Re: [Qemu-devel] [Qemu-block] [PATCH] qcow2: do lazy allocation of the L2 cache

[Stefan Hajnoczi]

On Fri, Apr 24, 2015 at 12:10 PM, Alberto Garcia <address@hidden> wrote:
> On Fri 24 Apr 2015 11:52:14 AM CEST, Kevin Wolf <address@hidden> wrote:
>
>>> The posix_memalign() call wastes memory.  I compared:
>>>
>>>   posix_memalign(&memptr, 65536, 2560 * 65536);
>>>   memset(memptr, 0, 2560 * 65536);
>>>
>>> with:
>>>
>>>   for (i = 0; i < 2560; i++) {
>>>       posix_memalign(&memptr, 65536, 65536);
>>>       memset(memptr, 0, 65536);
>>>   }
>>
>> 64k alignment is too much, in practice you need 512b or 4k, which
>> probably wastes a lot less memory.
>
> My tests were with 512b and 4k and the overhead was around 4k per page
> (hence 2560 * 4 = 10240, the 10MB I was talking about).
>
>> But I just looked at the qcow2 cache code and you're right anyway.
>> Allocating one big block instead of many small allocations in a loop
>> looks like a good idea either way.
>
> The problem is that I had the idea to make the cache dynamic.
>
> Consider the scenario [A] <- [B], with a virtual size of 1TB and [B] a
> newly created snapshot. The L2 cache size is 128MB for each image, you
> read a lot of data from the disk and the cache from [A] starts to fill
> up (at this point [B] is mostly empty so you get all the data from
> [A]). Then you start to write data into [B], and now its L2 cache starts
> to fill up as well.
>
> After a while you're going to have lots of cache entries in [A] that are
> not needed anymore because now the data for those clusters is in [B].
>
> I think it would be nice to have a way to free unused cache entries
> after a while.

Do you think mmap plus a periodic timer would work?

I'm hesitant about changes like this because they make QEMU more
complex, slow down the guest, and make the memory footprint volatile.
But if a simple solution addresses the problem, then I'd be happy.

Stefan