Re: [Qemu-devel] [RFC] parallelize migration_bitmap_sync()

Re: [Qemu-devel] [RFC] parallelize migration_bitmap_sync()

[Paolo Bonzini]

Il 01/03/2013 00:22, Michael R. Hines ha scritto:te
> Hi,
> 
> Currently migration_bitmap_sync() is very expensive: on the order of
> 15-20 milliseconds by my count using timestamps (for a simple 2GB ram
> virtual machine).
> Until new EPT processor versions come out in 2014, we need software
> support for cutting this time down much lower........by at least an
> order of magnitude.
> 
> Would anyone be opposed to me writing a patch that creates N threads and
> dividing up the migration_bitmap_sync() function to have the dirty page
> scanning run in parallel?

Yes, that's a possibility.  You can make a quick prototype using OpenMP.

But Juan is working on making the dirty bitmap really a bitmap (not a
"bytemap").  That should speed up migration_bitmap_sync by a factor of
64 (i.e. sizeof(long)*8).

Paolo

Re: [Qemu-devel] [RFC] parallelize migration_bitmap_sync()

[Michael R. Hines]

Oh, that's fantastic - thanks for the response.

- Michael

On 03/01/2013 04:34 AM, Paolo Bonzini wrote:
> Il 01/03/2013 00:22, Michael R. Hines ha scritto:te
>> Hi,
>>
>> Currently migration_bitmap_sync() is very expensive: on the order of
>> 15-20 milliseconds by my count using timestamps (for a simple 2GB ram
>> virtual machine).
>> Until new EPT processor versions come out in 2014, we need software
>> support for cutting this time down much lower........by at least an
>> order of magnitude.
>>
>> Would anyone be opposed to me writing a patch that creates N threads and
>> dividing up the migration_bitmap_sync() function to have the dirty page
>> scanning run in parallel?
> Yes, that's a possibility.  You can make a quick prototype using OpenMP.
>
> But Juan is working on making the dirty bitmap really a bitmap (not a
> "bytemap").  That should speed up migration_bitmap_sync by a factor of
> 64 (i.e. sizeof(long)*8).
>
> Paolo
>

Re: [Qemu-devel] [RFC] parallelize migration_bitmap_sync()

[Michael R. Hines]

Juan,

Can I try this patch that Paolo mentioned?

Does it go directly from GET_LOG_DIRTY => migration_bitmap?

Or is there still an intermediate sync in your patch?

Thanks,
- Michael

On 03/01/2013 04:34 AM, Paolo Bonzini wrote:
> Il 01/03/2013 00:22, Michael R. Hines ha scritto:te
>> Hi,
>>
>> Currently migration_bitmap_sync() is very expensive: on the order of
>> 15-20 milliseconds by my count using timestamps (for a simple 2GB ram
>> virtual machine).
>> Until new EPT processor versions come out in 2014, we need software
>> support for cutting this time down much lower........by at least an
>> order of magnitude.
>>
>> Would anyone be opposed to me writing a patch that creates N threads and
>> dividing up the migration_bitmap_sync() function to have the dirty page
>> scanning run in parallel?
> Yes, that's a possibility.  You can make a quick prototype using OpenMP.
>
> But Juan is working on making the dirty bitmap really a bitmap (not a
> "bytemap").  That should speed up migration_bitmap_sync by a factor of
> 64 (i.e. sizeof(long)*8).
>
> Paolo
>