context switch overhead should be little enough.

[-- Attachment #2: Type: text/html, Size: 1164 bytes --]

Re: [Qemu-devel] How to precisely monitor all the memory references in QEMU to feed the cache model

[Blue Swirl]

On 2/27/10, shocklink99@yahoo.com.tw <shocklink99@yahoo.com.tw> wrote:
> Hi,
> I'm adding a cache model  into QEMU 0.12
> I have encountered a problem that the cache miss error rate was high
> compared to real platform Creator(arm926) .
> I used the QEMU integrator board to run the experiment.
>  I've modified
> softmmu_header.h
> softmmu_template.h
> target-arm/translate.c
>
> e.g. in softmmu_header.h
> glue(glue(ld, USUFFIX), MEMSUFFIX)(target_ulong ptr)
> I monitored the ptr, I know that ptr is the access address
>
> e.g. in softmmu_template.h
> glue(glue(__ld, SUFFIX), MMUSUFFIX)(target_ulong addr
>
>
> e.g. target-arm/translate.c
> static inline TCGv gen_ld8s(TCGv addr, int index)
> {
>     TCGv tmp = new_tmp();
>
>      gen_helper_cache_access(addr , tcg_const_i32(1) );
>
>     tcg_gen_qemu_ld8s(tmp, addr, index);
>     return tmp;
> }
>
> I had taken care of all the related function.
> Is there  anything I ignored when running with the model?
>
> I have also reduce the timer interrupt to make it closed enough with the
> real platform.
> SO the context switch overhead should be little enough.

Interesting approach, this could be useful for modeling caches on
other architectures.

If your cache statistics merge I/D cache values, this could be one
source of error. I'd suppose pure data cache statistics should be
closer to reality, but QEMU's mode of operation for instruction
accesses differs a lot from real CPU:
 * on some architectures, translation may access some instructions in
the basic block which would not be executed by the CPU
 * we have a large TB cache, its size could be different from real CPU cache
 * TBs are flushed by QEMU (and that logic is different)
 * cache flushes by guest are ignored (this also applies to data caches)