Re: [Qemu-devel] [PATCH v4 25/26] tcg: Check for overflow via highwater mark

[Richard Henderson]

On 10/01/2015 02:50 AM, Aurelien Jarno wrote:
> On 2015-09-30 15:09, Richard Henderson wrote:
> > We currently pre-compute an worst case code size for any TB, which
> > works out to be 122kB.  Since the average TB size is near 1kB, this
> > wastes quite a lot of storage.
>
> The code generation buffer is currently computed as 1/4 of the guest
> RAM in softmmu mode (so 32MB for the default 128MB of RAM) or 32MB in
> user mode. 122kB is therefore less than 0.4% of waster memory, I am not
> therefore sure we need to add so much code just for that.
>
> BTW, I wonder if it is really a good idea to scale the code generation
> buffer with the size of the RAM, as the two do not seem that related. It
> happens that at some point we don't really increases performances
> anymore, and always defining it as 32MB might actually be a good idea.
> Personally I am using a patch that limits it to 128MB.

It depends on the guest and the workload I suppose.

For alpha, which almost never flushes the buffer, allocating 2GB to the guest 
(and thus 512MB to the buffer) seems to reach a nice steady-state during a 
guest gcc bootstrap where no more code generation is required.

If I were to limit the buffer to significantly less, as you're suggesting, I 
believe that the working set would overflow the buffer, requiring more code 
generation.

I guess I could always use -tb-size myself to override...



r~